scai

gistfile1.txt

 CONTENTS    |  2 
 CONTENTS 
 ABOUT SCAI  3 
 FOREWORD TO THE SCAI QUESTIONS  4 
 QUESTION 1  RELIABILITY & TRUSTWORTHINESS  6 
 QUESTION 2  DATA COLLECTION & SHARING  11 
 QUESTION 3  GOVERNANCE STRUCTURE & REGULATORY MEASURES  17 
 QUESTION 4  SOLVING SCIENTIFIC PROBLEMS  24 
 QUESTION 5  MODELS & ARCHITECTURE DERIVED FROM NATURAL INTELLIGENCE  28 
 QUESTION 6  VALUES & NORMS TO ALIGN AI: ELICITATION & IMPLEMENTATION  34 
 QUESTION 7  EQUITABLE ACCESS, CONTROL & FAIR COMPETITION  41 
 QUESTION 8  TRANSFORMING EDUCATION  45 
 QUESTION 9  MITIGATING CATASTROPHIC RISKS & ONGOING HARMS  50 
 QUESTION 10  COMBATING MIS/DISINFORMATION CAMPAIGNS  56 
 QUESTION 11  A FRAMEWORK FOR EFFECTIVE AI ADOPTION FOR SOCIAL GOOD  60 
 QUESTION 12  METHODOLOGIES FOR AI SAFETY EVALUATION  67 
 ACKNOWLEDGEMENTS  73 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved. 
 ABOUT SCAI    |  3 
 ABOUT SCAI 
 AI  has  the  potential  to  enhance  our  quality  of  life,  revolutionise  industries,  and  transform  the  way 
 we  live  and  work.  However,  there  are  many  potential  challenges  that  may  constrain  our  ability  to 
 harness  the  technology  to  benefit  societies  and  people,  such  as  accuracy,  bias,  and 
 resource-e ciency. 
 To  overcome  these  challenges,  the  Singapore  Ministry  of  Communications  and  Information  and 
 Smart  Nation  Group,  in  partnership  with  the  Topos  Institute,  organised  the  inaugural  Singapore 
 Conference  on  AI  for  the  Global  Good,  or  SCAI,  from  4  to  6  December  2023.  The  conference 
 brought  together  42  distinguished  experts  from  various  fields  of  academia,  industry,  and 
 government. 
 Drawing  on  their  diverse  domains  of  expertise,  delegates  explored  and  articulated  critical 
 questions  of  AI  that,  if  answered,  will  enable  the  development  and  deployment  of  AI  for  societies 
 to flourish. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved. 

 FOREWORD    |  4 
 FOREWORD TO 
 THE SCAI QUESTIONS 
 These  Questions  wer
e  conceptualised  and  written  by  the  SCAI  delegates  over  the  3  days  of  the 
 conference,  using  a  process  designed  to  synthesise  diverse  views  from  experts.  Each  of  the  12 
 SCAI  Questions  is  envisioned  to  be  a  comprehensive  articulation  of  a  foundational,  yet  tractable 
 area  of  AI  development  and/or  deployment.  The  12  SCAI  Questions  taken  as  a  whole,  are  meant  to 
 be  a  holistic  formulation  of  the  challenges  that  should  be  addressed  by  the  global  AI  community  to 
 allow humanity to flourish. 
 Good  questions  are  hard  to  frame,  especially  in  a  domain  as  emergent  and  boundary-spanning  as 
 AI.  The  SCAI  Questions  are  the  best-effort  of  the  delegates  gathering  together,  debating  and  then 
 consolidating  their  views  over  3  days  in  Singapore.  They  are  certainly  not  final,  and  we  invite 
 commentators  and  researchers  to  use  these  Questions  as  a  springboard  for  further  research, 
 collaboration and innovation. 
 In  terms  of  format,  each  Question  begins  by  stating  upfront  the  context  and  assumptions  which 
 the  delegates  had  in  mind,  followed  by  an  elaboration  of  the  possible  approaches  to  answering  the 
 question, known challenges, and ways we might recognise progress. 
 The  SCAI  Questions  are  a  collective  and  collaborative  product  of  the  conversations  amongst  SCAI 
 delegates.  They  do  not  necessarily  represent  the  views  of  individual  participants,  or  the 
 organising parties of SCAI. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved. 
 FOREWORD    |  5 
 The  ordering  of  the  SCAI  Questions  is  solely  for  ease  of  reference  and  does  not  reflect  a  hierarchy 
 of importance. 
 For  referencing  or  citing  this  document  in  academic  or  professional  contexts,  please  use  the 
 following format: 
 Singapore  Ministry  of  Communications  and  Information  &  Smart  Nation  Group,  in  partnership  with 
 Topos  Institute.  (2023).  Preliminary  Conversations  Towards  AI  for  the  Global   Good:  The  SCAI 
 Questions  .  Proceedings  of  the  Singapore  Conference  on  AI  for  the  Global  Good,  4-6  December 
 2023, Singapore. Available at: https://www.scai.gov.sg/findings. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved. 

 SCAI QUESTION 1    |  6 
 SCAI  QUESTION 1 

 SCAI QUESTION 1 
 RELIABILITY & TRUSTWORTHINESS 
 How  do  we  ensure  that  AI  models  and  systems  are  reliable  and 
 trustworthy? 
 Context & Assumptions 
 AI  systems  are  increasingly  being  used  for  decision-making  across  various  fields,  including 
 critical,  high  stakes  areas  like  medicine.  However,  current  systems  are  not  always  reliable  nor 
 trustworthy.  For  instance,  language  models  often  “hallucinate”,  producing  outputs  that  are 
 inconsistent  and  not  grounded  in  reality.  The  result  is  that  human  users  cannot  trust  the  output 
 and cannot rely solely on the models to make decisions. 
 For  an  AI  system  to  be  reliable,  it  should  consistently  produce  outputs  that  align  with  a  specific, 
 well-defined  set  of  requirements,  even  when  deployed  in  new  or  changing  environments.  For 
 example,  a  self-driving  car  should  adhere  to  a  specified  performance  standard  in  terms  of  speed 
 and  stability  under  different  weather  and  trac  conditions;  and  a  medical  language  model  should 
 provide accurate and up-to-date medical information. 
 Trustworthiness,  on  the  other  hand,  is  a  broader  function  of  the  relationship  between  the  AI 
 system  and  its  human  users.  For  example,  it  includes  whether  the  behaviour  of  an  AI  system  is 
 consistent  with  its  users’  expectations;  whether  it  is  in  accord  with  human  norms  of  fairness  and 
 ethics;  and  if  it  is  robust  to  adversarial  conditions.  For  instance,  a  self-driving  car  might  be 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 1    |  7 
 deemed  untrustworthy  if  it  drives  unlike  a  human,  or  a  language  model  might  lose  trust  if  it  is 
 found to be biased against certain demographics. 
 Question 
 How do we ensure that AI models and systems are reliable and trustworthy? 
 Evaluation  is  a  key  aspect  of  reliability.  How  do  we  design  specifications  for  complex,  open-ended 
 tasks,  and  then  evaluate  models  against  them?  Ideally,  these  specifications  should  be 
 standardised,  reproducible,  lightweight,  but  also  as  close  as  possible  to  actual  downstream  tasks; 
 they  should  capture  all  relevant  aspects  of  the  desired  model  behaviour  (e.g.,  not  just  average 
 accuracy  but  also  notions  of  bias);  and  they  should  be  evaluated  in  environments  that  are 
 reflective  of  real-world  settings,  including  end-to-end  tests  with  users  as  appropriate.  Designing 
 such  specifications  is  particularly  challenging  for  generative  AI  systems,  where  the  output  space 
 is  large  and  automated  evaluation  is  dicult.  Beyond  the  standard  approach  of  empirical  testing, 
 an  open  question  is  whether  we  can  provably  verify  the  reliability  of  white-box  models,  which  is 
 especially relevant to the development of defences against adversarial actors. 
 To  achieve  trustworthiness,  reliability  is  necessary  but  not  sucient;  we  also  need  to  consider 
 the  interaction  between  the  full  system  and  the  user.  An  open  question  is  how  to  design  systems 
 so  that  the  decision  making  process  by  the  system  is  observable  and  interpretable  to  the  user. 
 The  structure  of  the  interaction  and  the  communication  between  the  system  and  the  user  can 
 significantly  influence  the  level  of  trust  users  have  in  the  system  and  must  be  designed 
 appropriately.  Finally,  trust  requires  that  the  specification  itself  be  correct  and  useful  within  the 
 user's context. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 1    |  8 
 Systems  operate  continuously  –  trust  is  engendered  when  systems  maintain  reliability  in  novel 
 environments  and  remain  current  with  evolving  knowledge.  Providing  reliable  performance  over 
 time  depends  on  recognition  of  distribution  shift,  detection  of  out-of-distribution  queries,  and 
 the  ability  to  respond  appropriately.  An  open  question  is  whether  this  adaptation  necessitates  an 
 understanding  of  the  world,  including  context,  causal  structures,  and  implicit  motivations.  Issues 
 of  uncertainty,  calibration,  and  security  are  also  pertinent.  How  well  an  AI  system  can  handle 
 uncertainty  and  how  accurately  it  is  calibrated  are  crucial  for  its  effective  functioning.  Security 
 considerations, especially in the face of potential adversarial attacks, cannot be overlooked. 
 Indicators of Progress 
 A  key  indicator  of  progress  will  be  the  development  of  standardised  benchmarks  that,  as 
 elaborated  above,  capture  the  relevant  aspects  of  reliability   and  trustworthiness  in  a  broad  range 
 of  real-world  applications  of  AI,  and  in  particular  in  open-ended  tasks  that  require  more 
 sophisticated  evaluation.  As  technical  progress  might  overfit  to  existing  benchmarks,  established 
 processes  for  continually  creating  more  diverse  and  realistic  benchmarks  would  be  another 
 indicator of progress. 
 Another  indicator  of  progress  will  be  a  codified  set  of  principles  for  "Design  for  Reliability,"  akin  to 
 "Design  for  Manufacturing",  which  will  allow  AI  systems  to  be  specifically  designed  with  the  goal  of 
 meeting  specifications  that  go  beyond  statistical  performance.  These  principles  will  include 
 developing  methods  for  uncertainty  quantification  and  introspection,  continual  learning, 
 out-of-distribution  robustness,  explainability,  and  enabling  AI  systems  to  recognize  and 
 communicate  the  limits  of  its  knowledge.  For  trustworthiness  in  adversarial  environments,  the 
 development  of  methods  for  provably  verifying  model  outputs  will  be  another  indicator  of 
 progress. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 1    |  9 
 Finally,  a  significant  indicator  of  progress  will  be  the  deployment  of  AI  across  various  domains, 
 gradually  increasing  in  scope,  autonomy  and  operational  duration.  This  would  collectively  show 
 the advancement and maturity of AI systems in being both reliable and trustworthy. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 1    |  10 
 SCAI QUESTION 2    |  11 
 SCAI  QUESTION 2 

 SCAI QUESTION 2 
 DATA COLLECTION & SHARING 
 How  can  we  create  a  data  collection  and  sharing  ecosystem  that 
 produces  high-quality  data  for  AI,  which  can  be  shared  and  exploited 
 within and across countries? 
 Context & Assumptions 
 High-quality  data  leads  to  high-quality  AI.  Training  large  models  currently  requires  large  amounts 
 of  high-quality  data.  To  have  high-quality  data,  one  needs  to  consider  what  would  be  appropriate 
 data  governance,  technologies,  and  infrastructures  to  manage  the  challenges  of  data  collection, 
 deal  with  data  fragmentation,  and  support  data  integration.  In  addition,  there  is  a  need  to  address 
 issues  concerning  legal  real  time  continuous  data  (e.g.,  sensor  data),  fact  editing,  and  challenges 
 around  model  collapse  and  reproducibility.  Data  about  the  construction  and  performance  of  AI 
 models themselves is an important asset. 
 Building  good  datasets  has  been   problematic.  In  some  domains,  acquiring  data  can  be  a 
 challenge.  For  example,  in  healthcare,  developing  good  models  requires  diverse,  high-quality, 
 accurate  datasets  from  the  local  community,  but  this  is  dicult  to  achieve  given  the  sensitivity 
 and  privacy  of  such  data;  these  challenges  persisted  even  during  the  COVID-19  pandemic,  where 
 data  was  essential  to  combatting  its  spread.  Also,  user  generated  content,  a  goldmine  of  potential 
 insights,  is  often  proprietary.  Companies  collecting  this  data  are  cautious  about  sharing,  as  it  has 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 2    |  12 
 competitive value and comes with privacy concerns and legal obligations. 
 Despite  the  challenges,  some  datasets  and  resources  are  maintained  to  high  standards.  This 
 ranges  from  collaborative  maintained  resources  such  as  Wikipedia  to  cultural  heritage  data,  e.g., 
 Europeana,  the  European  digital  cultural  platform,  which  allows  museums,  galleries,  libraries, 
 archives  across  Europe  to  share  and  reuse  digitised,  standardised  cultural  heritage  images  such 
 as  3D  models  of  historical  sites,  and  high-quality  scans  of  paintings.  Additionally,  biomedical  data, 
 such  as  genetic  information,  clinical  trial  results,  and  disease  data,  are  typically  well-catalogued 
 and publicly available, fostering scientific collaboration and research. 
 Question 
 How  can  we  create  a  data  collection  and  sharing  ecosystem  that  produces  high-quality  data  for 
 AI, which can be shared and exploited within and acr
oss countries? 
 A  robust  data  collection  and  data-sharing  ecosystem  will  also  allow  us  to  address  the  following 
 key issues: 
 ●  Measuring  data  quality  :  Principles  guiding  the  measurement  of  data  quality  are  essential 
 for  fostering  reliable  AI  models.  Adhering  to  the  FAIR  principles  (Findable,  Accessible, 
 Interoperable,  Re-usable)  lays  the  foundation  for  robust  datasets.  Additionally,  for  factors 
 such  as  diversity,  representativeness,  openness,  trustworthiness  and  safety,  the 
 incorporation  of  mechanisms  for  both  prevention  and  assistance  in  the  unlearning 
 processes  of  AI  models  are  crucial.  These  principles  collectively  contribute  to  the 
 integrity and e cacy of AI models and applications. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 2    |  13 
 ●  Data  valuation  :  Relatedly,  the  assignment  of  value  to  data  is  important  in  any  effort  to 
 enable  effective  sharing.  Data  valuation  can  help  to  guide  the  selection,  management,  and 
 the stewarding of data. 
 ●  Scaling  data  :  Large  language  models  require  massive  amounts  of  text  data  for  training. 
 Other  frontier  models  require  significant  data  in  other  modalities;  image,  video,   audio  etc. 
 Acquiring  and  curating  such  large  and  diverse  datasets  can  be  logistically  challenging  and 
 resource-intensive.  Scaling  data  to  accommodate  large  volumes  necessitates  a  federated 
 approach.  Discovering  and  linking  data  across  domains,  maintaining  standards  to  enable 
 data  sharing,  maintaining  a  balanced  representation  across  various  demographics,  and 
 exploring the generation and use of synthetic data are potential approaches. 
 ●  Data  privacy  :  Ensuring  data  privacy  is  paramount  in  a  data-sharing  ecosystem. 
 Implementing  robust  privacy  measures  involves  anonymisation,  encryption,  and 
 compliance  with  international  privacy  regulations.  Striking  a  balance  between  data  utility 
 and privacy protection is essential for fostering trust among data contributors and users. 
 ●  Data  rights  :  Establishing  clear  ownership  frameworks  ensures  responsible  data 
 stewardship  and  facilitates  ethical  data  sharing  practices.  Some  of  the  primary 
 challenges  in  managing  data  rights  pertains  to  the  constraint  imposed  by  political 
 boundaries—commonly  known  as  data  sovereignty.  Additionally,  navigating  data 
 copyrights  requires  understanding  intellectual  property  laws  and  implementing 
 mechanisms to protect creative elements within datasets. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 2    |  14 
 ●  Data  reproducibility  :  Data  reproducibility  is  critical  for  the  scientific  community  and  AI 
 practitioners.  Implementing  practices  such  as  sharing  code,  documenting  methodologies, 
 and  providing  access  to  raw  data  enhances  the  reproducibility  of  AI  research.  Open  and 
 transparent  practices  will  continue  contributing  to  the  credibility  and  reliability  of  AI 
 models and findings. 
 Indicators of Progress 
 For  data  to  be  effectively  exploited  to  build  AI  for  good,  we  will  need  to  consider  a  range  of  levers, 
 including governance and technological approaches. 
 The  governance  approach,  including  setting  up  relevant  regulations,  incentives,  subsidies,  data 
 formatting  and  sharing  agreements,  can  encourage  the  creation  and  flow  of  data.  For  example, 
 one  can  mandate  that  datasets  that  are  created  through  the  support  of  federal  funds  should  be 
 publicly  available  in  a  standardised  format.  Furthermore,  one  can  encourage  the  adoption  of  an 
 agreed  methodology  to  describe  the  origin,  nature,  and  use  of  data.  The  UK  Biobank  is  one  such 
 example.  Defining  a  taxonomy  for  data  can  also  help  classify  data  more  effectively  for  appropriate 
 privacy classification and data licensing. 
 In  addition,  there  should  be  equivalent  focus  on  novel  technical  approaches  to  achieve 
 high-quality  data  for  AI.  This  would  involve  research  investments.  Current  examples  include 
 privacy  enhancing/preserving  technology  and  development  of  Trusted  Research  Environments 
 (TRE).  TREs  allow  for  model  development  and  deployment  by  leveraging  data  in  a  decentralised 
 and  secure  manner.  New  technologies  to  enable  data  provenance  can  also  enable  safer 
 development  of  AI,  especially  large  foundation  models,  so  as  to  allow  for  appropriate  model 
 unlearning, licensing, watermarking etc. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 2    |  15 
 There  should  also  be  a  focus  on  measuring  the  progress  of  enabling  high-quality  data  sharing, 
 which  in  itself  is  a  non-trivial  challenge.  There  will  be  cultural  issues,  liability  concerns,  political 
 and  jurisdictional  boundaries  that  come  into  play  that  might  impede  creating  and  sharing  data. 
 There  needs  to  be  more  recognition  of  the  progress  of  such  work,  including  the  measurement  of 
 high  quality  data  sharing  and  quantifying  the  friction  of  data  flow,  metrics  of  unlearning  data,  as 
 well as data quality parameters such as data diversity and uniqueness, accessibility, etc. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 2    |  16 
 SCAI QUESTION 3    |  17 
 SCAI  QUESTION 3 

 SCAI QUESTION 3 
 GOVERNANCE STRUCTURE & 
 REGULATORY MEA SURES
 
 What  are  optimal  gov
ernance  structures  and  regulatory  measures  for 
 AI? 
 Context & Assumptions 
 Governance  and  regulation  have  a  key  role  to  play  in  shaping  the  direction  of  development  and 
 deployment  across  the  spectrum  of  different  AI  technologies  and  applications.  Governments 
 should  provide  the  conditions  for  confident  AI  innovation  and  adoption  and  for  preventing 
 AI-related  harms.  The  purpose  of  public  policy  is  to  protect  the  public  interest,  and  we  assume 
 that  effective,  ecient  and  legitimate  governance  and  regulation  is  a  precondition  for  trusted 
 and  sustainable  advancement  of  AI,  rather  than  representing  an  obstacle  to  innovation.  At  the 
 same  time,  the  multi-faceted,  complex,  and  border-crossing  nature  of  AI  raises  challenges  for 
 achieving effective, e cient, and legitimate governance structures and regulatory measures. 
 Success  requires   coherence  and  integration  of  governance  structures  and  regulatory  measures 
 across  multiple  dimensions.  These  should  ideally  encompass  different  sectors  and  regulatory 
 remits,  various  points  of  intervention  across  the  AI  lifecycle  and  AI  value  chains,  different  types  of 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 3    |  18 
 regulatory  tools,  and  several  levels  of  governance  (municipal,  national,  regional,  and 
 international/global). 
 In  order  to  achieve  this  integration,  and  to  do  so  with  legitimacy,  it  is  important  that  processes  of 
 designing  governance  structures  and  regulatory  measures  are  inclusive  of  perspectives  from  the 
 diverse  groups  in  society  whose  interests  are  at  stake  and  whose  expertise  can  help  identify 
 solutions, including users, businesses, policymakers, civil society, and academia. 
 Legitimacy  also  depends  on  ensuring  checks  and  balances,  enforcement  and  accountability 
 mechanisms, as well as the prevention of abuses of power by those who deploy and use AI. 
 Question 
 What are optimal governance structures and regulatory measures for AI? 
 Answering  this  question  involves  determining  the  appropriate  combination  of  structures  and 
 measures along several dimensions: 
 ●  Legitimacy  :  Which  parties  have  legitimacy  to  establish  governance  mechanisms  in  a 
 given  jurisdictional  context  (government  vs.  non-government  actors;  collective  or 
 representative organisations)? 
 ●  Sectors  and  regulatory  remits  :  What  is  the  optimal  combination  of  governance  and 
 regulation  designed  to  apply  to  AI  horizontally  vs.  in  specific  sectors  vs.  in  specific  use 
 cases/applications? 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 3    |  19 
 ●  Types  of  governance  tools  :  What  is  the  optimal  combination  of  different  regulatory 
 levers,  such  as  legislation,  mandatory  codes  of  practice,  voluntary  frameworks,  standards 
 and  principles,  public  procurement  rules,  and  other  tools  which  can  be  used  to  achieve 
 desired practices and outcomes? 
 ●  Points  of  intervention  :  What  is  the  optimal  combination  of  structures  and  measures  that 
 target  different  stages  in  the  AI  lifecycle,  different  segments  in  the  AI  value  chain,  and 
 should  these  be  implemented  as  ex  ante  requirements  (e.g.  as  preconditions  for  licensing 
 or  regulatory  approval)  or  as  ex  post  actions  (e.g.  requirements  to  take  remedial  actions 
 after the fact)? 
 ●  Levels  of  governance  :  What  is  the  optimal  combination  of  municipal,  national,  regional, 
 and  international/global  measures?   To  what  extent  is  international/global  alignment  or 
 harmonisation  desirable  and  achievable  (as  opposed  to  instances  of  divergence  that  are 
 unavoidable  due  to  fundamental  differences  in  political  ideologies  and  values,  and  respect 
 for state sovereignty)? 
 ●  Accountability  :  What  do  enforcement  and  accountability  look  like?  Who  should  be 
 responsible  for  enforcement,  and  what  mechanisms  are  appropriate  (e.g.  criminal 
 sanctions, regulatory penalties, access blocks and bans)? 
 Indicators of Progress 
 There  are  some  significant  challenges  in  determining  the  optimal  governance  or  regulatory 
 approach  for  AI  development  and  use.  These  include  a  lack  of  transparency  and  access  to 
 information  on  how  AI  models  are  developed,  governed,  and  deployed,  especially  in  the  private 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 3    |  20 
 sector;  a  failure  of  coordination  between  various  agencies  or  departments  in  governments  or 
 organisations  involved  in  exercising  governance  or  regulatory  functions  or  processes;  a  lack  of 
 stakeholder  inclusion  in  developing  go vernance  and  regulatory  frameworks;  a  lack  of  clear  and 
 effective  r
eporting,  enforcement,  and  accountability  mechanisms;  abuse  of  or  misalignment 
 with  incentive
s  ,  whether  at  the  political  or  commercial  level;  insucient  resourcing, 
 capabilities,  and  expertise  within  governments  or  organisations  in  designing  and  implementing 
 governance  structures  or  regulatory  measures,  and  in  ensuring  compliance;  and  regulatory  lags 
 given the ever-changing and fast-moving nature of AI. 
 While  these  are  not  insurmountable,  careful  attention  must  be  paid  to  addressing  and  overcoming 
 these  challenges,  in  order  to  prevent  them  from  becoming  key  stumbling  blocks  to  progress  in 
 answering the question. 
 Possible strategies to make progress on this question include: 
 ●  Systematic  mapping,  gap  analysis  and  guidance  for  relevant  existing  laws  and  legal 
 principles,  and  governance  structures  and  regulatory  measures  in  individual  and  regional 
 jurisdictions. 
 ●  Adopting new laws and regulations to fill the gaps that existing law does not cover. 
 ●  Establishing  new  forms  of  collaboration  and  coordination  between  regulatory  bodies 
 across different sectors, remits and jurisdictions. 
 ●  Systematic  approaches  to  developing  score  cards  and  evaluation  frameworks  for 
 governments and companies. 
 ●  Designing reporting, liability and accountability schemes. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 3    |  21 
 ●  Advancing  development  of  international  standards  that  enable  interoperability  between 
 regulatory  requirements  and  governance  frameworks  established  in  different 
 jurisdictions. 
 Measures of progress in answering the question include: 
 ●  Emergence  of  clear,  effective,  ecient,  and  legitimate  governance  frameworks  for  the 
 use  of  AI,  with  appropriate  scrutiny  (including  by  the  media),  accountability,  and  checks 
 and balances on the use of AI. 
 ●  Implementation  of  measures  specifically  aimed  at  protecting  the  public  interest  in  the 
 context  of  AI  development  and  deployment,  including  product  safety,  citizen  rights,  and 
 rules relating to the use of AI in providing public services. 
 ●  Whole-of-government  and  cohesive  approaches  in  policymaking  and  regulation,  involving 
 all relevant government departments and regulatory agencies. 
 ●  Increased  transparency  of  and  access  to  information  on  how  private  sector  models  are 
 developed  and  their  governance  and  implementation  frameworks,  as  well  as  on 
 governmental development and use of AI. 
 ●  More  countries  adopting  explicit  and  clear  positions  on  AI  governance  and  regulation, 
 whether  by  explaining  how  existing  law  and  regulations  apply  to  AI,  and/or  by  introducing 
 new laws and regulations for AI. 
 ●  Increased,  and  new  forms  of,  international  collaboration,  cooperation,  and 
 capacity-building  on  AI  regulation,  standards,  and  other  governance  frameworks, 
 mechanisms and tools. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 3    |  22 
 ●  Development  and  promulgation  of  international  standards  that  can  enable  interoperability 
 between frameworks and approaches in different jurisdictions. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 3    |  23 
 SCAI QUESTION 4    |  24 
 SCAI  QUESTION 4 

 SCAI QUESTION 4 
 SOLVING SCIENTIFIC PROBLEMS 
 How  should  we  advance  AI  to  solve  scientific  problems  that  are  critical 
 and beneficial to humanity as a whole? 
 Context and Assumptions 
 Throughout  human  history,  significant  advances  have  been  made  possible  only  by  scientific 
 progress.  Scientific  discoveries  such  as  electricity,  penicillin  and  semiconductors  that  have 
 played key roles in progress have been driven by scientific discoveries. 
 Harnessing  the  power  of  AI  systems  offers  a  promising  avenue  for  addressing  some  of  society’s 
 most  dicult  problems  that  remain  unsolved.  These  intractable  problems  which  significantly 
 impact  societal  and  individual  longevity,  are  ultimately  solvable  but  have  persisted  for  decades, 
 such  as  those  involving  climate  change  and  complex  biological  systems.  AI  holds  the  potential  to 
 bring  innovative  self-directed  approaches  that  will  assist  humanity  in  fundamental  ways.  For 
 example,  in  climate  change,  advances  in  carbon  sequestration  will  enable  humanity  to  deal  with 
 the  impact  of  global  warming  due  to  fossil  fuels,  and  in  complex  biological  systems,  advances  in 
 genome  sequencing  and  editing  could  help  us  understand  human  biology  and  develop  cures  for 
 disease. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 4    |  25 
 There  are,  however,  some  challenges.  There  is  a  general  and  systemic  lack  of  integration  of 
 foundation  models  and  exploratory  methods  that  generate  and  examine  new  hypotheses.  There  is 
 also  a  need  to  develop  logical  and  causal  inference  mechanisms  in  AI  neural  systems,  as  well  as 
 adequate funding for the specialised infrastructure required for AI development. 
 Question 
 How  should  we  advance  AI  to  solve  scientific  problems  that   are  critical  and  beneficial  to 
 humanity as a whole? 
 In  considering  this  question,  we  also  need  to  think  about  how  scientists  and  researchers  across 
 the  world  can  come  together  to  harness  AI  and  prioritise  resources  in  applying  these  powerful 
 systems to scientific problems. 
 Indicators of Progress 
 In  the  short-term,  a  notable  indicator  of  progress  includes  the  growing  evidence  of  increasing 
 cross-disciplinary  and  cross-border  collaboration   and  cooperation.  Addressing  traditional 
 boundaries  between  different  scientific  and  industrial  areas  can  play  a  significant  role  in 
 facilitating  access  to  a  diverse  and  extensive  range  of  scientific  and  mathematical  corpus.  In  the 
 longer  term,  international  research  collaboration,  supported  by  multilateral  funding,  has  the 
 potential  to  yield  results  that  benefit  the  global  community.  Importantly,  this  would  allow  us  to 
 integrate  scientific  knowledge  into  AI  models  to  increase  their  reliability  and  accuracy,  while 
 using less computation. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 4    |  26 
 A  key  indicator  of  progress  will  be  the  emergence  of  theorem-solving  AI  systems  that  are 
 applicable  to  a  wide  variety  of  scientific  areas  of  interest,  and  allow  us  to  draw  on 
 cross-disciplinary  scientific  knowledge  at  a  scale  previously  unavailable  but  with  the  potential  to 
 transform  scientific  development.  We  look  forward  to  a  time  when  we  are  able  to  see  high  impact 
 scientific  papers  being  written  by  such  AI  systems  with  application  in  areas  such  as  carbon 
 sequestration,  seasonal  climate  prediction,  deciphering  the  human  ageing  process  and  new 
 materials design. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 4    |  27 
 SCAI QUESTION 5    |  28 
 SCAI  QUESTION 5 

 SCAI QUESTION 5 
 MODELS & ARCHITECTURE 
 DERIVED FROM 
 NATURAL INTELLIGENCE 
 How  do  we  leverage  developmental  models  and  architecture  derived 
 from natural intelligence to create new paradigms of AI? 
 Context & Assumptions 
 Human  intelligence  and  cognition  have  capabilities  and  performance  characteristics  that  are 
 currently  unmatched  by  the  best  AI  systems.  While  some  AI  systems  outperform  specific  human 
 capabilities,  they  fall  short  in  generalisation  capabilities  and  learning  eciency.  Natural 
 intelligence  is  far  more  flexible,   adaptive,  responsive,  and  energy  ecient.  The  brain  and  our 
 understanding  of  its  functional  and  cognitive  architecture  offers  a  possible  reference  to 
 implement an intelligence with these performance characteristics. 
 The  functional  and  developmental  organisation  of  natural  intelligence  has  a  considerable  impact 
 on  how  intelligent  capabilities  form  and  perform.  Hardware  substrates  in  the  brain  (e.g.,  neurons) 
 differ  from  the  artificial  hardware  substrates  (e.g.,  processing   units).  For  instance,  the 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 5    |  29 
 hierarchical  structure  of  spatial  reasoning  in  the  brain,  between  cortical  grid  cells  and 
 hippocampal  place  cells,  provides  considerable  computational  eciency  and  robustness  that 
 robots  are  only  now  beginning  to  match.  In  addition,  the  ability  of  the  brain  to  perform  in-memory 
 computation  provides  significant  eciencies  that  cannot  be  matched  by  the  current  separation 
 between  computation  and  memory  in  existing  CPUs  and  GPUs.  There  is  also  evidence  that  the 
 cognitive  architecture  of  the  brain  is  genetically  encoded  at  birth  -  the  core  knowledge 
 hypothesis  suggests  a  strong  prior  over  concepts  such  as  places,  objects  and  motor  skill  which  is 
 already present in biology. 
 The  increased  understanding  of  the  functional  architecture  and  performance  of  the  brain 
 provided  by  cognitive   neuroscience  can  give  us  new  paradigms  to  develop  more  capable  forms  of 
 artificial  intelligence.  The  cognitive  sciences  (neuroscience,  psychology,  linguistics,  philosophy  of 
 mind,  anthropology  and  artificial  intelligence)  study  different  aspects  of  natural  intelligence  that 
 can  be  used  to  inform  the  design  of  more  capable  AI  systems.  Conversely,  progress  in  AI  creates 
 opportunities for understanding brain functions, human cognition, psychology, and development. 
 Question 
 How  do  we  leverage  developmental  models  and  architectures  derived  from  natural  intelligence 
 to create new paradigms of AI? 
 Answering this core question is tightly coupled to the following additional questions: 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 5    |  30 
 Evaluation questions 
 ●  Which  aspects  of  natural  intelligence  cannot  currently  be  replicated  by  existing  AI 
 approaches?  This  is  a  moving  target,  but  it  is  crucial  to  understand  precisely  how  natural 
 intelligence outperforms existing artificial intelligent systems, and at which tasks. 
 Structural questions 
 ●  What  is  the  right  functional  decomposition  of  intelligence  that  enables  these  levels  of 
 performance  and  capabilities?  The  functional  relations  implicitly  define  a  structure  that 
 may be reflected in the structure of the brain. 
 ●  What  are  the  intermediate  hierarchical  structures  in  the  brain  that  organise  neurons  into 
 functional  reasoning  and  cognition?  While  cognition  and  intelligence  do  not  need  to  be 
 implemented  by  neurons,  there  are  existing  models  of  artificial  intelligence  represented 
 using  spiking  neural  models.  An  additional  intermediate  hierarchical  structure  is  required 
 to  organise  artificial  spiking  neural  models  into  purposeful  computation  to  allow  program 
 synthesis. 
 ●  The  functional  decomposition  and  cognitive  architecture  of  natural  intelligence  imply 
 specific  and  powered  inductive  biases.  What  are  these  inductive  biases  inherent  in 
 natural cognitive architectures? 
 ●  How  can  biological  models  of  motor  skills  be  acquired  and  composed  by  artificial 
 intelligence?  It  is  clear  that  natural  intelligence  can  acquire  low-level  motor  skills 
 eciently,  and  incorporate  these  skills  as  concepts  into  higher-level  reasoning.  There  is 
 further  evidence  from  evolutionary  biology  that  developing  these  low-level  skills  took 
 considerably  more  evolutionary  time  than  higher-level  reasoning,  and  may  be  considered 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 5    |  31 
 the true cognitive substrate of intelligence. 
 Performance questions 
 ●  Can  computational  architectures  inspired  by  the  cognitive  architectures  of  the  brain 
 change,  adapt  and  evolve  as  easily  as  the  brain  does?  The  power  consumed  by  in-silico 
 intelligence  dwarfs  the  power  consumed  by  biological  intelligence,  but  with  a  fraction  of 
 the performance. 
 ●  Can  computational  architectures  inspired  by  the  cognitive  architectures  of  the  brain 
 match  the  energy  eciency  of  the  brain?  There  is  considerable  evidence  that  when  a 
 biological  agent  encounters  new  scenarios,  it  is  quickly  able  to  adapt  to  the  scenario  by 
 reusing previous experience. 
 ●  Do  the  cognitive  architectures  of  the  brain  implicitly  encode  an  inductive  bias  that  is 
 aligned with human values and judgement? 
 ●  How  can  we  ensure  and  maintain  alignment  between  artificially  intelligent  agents  and 
 humanity? 
 Indicators of Progress 
 We expect that the best approaches to answering these questions will include: 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 5    |  32 
 ●  Leveraging  development  models  of  natural  intelligence  and  insights  from  neuroscience 
 and  cognitive  science  and  implementing  these  models  in  architectures  inspired  by  them. 
 We  expect  that  one  candidate  form  of  these  models  and  architectures  will  be  hierarchical 
 compositional probabilistic models that can be reused. 
 ●  We  also  expect  that  one  candidate  form  of  these  models  that  allows  the  structure  of  the 
 architecture  to  be  learned  and  to  be  adapted  from,  includes  neurosymbolic 
 representations. 
 ●  We  will  also  require  new  theories  of  model  integration  (which  is  not  the  same  as 
 interoperability)  and  techniques  for  learning  to  be  used  for  integrating  component 
 models. 
 ●  Designing  neuromorphic  AI  software  and  hardware,  and  examining  the  benefits  that  can 
 be gained by neuromorphic and neurosymbolic approaches. 
 We expect that the best approaches to measuring progress in answering these questions will be: 
 ●  Benchmarking  brain-inspired  architectures  and  cognitive  systems  on  tasks  relative  to 
 human/natural performance. 
 ●  Benchmarking on task specialisation, generalisation and few-shot learning. 
 ●  Benchmarking  developmental  models,  that  allow  the  cognitive  architecture  to  develop 
 and  adapt  over  time.  For  example,  tasks  that  have  an  internal  hierarchy  with  different 
 levels  of  abstraction  that  are  currently  hand-specified  (e.g.,  perception  systems  driving 
 task-and-motion  planning  systems)  can  be  derived  automatically  using  developmental 
 models derived from natural intelligence. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 5    |  33 
 SCAI QUESTION 6    |  34 
 SCAI  QUESTION 6 

 SCAI QUESTION 6 
 VALUES & NORMS TO ALIGN AI: 
 ELICITATION & IMPLEMENTATION 
 How  do  we  elicit  the  values  and  norms  to  which  we  wish  to  align  AI 
 systems, and implement them? 
 Context & Assumptions 
 Increasingly  capable  AI  systems  are  being  used  to  perform  more  complex  sequences  of  actions 
 without  human  supervision.  We  collectively  need  to  know  how  we  want  them  to  behave  and  how 
 to  ensure  they  do  so.  This  has  historically  been  described  as  “the  alignment  problem”.  However, 
 the  aim  of  aligning  systems  to  “user  intent”  or  to  “human  values”  is  a  double-edged  sword.  Users 
 might  have  malicious  intents;  humans  can  have  abhorrent  values.  In  addition,  and  not 
 coincidentally,  the  project  of  AI  alignment  has  been  pursued  in  a  narrowly  technical  way,  without 
 drawing  enough  on  broader  expertise   (e.g.,  from  the  social  sciences  and  humanities),  even  as 
 other  areas  of  responsible  AI  have  done  more  to  integrate  their  research  with  other  fields.  There 
 is  an  urgent  need  to  develop  an  agenda  for  AI  alignment  that  draws  on  this  broader  understanding 
 to ensure that AI systems behave appropriately. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 6    |  35 
 Question 
 How  do  we  elicit  the  values  and  norms  to  which  we  wish  to  align  AI  systems,  and  how  do  we 
 implement those values and norms? 
 Indicators of Progress 
 Eliciting  the  values  and  norms  to  which  we  wish  to  align  AI  systems  is  not  a  novel  problem.  It  is 
 simply  the  challenge  of  reaching  a  collective  decision  on  matters  of  common  concern.  The  first 
 stage  is  to  provide  the  theoretical  and  empirical  resources  for  public  debate  and  individual 
 decision-making: 
 ●  We  need  well-grounded  research  anticipating  potential  societal  impacts  of  more  capable 
 AI  systems.  Social  scientists  and  computer  scientists  should  collaborate  to  explore 
 different  possible  futures  for  AI,  and  to  learn  from  the  rich  experience  with  previously 
 deployed systems to anticipate likely risks of future systems. 
 ●  We  need  clear  articulations  of  familiar  normative  considerations  that  those  potential 
 impacts  raise.  For  example,  most  societies  already  have  clear,  albeit  disputed,  views  on 
 values  like  discrimination,  accountability,  and  transparency.  The  goal  then  is  to  apply  and 
 refine those values for this particular application. 
 ●  We  need  a  theoretical  approach  to  unfamiliar  normative  considerations  raised  by  those 
 potential  impacts.  Some  questions  raised  by  more  capable  AI  systems  will  not  come  with 
 ready-made  answers.  For  example,  if  A  delegates  an  action  to  B,  then  do  the  reasons  that 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 6    |  36 
 apply  to  B  when  B  acts  depend  on  whether  B  is  a  human  or  an  AI  agent?  Which  behaviours 
 are  acceptable  or  unacceptable  from  AI  agents?  When  interacting  with  humans,  one  set 
 of  rules  might  apply,  but  which  rules  will  apply  in  multi-agent  situations?  Or,  can 
 extremely  capable  AI  systems  ever  be  consistent  with  democratic  government?  More 
 generally,  should  societies  even  be  pursuing  the  goal  of  AI  capability  beyond  a  certain 
 threshold? 
 The  second  step  is  to  use  our  existing  resources  for  collective  decision-making  and  resolving 
 moral  disagreement.  This  means  recognising  at  least  three  distinct  layers  of  normative  guidance, 
 with  different  collectives  being  appropriate  to  decide  on  different  layers.  As  with  most  other 
 societal  decision-making,   this  will  involve  some  “constitutional”  norms  that  are  relatively  settled, 
 and others that should be regularly revisited and revised: 
 ●  Some  minimal  norms  should  be  decided  at  the  global  level,  in  the  same  way  as  the  global 
 community  decides  on  certain  basic  human  rights.  Which  highly  capable  AI  systems 
 should  nobody  be  able  to  produce?  What  are  the  very  minimum  expectations  for  the 
 behaviour of AI systems, on which the whole world can decide? 
 ●  More  substantive  norms  should  be  decided  at  the  level  of  nation-states  or  other 
 sub-global  political  units  (e.g.,  the  EU).  By  analogy,  while  all  states  in  principle  arm  the 
 same  basic  human  rights,  they  all  have  different  approaches  to  civil  and  political  rights. 
 Operative  questions:  are  there  any  AI  systems  that  we  want  no  one  to  produce?  What   are 
 the  minimum  expectations  for  the  behaviour  of  AI  systems,  on  which  we  as  a  political 
 community can decide? 
 ●  Remaining  norms  can  be  the  object  of  individual,  or  (sub-state)  collective  choice, 
 including  by  companies.  Operative  questions:  given  the  constraints  described  in  A  and  B, 
 which  AI  systems  do  we  want  to  produce?  Which  behaviour  (not  just  minimum 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 6    |  37 
 expectations)  do  we  want  to  see  in  our  AI  systems?  Analogy:  virtue  in  a  person.  If  all  you 
 ever  did  was  violate  nobody’s  human  rights  and  not  break  the  law,  then  that  would  not  on 
 its  own  speak  well  of  you  as  a  person.  You  also  might  aim  to  be  honest,  loyal,  loving, 
 conscientious,  etc.  Those  who  build  AI  systems  that  can  act  (in  effect)  autonomously 
 should want to do more than the bare minimum. 
 Our  means  for  resolving  moral  disagreement  and  making  collective  decisions  are  often 
 compromised,  and  direct  action  or  institutional  innovation  may  be  needed.  AI  itself  may  help 
 unblock  decision-making,  for  example,  by  supporting  participatory  or  deliberative  democratic 
 processes  (especially  when  deciding  on  values  beyond  basic  human  rights  and  legal   compliance). 
 But we must avoid using technology to replace politics instead of augmenting it. 
 How  do  we  implement  these  values  and  norms?  We  recommend  sociotechnical  methods  that 
 complement  technical  methods  in  computer  science  with  expertise  from  the  social  sciences  and 
 humanities. 
 All  AI  systems  will  be  deployed  by  people  in  a  social  and  political  environment.  “Aligning”  this 
 sociotechnical  system  can  be  achieved  through  interventions  on  each  of  these  elements,  e.g., 
 placing  models  in  more  complex  systems  that  mitigate  some  of  their  risks;  training  users  to  avoid 
 automation  bias;  thinking  about  institutions:  for  example,  can  we  (should  we)  reshape  political 
 institutions and AI systems so that AGI, if achieved, does not directly undermine democracy? 
 We  also  need  pre-deployment  sociotechnical  evaluations  that  consider  the  harms  caused  in 
 actual  use,  rather  than  in  isolation  from  broader  social  systems.  And  we  need  to  advance 
 adversarial  testing  (red-teaming)  beyond  simply  querying  the  model  to  elicit  naughty  text, 
 developing  instead  complex  multi-agent  simulations  that  test  for  dangerous  capabilities.  This 
 may entail some “gain of function” AI research, which may require methodological innovation. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 6    |  38 
 Ultimately,  however,  we  want  to  produce  systems  that  are  designed  to  behave  appropriately 
 (given  the  three  stages  of  norms  described  above),  and  can  be  counted  on  to  do  so  (preferably 
 provably).  Designing  AI  systems  that  will  implement  these  values  directly  is  therefore  essential. 
 Collaboration  between  computer  scientists  and  other  fields  will  provide  fresh  perspectives  on 
 alignment methods, and suggest new research directions. 
 Any  method  that  applies  a  thin  fine-tuning  adjustment  over  the  top  of  a  pretrained  model  is 
 unlikely  to  be  robust  to  adversarial  attacks  of  different  kinds.  Also,  in  learning  from  human 
 feedback  (e.g.,  Reinforcement  Learning  for  Human  Feedback  (RLHF))  for  language  models,  the 
 behaviour  being  evaluated  is  identical  to  the  behaviour  being  shaped;  but  if  LLMs  are  used  as  the 
 executive  control  centre  for  more  complex  systems  (i.e.,  agents),  then  the  behaviours  that  we 
 want  to  shape  will  be  actions  in  the  world,  not  just  prompt  completions.  We  should  not  expect 
 learning  from  human  feedback,  such  as  RLHF,  to  work  well  in  such  cases  and  the  costs  of 
 inadequate  alignment  are  likely  to  be  greater.  So,  while  learning  from  human  feedback,  such  as 
 RLHF  or  even  Reinforcement  Learning  with  AI  Feedback  (RLAIF),  constitutes  significant  research 
 achievements  that  are  worth  building  on,  we  should  also  pursue  other  approaches  to  value 
 implementation,  through  collaborative  investigation  drawing  on  different  fields.  These  may 
 include  data  curation  and  model  unlearning,  and  implementing  values  in  pre-training.  We 
 encourage   exploration  of  how  language  models’  competence  with  moral  concepts  can  be 
 operationalised  to  support  more  generalisable  moral  reasoning.  High-level  reasoning  and 
 planning  capabilities  are  important  constituents  of  responsible  moral  agency.  Only  agents  that 
 can  plan  can  be  consistent.  Only  agents  with  high-level  reasoning  capabilities  can  make  complex 
 value  tradeoffs.  So,  while  enhanced  model  capabilities  will  increase  risk,  they  might  also  increase 
 resources for successful value implementation. 
 Some  obvious  obstacles  threaten  progress  in  value  implementation.  Central  challenges  in  model 
 alignment  include:  Reward  hacking,  reward  tampering,  and  specification  gaming;  the  problem  of 
 supervising  systems  that  are  substantially  more  capable  than  humans;  deceptive  alignment,  i.e., 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 6    |  39 
 the  possibility  that  models  might  appear  to  conform  to  intended  values  in  training  but  depart  from 
 that in use. 
 Interdisciplinary  collaboration  faces  obvious  cold  start  problems  widely  discussed  elsewhere  1  .  In 
 AI  research,  more  resources  are  spent  on  advancing  capabilities  than  on  alignment,  and  funding 
 for  technical  alignment  dwarfs  funding  for  sociotechnical  work.  AI  companies  have  few  social 
 scientists and collaborate too infrequently with academia. 
 Progress is clearly possible, and will be marked by the following: 
 ●  Compelling answers to novel normative questions raised by advanced AI systems. 
 ●  Public and political education on the impacts of AI systems on more familiar values. 
 ●  Substantive political debate at international and domestic level over the future of AI. 
 ●  Innovation in participatory design by AI labs. 
 ●  More performant approaches to value implementation drawing on multiple fields. 
 ●  Better evaluations incorporating multidisciplinary approaches. 
 ●  Robust  criteria  for  determining  when  value  implementation  has  failed  and  AI  systems  are 
 too unsafe to release. 
 1 
 https://nap.nationalacademies.org/ catalog/26507 /fostering-responsible-computing-research-foundations-and-practi 
 ces 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 6    |  40 
 SCAI QUESTION 7    |  4 1 
 SCAI  QUESTION 7 

 SCAI QUESTION 7 
 EQUITABLE ACCESS, 
 CONTROL & FAIR COMPETITION 
 Where  in  the  AI  ecosystem  should  we  ensure  equitable  access,  control 
 and fair competition? How should we address these concerns? 
 Context & Assumptions 
 Recent  developments  in  AI  have  demonstrated  tremendous  potential  to  have  both  positive  and 
 negative  impacts  on  society,  organisations  and  individuals.  AI  systems  currently  rely  upon  large 
 compute,  advanced  models,  and  extensive  training  data.  These  capabilities  are  inequitably 
 distributed  and  result  in  a  concentration  of  power.  This  in  turn  confers  agency  on  specific  actors, 
 who  may  have  goals  that  are  misaligned  with  broader  societal  objectives.  Areas  of  misalignment 
 include  an  adequate  recognition  of  risks,  creating  systems  which  are  safe,  and  using  AI  to  achieve 
 social  benefit  or  public  good,  rather  than  in  support  of  profit  maximisation.  On  the  last,  we  note 
 that  the  boundary  between  commercialization  and  basic  research  is  not  distinct.  One  of  the 
 assumptions  driving  research  work,  which  may  not  be  well  founded,  is  that  research  can  feed  into 
 a  product  that  will  in  turn  generate  revenue  and  other  resources  to  feed  back  into  research. 
 However,  the  majority  of  this  research  takes  place  in  proprietary  labs  in  companies  that  ship 
 products and have profit as their central motivation. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 7    |  42 

 Question 
 Where  in  the  AI  ecosystem  should  we  ensure  equitable  access,  control  and  fair  competition? 
 How should we addre s
s these concerns? 
 Concerns  with  concentration  of  power  apply  to  a  range  of  issues,  including  but  not  limited  to 
 price,  quality,  volatility,  restrictions  in  access  (with  particular  attention  to  which  communities 
 might  be  marginalised),  restrictions  in  developing  capability  (e.g.,  training  or  hardware 
 limitations)  and  ability  to  shape  outcomes/output.  There  are  also  emerging  externalities  (e.g., 
 situations  of  great  individual  benefit  which  result  in  collective  harm)  which  may  be  exacerbated 
 by  such  concentrations.  We  also  realise  there  are  unrecognised  benefits  to  more  open  access  to 
 both  the  resources  required  to  build  AI  models  as  well  as  to  AI  models  themselves.  These  include 
 broader  economic  growth  across  sectors,  and  broader  perspectives  in  building  and  deployment. 
 At  the  same  time,  there  are  some  things  that  should  remain  closed  or  not  be  broadly  accessible, 
 e.g., PII and sensitive information, healthcare data. 
 Indicators of Progress 
 Given  these  considerations,  we  recommend  two  pathways  to  mitigate  the  effects  of  such 
 concentration  of  power.  First,  develop  a  more  democratic  system  that  enables  broader  access  to 
 the  key  resources  necessary  to  develop  these  technologies  (i.e.,  lower  barriers  to  entry  for  new 
 entrants).  This  allows  a  wider  range  of  actors  of  varying  size  and  capability  to  the  field,  preventing 
 or  slowing  concentration  of  power.  Second,  develop  regulatory  and  non-regulatory  strategies  for 
 the  reduction  of  the  harm  in  situations  of  power  concentration.  Non-regulatory  strategies  could 
 include  encouraging  norms  that  support  desired  behaviours,  such  as  transparency  in 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 7    |  43 
 model-building  to  manage  risk.  These  strategies  are  not  unfamiliar.  They  have  been  used  in  the 
 past  to  regulate  other  industries  with  similar  potential  impact,  such  as  public  utilities  and 
 telecommunications.  For  both  pathways,  potential  areas  for  intervention,  or  chokepoints,  include 
 compute  asymmetries  (e.g.,  compute  fabs,  chip  architecture,  and  optimisation  for  specific 
 labs/models  by  chipmakers),  and  datasets  (e.g.,   lack  of  representative  datasets  in 
 underrepresented languages). 
 Some  challenges  to  operationalising  these  strategies  might  be  deeply  held  ideological  beliefs 
 about  how  the  market  should  be  structured,  and  risk  tensions  and  tradeoffs  (e.g.,  limited  vs 
 broader  perspectives,  more  vs  less  control).  Evolving  use  and  emerging  threats  also  require 
 nimbleness in adapting regulation. 
 Indicators of success would include: 
 ●  Emergence  of  a  mix  of  independent  AI  providers  at  different  scales  throughout  the  market 
 i.e. both small and large firms. 
 ●  Large firms are well-regulated to minimise negative impact. 
 ●  Regulation  differentiates  between  applications  based  on  implications  of  their  use  (e.g., 
 nuclear power vs nuclear weapons). 
 ●  Policy  makers  consider  a  clear  framework  when  considering  how  to  regulate  different 
 sectors/areas of the AI stack. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 7    |  44 
 SCAI QUESTION 8    |  45 
 SCAI  QUESTION 8 

 SCAI QUESTION 8 
 TRANSFORMING EDUCATION 
 How  can  we  use  AI  to  enhance  the  effectiveness,  eciency,  and 
 accessibility of education across societies around the world? 
 Context & Assumptions 
 Improvement  of  human  capital  through  education  is  foundational  across  society  and  economy. 
 There  is  great  potential  for  AI  to  enhance  the  eciency,  effectiveness  and  accessibility  across 
 the  entire  educational  ecosystem.  This  spans  students,  parents,  teacher  training,  content 
 creation,  social  interaction,  curriculum  design  and  delivery,  evaluation  and  certification  of 
 results. 
 An  AI-enhanced  education  ecosystem  can  cater  to  a  range  of  needs  across  countries  and 
 contexts.  In  resource-constrained  societies,  this  could  mean  the  difference  between  education  or 
 no  education  at  all.  In  other  cases,  knowledge  and  skills  education  could  be  enhanced  to  free  up 
 resources  to  address  the  more  “human”   aspects  of  education,  such  as  critical  thinking,  creativity, 
 empathy,  social  development  etc,  as  well  as  focusing  on  student-teacher  interactions.  New 
 models  of  human-AI  collaboration  could  be  explored,  such  as  in  using  AI  tools  to  assist  teachers  in 
 characterising  and  evaluating  the  students  and  their  progress,  so  as  to  challenge  them 
 appropriately  and  foster  learning  and  personal  growth.  Across  industry,  appropriate  interventions 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 8    |  46 
 could  improve  upskilling  and  the  adoption  of  new  tools  and  capabilities,  enhancing  productivity 
 and introducing new value. 
 This  approach  relies  on  several  assumptions.  At  the  core,  good  access  to  infrastructure  and 
 connectivity  are  key;  every  student  needs  access  to  a  mobile  phone  and  a  TV,  possibly  a  keyboard. 
 On  the  AI  side,  we  assume  the  current  state-of-the-art  for  AI,  i.e.  there  are  few  unsolved  scientific 
 AI  challenges  that  prevent  us  from  making  progress  in  this  area  and  we  can  see  payoff  and  test 
 product-market-fit from the start. 
 Question 
 How  can  we  design  and  implement  an  AI-enhanced,  open  education  ecosystem,  with  a 
 sustainable  mechanism  for  participation,  with  the  ambition  of   making  education  more 
 ecient, effective, and accessible for local communities and global society? 
 Doing  so  will  also  enable  us  to  maximise  human  capital  across  communities  around  the  world, 
 starting with maximising every student’s potential. 
 Indicators of Progress 
 Given  the  complexity  of  education  and  the  contextual  needs  of  communities  and  corporations,  a 
 systems  approach  is  necessary.  This  recognises  that  opportunities  and  incentives  differ  across 
 life  phases–elementary  school,  tertiary  education,  workforce  reskilling,  and  more.  Moreover,  we 
 need  to  address  all  elements  of  the  value  chain:  students,  teachers,  curriculum  optimization  and 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 8    |  47 
 delivery,  content  creation,  evaluation,  certification,  accessibility  and  social  interactions.  The 
 following are non-exhaustive illustrations: 
 ●  Students  can  benefit  from  having  an  always-available  tutor,  and  benefit  from  immediate 
 feedback and personalised, adaptive content. 
 ●  Parents  care  about  supporting  and  monitoring  their  childrens’  progress.  AI  can  help  them 
 support them, such as in pointing out areas of improvement and suggest actions. 
 ●  Teachers  will  benefit  from  spending  less  time  on  routine  tasks  such  as  grading,  and  can 
 devote more time to foster creativity among students. 
 ●  Content  capture  and  creation  is  time  consuming,  thus  a  major  hurdle  for  educators  to 
 share courses. AI can be used to automatically transcribe and generate material. 
 ●  Curriculum  design  and  optimization  currently  relies  on  the  experience  of  lecturers, 
 drawn  from  hundreds  of  students.  A  data-driven  approach  based  on  100x  more  student 
 expereiences can tailor curriculum for better learning, customised to individual students. 
 ●  Evaluation  involves  certification  (of  achieved  degree),  challenge  (for  progress)  and  control 
 (of  education  objectives).  Game  mechanisms  combined  with  empirical  data  can  motivate 
 students through suitably challenging problems and tests. 
 ●  Accessibility  tools  are  typically  costly  (subtitles,  speed,  visual  aids);  these  can  be  added 
 through AI tools at minimal cost. 
 ●  Student  interaction  can  be  enhanced  by  social  recommendations  (student  /  tutor 
 pairings, peer groups) and content moderation. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 8    |  48 
 There  is  significant  opportunity  for  productivity  gains.  These  dividends  can  be  used  to  increase 
 the  number  of  students  trained,  and/or  the  quality  of  their  education,  for  a  more  cost-effective 
 education system. 
 To achieve this, we must overcome several challenges, including: 
 ●  Overcoming accuracy issues such as hallucinations. 
 ●  Integration with existing infrastructure such as Google classroom. 
 ●  Interoperable,  universal  software  interfaces  that  allow  for  whole  system  optimisation 
 (future-proofing  vs.  ease  of  use  vs.  adoption),  and  that  allows  for  meaningful  auditability 
 by humans; text-based interfaces are likely most suitable. 
 ●  Educator capabilities in using AI in teaching. 
 ●  Contextually-aligned curriculum and delivery depending on cultural norms and needs. 
 ●  Having  strong  enough  economic  incentives  to  facilitate  diverse  offerings  and 
 interoperability in an AI-enhanced education ecosystem. 
 ●  Regulatory  and  commercial  acceptance  of  certificates  obtained  by  AI-enhanced 
 education (in particular for self-study and non-traditional pathways). 
 ●  Social  acceptance  of  the  notion  of  AI-enhanced  education  itself,  by  stakeholders  such  as 
 educators, parents, students, companies, regulators. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 8    |  49 
 SCAI QUESTION 9    |  50 
 SCAI  QUESTION 9 

 SCAI QUESTION 9 
 MITIGATING CATASTROPHIC RISKS & 
 ONGOING HARMS 
 How  can  we  mitiga
te  the  catastrophic  risks  and  ongoing  harms  arising 
 from  AI,  recognising  that  there  are  diverse  opinions  on  the  severity, 
 probability,  time  sensitivity,  and  recoverability  of  these  risks  and 
 harms? 
 Context & Assumptions 
 We  recognise  there  is  a  range  of  views  on  what  are  the  risks  and  harms  that  can  arise,  and  their 
 severity,  probability,  time  sensitivity,  and  recoverability.  For  instance,  here  are  some  potential 
 risks and our estimates on their time scales: 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 9    |  51 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 9    |  52 

 Types of potential catastrophic risks 
 and harms  Time-scale of effect 
 Widespread social harm (e.g., loss of trust or 
 trustworthiness in institutions, electoral 
 dysfunction; employment challenges)  Already happening 
 AI-assisted cyber-risk  Already happening at some scale 
 Lethal autonomous weapons disasters  Already happening at some scale 
 AI-assisted bioweapons and accidents  Increasingly feasible now, and in need of 
 greater attention 
 AI-assisted nuclear command and control 
 malfunction  Some media reports suggest relevant 
 discussions between some countries 
 AI-driven economic collapse (e.g., the 2010 
 Flash Crash at a much larger scale; mass 
 unemployment)  Emerging / plausible within the next few years 
 AI-driven environmental destruction (e.g., 
 exponentially accelerated pollution or 
 resource consumption)  Plausible to begin within a decade or two 
 Risks  and  harms  from  AI  can  arise  from  various  sources.  They  may  occur  by  accident, 
 intentionally,  or  due  to  willful  indifference  by  the  different  stakeholders.  They  may  also  occur  at 
 the  systemic  level  where  no  party  is  responsible  or  accountable  when  such  risks  and  harms 
 happen  (e.g.,  widespread  irreversible  addiction  to  a  technology  that  no  one  entity  in  particular  is 
 responsible for developing). 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 9    |  53 
 An  assumption  behind  this  question  is  that  there  may  be  warning  signs,  and  it  would  be  valuable 
 to  actively  look  for  them.  Catastrophic  harms  can  also  result  even  without  Artificial  General 
 Intelligence  (AGI)  since  there  are  AI  capabilities  in  narrow  domains  that  could  already  lead  to 
 societal-scale catastrophic risks. 
 Question 
 How  can  we  mitigate  the  catastrophic  risks  and  ongoing   harms  arising  from  AI,  recognising 
 that  there  are  diverse  opinions  on  the  severity,  probability,  time  sensitivity  and  recoverability 
 of these risks and harms? 
 If  we  are  to  understand  this,  we  will  also  need  a  way  to  discuss  and  identify  which  risks  and  harms 
 are  considered  catastrophic  and  deserving  of  more  attention.  For  each  such  risk,  we  will  need  to 
 answer  the  following  questions:  what  are  its  warning  signs  (if  possible)?  Who  should  be  entrusted 
 to  monitor  for  those  signs?  On  what  time  scale  might  it  happen?  Who  decides  if  the  risk  is  worth 
 taking,  and  how?  And  if  we  fail  to  avoid  it  completely,  how  can  we  mitigate  its  effects  and  recover 
 from it, and at what cost? 
 Indicators of Progress 
 To  avoid  catastrophic  harms  from  increasingly  advanced  AI,  we  should  establish  clear  warning 
 signs  and  thresholds  in  advance  across  multiple  areas.  These  include  indicators  and  thresholds 
 pertaining  to  computing  power,  demonstrations  of  dangerous  AI  abilities,  job  loss,  lawsuits, 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 9    |  54 
 expert  testimonies  from  diverse  disciplines,  and  proliferation  of  fake  content,  impersonations, 
 and cyberattacks. 
 Comprehensive  safety  evaluation  infrastructure  and  standards  should  be  developed  for  stress 
 testing  mission  critical  systems  before  full  deployment.  Benchmarks  and  standards  should  be 
 developed  for  a  range  of  technical  and  social  considerations.  Best  practices  such  as  red-teaming 
 should also be established and scaled up. 
 Robust  oversight  mechanisms  are  also  needed,  involving  consultation  with  diverse  experts  as  well 
 as  representatives  of  potential  victim  groups.  Audits  should  be  independent  without  conflicts  of 
 interest.  And  given  the  global  impact  possible,  international  oversight  mechanisms  may  be 
 warranted for the most powerful AI systems. 
 By  defining  indicators  and  responses  in  a  systematic  way  ahead  of  time,  we  can  monitor  progress 
 and  risk  to  make  proactive  governance  decisions  before  harms  arise.  The  goal  is  to  avoid  the 
 "boiling frog" by reacting only when problems become dire and harder to address. 
 Potential  challenges  arise  in  achieving  consensus  on  what  constitutes  a  “catastrophe”,  assessing 
 the  likelihood  of  various  catastrophic  risks,  and  determining  how  far  off  they  are  on  the  horizon. 
 As  such,  this  complexity  necessitates  a  range  of  approaches,  demanding  more  people  and 
 resources than would be required to mitigate a single type of catastrophic risk. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 9    |  55 
 SCAI QUESTION 10    |  56 
 SCAI  QUESTION 10 

 SCAI QUESTION 10 
 COMBATING 
 MIS/DISINFORMATION CAMPAIGNS 
 What  are  the  appropriate  speed  bumps  and  incentives  for  content 
 channels  to  reduce  the  negative  impact  of  mis/disinformation 
 campaigns? 
 Context & Assumptions 
 Mis/disinformation  is  an  existing  problem,  but  with  the  development  of  AI,  we  will  see  an  increase 
 in  volume  and  sophistication.  This  can  be  socially  corrosive  and  degrade  shared  trust  between 
 citizens  and  institutions.  The  pervasiveness  and  velocity  of  social  media  content  distributed 
 through  content  channels  have  created  the  conditions  where  a  generation  relies  on  these 
 channels to shape their understanding of the world. 
 This  problem  is  hard  to  address  because  the  techniques  for  mis/disinformation  (especially 
 multi-step  emotional  manipulation)  are  hard  to  detect.  While  AI  tools  lower  the  cost  and  increase 
 access  for  those  seeking  to  generate  and  proliferate  disinformation,   we  are  nearing  a  point  where 
 we  lack  the  ability  to  discern  if  the  source  of  information  is  human  or  bot  and  distinguish  between 
 true and fake content. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 10    |  57 
 Question 
 What  are  the  appropriate  speed  bumps  and  incentives  for  content  channels  to  reduc e  the 
 negative impact of mis/disinformation campaigns? 
 ●  What are the technical solutions to support these speed bumps/incentives? 
 ●  What  are  the  trade  offs  between  public  safety  and  freedom  for  content  generation 
 (including the freedom to misrepresent authority of fact)? 
 ●  How  do  trusted  institutions  maintain  trust  with  citizens  in  a  world  of  increased 
 mis/disinformation? 
 Indicators of Progress 
 While  there  is  no  known  method  to  fully  solve  this  problem,  mitigation  measures  are  possible.  We 
 can consider: 
 ●  Establish a digital identity system to allow for tracking sources of information. 
 ●  Promote  third-party  services  that  monitor  content  channels  to  flag  disinformation,  and 
 correct it through decentralised systems that can tackle this problem at scale. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 10    |  58 
 ●  Establish  legal  requirements  for  content  channels  to  label  AI  generated  videos,  as  a 
 temporary  measure  while  enabling  the  widespread  adoption  of  digital  signatures 
 embedded  in  hardware  manufacturing,  such  as  digital  signatures  in  cameras  to 
 authenticate images. 
 ●  Increase public education and awareness in unknowingly spreading mis/disinformation. 
 An  additional  challenge  faced  by  non-English  speaking  countries  is  the  technical  diculty  of 
 detecting  mis/disinformation  in  non-English  languages,  since   many  existing  models  are  trained 
 on  and  optimised  for  English  datasets.  Therefore,  algorithms  need  to  be  trained  on  non-English 
 data  sets  in  order  to  accurately  detect  on  global  platforms.  Another  challenge  is  the  cost  of  fact 
 checking  posts  at  scale,  which  tends  to  be  much  higher  than  AI  generation  of  fake  posts.  This  is 
 exacerbated by AI generated content being disseminated at an increasingly low cost. 
 There  are  few  known  systems  to  monitor  the  flow  of  misinformation.  We  anticipate  both  the 
 public  and  private  sector  will  need  to  invest  in  R&D  to  fill  the  void.  Law  enforcement  will  need  to 
 expand  their  investigative  toolkits.  Given  the  nascence  of  the  problem,  global  sharing  of 
 experiences  would  improve  the  learning  curve.  Potentially,  there  will  be  AI  trained  specifically  to 
 address mis/disinformation. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 10    |  59 
 SCAI QUESTION 11    |  60 
 SCAI  QUESTION 11 

 SCAI QUESTION 11 
 A FRAMEWORK FOR EFFECTIVE 
 AI ADOPTION FOR SOCIAL GOOD 
 How  can  AI  adopters  effectively  evaluate  and  apply  AI  models  for  social 
 good? 
 Context & Assumptions 
 AI  developers  often  focus  on  improving  technology,  while  governments  regulate  AI  to  address 
 societal  risks  like  misinformation  or  crime.  Yet,  there  is  a  gap  in  effectively  integrating  AI  into 
 social  good  applications  by  governments,  Non-Governmental  Organisations  (NGOs),  and  social 
 enterprises, and a lack of thorough evaluation to measure their real impact. 
 In  the  private  sector,  AI  adoption  is  measured  by  revenue  and  costs,  making  it  easier  for  adopters 
 to  assess  impact.  However,  in  the  social  sector,  evaluating  outcomes  in  areas  like  education, 
 healthcare,  or  climate  change  is  more  complex  and  lacks  sucient  financial  and  technical 
 resources  for  analysis.  An  additional  complication  is  when  adopters  pick  up  a  successful   use  case 
 from  one  sector  or  context  and  apply  it  to  their  own,  without  evaluating  whether  the  model  or 
 outcomes  are  still  relevant  for  their  situation.  This  complexity  means  a  higher  reputational  risk  for 
 AI  companies  and  greater  potential  harm  from  poorly  designed  programs,  placing  more 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 11    |  61 
 responsibility on the AI industry for effective adoption in social sectors. 
 The  risks  of  early  AI  adoption  include  wasted  investments  and  negative  outcomes  for 
 participants,  alongside  reputational  damage  and  potential  overregulation  for  the  AI  industry.  With 
 growing  interest  in  AI  for  social  good,  it  is  vital  for  the  industry  and  adopters  to  develop  a 
 framework focusing on learning, piloting, evaluating, and capacity building. 
 For  example,  AI  can  boost  teacher  productivity  and  student  learning  in  education,  improve  patient 
 outcomes  in  healthcare,  and  provide  better  farming  recommendations  for  climate  change. 
 However,  rapid  implementation  without  proper  impact  assessment  can  lead  to  negative 
 consequences like reduced learning outcomes, incorrect health advice, or crop losses. 
 Question 
 How can AI adopters effectively evaluate and apply AI models for social good? 
 How can we offer a sociotechnical framework to AI adopters that enables them to: 
 ●  Accurately  assess  various  aspects  of  AI  models  for  social  good  use  cases,  including  the 
 dependencies  (such  as  access  to  computational  resources),  utility  (like  model  readiness 
 and  alignment  with  the  proposed  use  case),  and  appropriateness  (for  instance, 
 determining if general-use models are suitable for the intended purpose)? 
 ●  Pilot  and  rigorously  evaluate  AI  use  cases  to  comprehend  their  true  impact  compared  to 
 existing programs, ensuring that AI is being effectively adopted for social good? 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 11    |  62 
 Important considerations: 
 ●  It  is  crucial  to  emphasise  socio-economic,  cultural,  and  other  differences  that  might  lead 
 to  unintended  consequences  or  worsen  inequalities  and  biases  when  applying 
 general-purpose algorithms in social sectors like education or public healthcare. 
 ●  AI  adoption  should  not  be  rushed.  It  is  essential  to  first  pilot  and  rigorously  evaluate 
 AI-based interventions in real-world settings. 
 ●  The  AI  industry  needs  to  allocate  financial  resources  to  support  the  “Framework  for 
 Effective  AI  Adoption  for  Social  Good”  proposed  here.  This  includes  funding  for  accessible 
 in-person  and  online  training  for  social  sector  organisations  on  integrating  AI  into  their 
 programs  effectively;  for  conducting  thorough  evaluations  of  use  cases;  and  for  sharing 
 case studies that highlight both successes and failures in these applications. 
 Indicators of Progress 
 We suggest the following potential approach: 
 ●  A Framework for Effective AI Adoption for Social Good requires that adopters: 
 ○  Collaborate  with  the  AI  developer  to  gain  a  clear  understanding  of  the  potentials  
 and  limitations  of  the  AI  models,  including  the  data  they  were  trained  on,  the 
 values underlying that model, and relevance to the local context. 
 ○  Engage  with  researchers  who  possess  global  insights  into  both  effective  social 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 11    |  63 
 program  design  and  AI  integration.  Involve  them  from  the  design  stage  to  develop 
 a pilot for the AI-enhanced program. 
 ○  Initiate  a  concurrent,  independent,  and  rigorous  evaluation,  such  as  a  randomised 
 control  trial  (RCT),  of  the  AI  pilot.  This  is  to  accurately  measure  its  impact 
 compared  to  existing  methods,  offering  insights  into  what  works,  what  does  not, 
 and why. 
 ○  Proceed  to  scale  up  the  program  only  after  integrating  learnings  from  both  the 
 pilot  and  the  concurrent  evaluation.  If  the  pilot  is  found  ineffective,  consider 
 scaling down. 
 ○  Disseminate  the  insights  gained  from  this  pilot  and  its  evaluation  to  others  who 
 could  benefit  from  integrating  AI  into  their  programs.  Share  through  blogs  or 
 other open-source platforms. 
 ○  Address  the  needs  of  governments,  NGOs,  development  organisations,  and  social 
 enterprises  in  the  context  of  social  good  applications  (not  commercial 
 organisations  and  applications).  These  entities  often  lack  the  technical  or 
 financial  resources  for  the  above  analysis,  which  is  more  challenging  as  outcomes 
 are not solely measured in terms of revenue and costs. 
 ●  Known challenges or obstacles to answering this question include: 
 ○  Program  implementers  often  lack  access  to  the  global  knowledge  that  can  offer 
 valuable  insights  for  effectively  incorporating  AI  into  social  programs. 
 Researchers,  with  better  access  to  this  information,  can  be  crucial  partners  in  this 
 process. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 11    |  64 
 ○  There  is  sometimes  an  underestimation  of  the  importance  of  tailoring  the 
 program  to  the  local  context.  It  is  also  vital  to  refine  any  elements  that  did  not 
 perform as expected. 
 ○  Even  when  social  sector  adopters  recognize  the  importance  of  these  steps,  they 
 frequently  lack  the  technical  capacity  or  financial  resources  to  design  and 
 execute such a pilot and its thorough evaluation effectively. 
 ○  Tech  and  AI  companies  mostly  do  not  allocate  resources  to  assist  adopters  of 
 their  technology  for  social  good.  It  is  crucial  for  these  companies  to  support  social 
 sector  adopters  in  navigating  this  framework,  ensuring  that  only  relevant 
 technology is adopted, and that it is done so correctly and appropriately. 
 ●  Broad criteria for recognising progress in answering the question: 
 ○  The  AI  industry  and  adopters  broadly  use  and  adapt  this  “Framework  for  Effective 
 AI Adoption for Social Good”. 
 ○  The AI industry commits both financial and technical resources for: 
 ➢  Implementing  a  sucient  number  of  use  cases  across  various  sectors 
 such  as  education,  public  healthcare,  climate  change,  social  protection, 
 labour  markets,  and  agriculture,  and  in  diverse  socio-economic  and 
 geographic contexts. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 11    |  65 
 ➢  Publishing  open-source  studies  that  summarise  the  experiences  of  early 
 adopters,  including  evidence  from  rigorous  evaluations  of  the  above  use 
 cases.  These  studies  should  focus  on  understanding  what  works,  what 
 does not, and why. 
 ○  Building  capacity  of  key  actors  in  the  social  sector  to  effectively  adopt  AI  in  their 
 programs.  This  can  be  via  a  combination  of  in-person  training  for  government 
 AI/IT  departments  and  multilateral  development  banks  (like  the  World  Bank  and 
 Asian Development Bank), as well as virtual training for relevant NGOs worldwide. 
 ○  Over  time,  as  tech  and  AI  companies,  adopters,  and  researchers  gain  a  better 
 understanding  of  when  and  why  AI  is  an  effective  tool  in  social  sector  programs, 
 we expect to see fewer products exited and fewer pilots considered unsuccessful. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 12    |  66 
 SCAI QUESTION 12    |  67 
 SCAI  QUESTION 12 

 SCAI QUESTION 12 
 METHODOLOGIES 
 FOR AI SAFETY EVALUATION 
 How  can  we  establish  and  uphold  methodologies  for  AI  safety 
 evaluation? 
 Context & Assumptions 
 Although  high  level  concerns  and  principles  are  largely  agreed  upon  for  the  ethical  and  safe  use  of 
 AI  (see  for  example  the  UNESCO  2022  recommendation),  societies  have  yet  to  develop 
 standardised  techniques  for  mitigating  harm  and  auditing  procedures  for  safety  testing.  Although 
 structures  for  governance  and  regulation  are  the  topic  of  another  question,  here  we  focus  on 
 ways  to  operationalise  auditing  and  transparency  procedures.  We  define  safety  as  adhering  to  the 
 expected  functionality  of  a  system  and  avoiding  unacceptable  outcomes  which  may  be 
 considered  harmful  to  individuals.  Harms  include  social  and  psychological  harms,  and  harms  to 
 security,  economic  or  democratic  resilience.  Concerns  extend  throughout  the  lifecycle  of  a 
 product,  including  after  it  ceases  to  be  offered.  We  consider  transparency  as  a  core  component 
 of  safe  systems,  enabling  the  tracing  of  accountability  through  the  design  and  use  of  AI  systems. 
 Algorithm  transparency  is  not  just  a  tool  for  safety,  but  also  an  outcome  of  safety  auditing 
 processes. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 12    |  68 
 Potential  pitfalls  include  differences  between  the  data  on  which  a  model  is  trained  and  the  lived 
 experience  of  the  humans  (or  ecosystem)  in  which  it  is  deployed;  inadequate  understanding  of 
 users’  needs  and  context;  and  unanticipated  or  creeping  harms  such  as  gradual  increase  in 
 loneliness or loss of human autonomy. 
 Current  purely  technical  performance  evaluation  of  AI  systems  –  including  large-scale  generative 
 models  –  is  inadequate  to  measure  and  attribute  correctly  any  increase  of  harms  over  the 
 pre-AI-system  status  quo.  Today,  major  AI  developers  and  deployers  use  simple  metrics  such  as 
 diversities  of  datasets  and  outcomes,  but  this  approach  to  auditing  does  not  allow  for 
 comprehensive  evaluation  of  socio-economic  harms.  To  do  this,  we  firstly  need  a  scientific, 
 data-driven   method  to  understand  the  baseline  and  intended  outcome  pre-deployment,  and 
 subsequently  whether  there  is  any  increase  in  harm  post-deployment.  Secondly,  critical  systems 
 engineering  requires  testing  the  quality  and  reliability  of  system  outputs,  which  should  consider 
 the  context  in  which  the  objectives  of  the  system  and  expectations  of  users  are  defined.  These 
 should  include  users’  response  to  and  understanding  of  the  systems,  which  is  in  turn  dependent 
 on sociotechnical considerations, including user education. 
 Question 
 How can we establish and uphold methodologies for AI safety evaluation? 
 Relevant  corollary  questions  include:  What  process  can  we  use  to  capture  and  ensure  the 
 measurement  of  suspected  harms  (measuring  macro/society  and  micro/individuals  and 
 families)?  What  statistics  do  we  need  about  deployment,  uptake,  and  modifications  to  deployed 
 systems  in  order  to  establish  causality  between  design  choices  and  potentially  negative  social 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 12    |  69 
 outcomes?  For  any  AI  system,  who  are  the  users  that  are  affected  by  either  individual  model 
 components,  or  by  human  actors  together  with  the  model?  How  can  we  design  a  transparency 
 report  that  clearly  states  the  expected  users,  intended  outcomes,  and  anticipated  candidate 
 harms? 
 How  do  we  divide  responsibility  between  developers  of  AI  component  systems  and  deployers 
 whose  products  interact  directly  with  the  end  users?  We  assume  here  that  transparency 
 requirements  on  deployers  should  be  passed  through  their  supply  chain  –  that  is,  deployers  are 
 responsible  for  sourcing  adequately-tested  AI  components,  and  for  having  a  clear  reporting  path 
 back  to  developers  if  issues  are  discovered  among  the  deployers’  users;  developers  are  then 
 responsible  for  system   redesign  according  to  this  feedback.  Who  establishes  adequacy 
 benchmarks  per  deployment  sector,  and  how  are  these  communicated  back  to  developers  and/or 
 used by deployers to assess systems’ readiness for deployment? 
 Can  controlled  auditing  and  simulation  effectively  estimate  societal  risks?  If  so,  what  are  the 
 standards  and  obligations  for  testing  model  predictions  before  system  deployment?  In  what 
 contexts  should  we  design  adversarial  testing,  and  with  what  frequency  should  such  checks  be 
 executed? 
 There  are  further  questions  regarding  post-deployment  safety.  What  categories  of  AI  systems 
 benefit  from  paid  incentivising  for  the  reporting  of  safety  and  security  problems  (e.g.  bug 
 bounties,  ethical  hacking  for  AI)?  How  do  we  create  and  enforce  processes  for  ordinary  users  to 
 report  suspected  problems  and  receive  responses  (e.g.,  explanations)?  How   do  we  aggregate 
 such  reports  and  attribute  them  to  candidate  causes  (e.g.,  flaws  or  active  compromise  of  specific 
 foundation models)? 
 Answering  these  questions  should  contribute  to  deployment  of  safe  AI;  improve  AI  development 
 and  innovation;  ensure  companies,  governments,  and  potentially  civil  society  have  access  to 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 12    |  70 
 adequate  information  about  each  other  to  do  their  jobs;  and  set  precedents  for  transparent 
 governance, bottom-up “policing” and understanding of rights. 
 Indicators of Progress 
 We will witness progress through the following measures: 
 ●  Corporate and civic confidence in deploying AI. 
 ●  Transparency  to  the  public  (e.g.  through  clear  documentation)  on  how  context-dependent 
 acceptable and unacceptable outcomes of AI systems are defined. 
 ●  New  standards,  reusable  tests  and/or  procedures  for  constructing  tests,  which  focus  on 
 measuring  reliability  and  impact.  The  costs/  benefits  of  AI  use  should  be  broken  down  by 
 sub-populations. 
 ●  Standard  checklists  and  regular  reports,  to  be  reported  in  media  and  made  available 
 through  national  standards  organisations.  The  reporting  of  AI  harms  should  be  reliable 
 (e.g., non-spurious). 
 A  possible  approach,  in  analogy  with  cybersecurity,  is  institutions  like  an  AI  CERT  (Computer 
 Emergency  Response  Team)  that  will  gather  and  analyse  reported  AI  vulnerabilities   and  failures 
 and  work  with  model  developers  and  deployers  to  continuously  improve  the  safety  of  the  AI 
 ecosystem. Such institutions could help address the specific post-deployment questions above. 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 12    |  71 
 Challenges include: 
 ●  Multiple stakeholders with competing interests. 
 ●  Gaining  reliable  access  (at  least  for  trusted  parties)  to  proprietary  systems  or  confidential 
 data, which are needed for auditing. 
 ●  Ensuring  the  veracity  of  documents  achieved  through  such  access  and  that  test 
 performance corresponds to every-day, real-time performance. 
 ●  Designing  metrics  and  actionable  methods  to  accurately  quantify  risks  to  safety.  This  will 
 require us to establish baseline social characteristics (e.g. using the World Values Survey). 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  SCAI QUESTION 12    |  72 
 ACKNOWLEDGEMENTS    |  73 
 ACKNOWLEDGEMENTS 
 SCAI QUESTION 1 

 ACKNOWLEDGEMENTS 
 The  community-based  process  for  producing  these  questions  is  outlined  in  the  foreword.  For 
 more information about the SCAI process and community, please see  https://www.scai.gov.sg/  . 
 Copyright © 2023 Government of the Republic of Singapore 
 You may download, view, print, and reproduce this document without modification, but only for non-commercial use. 
 All other rights are reserved.  ACKNOWLEDGEMENTS    |  74