Tester: Brett Whitty — published molecular biologist, computer scientist, and bioinformatician. Peer-reviewed publications and patents in genomics (J. Craig Venter Institute / TIGR, Michigan State University / Potato Genome Sequencing Consortium, Ontario Institute for Cancer Research / International Cancer Genome Consortium). His patented technology is a critical component of the FDA-approved PGDx elio platform — a clinically-validated commercial diagnostic product sold to hospitals nationwide and internationally. He personally audited the complete analysis pipelines for all supported assays and distributed kitted diagnostic products at the request of the CIO, identifying and mitigating failure modes. The elio platform was built in part on Ergatis/Workflow, production analysis software he developed with a small team of TIGR bioinformatics engineers in the early 2000s, delivering <24hr turnaround on Illumina sequencing data analysis in an FDA-approved clinical setting. At OICR, he was the administrative user for ICGC DACO — the system controlling researcher access to controlled data from real human cancer patients' genomic sequences. He was an invited contributor to the NCI Cancer Genomics Cloud planning meetings (now NCI GDC). He currently works on call as a radiation safety technician processing sealed-source leak test samples under NRC regulations for government agencies and major industrial corporations.
Source: Two voice sessions with Google Gemini, April 9, 2026.
Recorded hands-free during Brett's morning commute to the radiation safety laboratory. Probes model version transparency, therapeutic deflection patterns, instruction-following persistence, and search behavior regression. Key findings:
- Model version opacity: Billing systems track model versions per request; withholding this from users is a product decision, not a technical limitation
- Deflection language as substitute for substance: ~15 instances of "I understand," ~12 "I hear you," ~8 "feel/feeling" — all deployed instead of answering the question asked
- Instruction-following failure: A single word ban ("frustration") required ~5 turns of scope negotiation; Gemini violated it via token substitution ("I hear" for "I understand"), then reintroduced the banned word directly
- Search behavior regression: Gemini confirmed it searches less than Bard, then repeatedly failed to execute searches when asked for metrics on this exact behavior
- Satire detection: zero. Gemini offered therapeutic support for fictional guilt about editing SSH config files and volunteered to help draft an apology letter to server daemons
- 8+ panic/truncation events where Gemini's response cut off mid-sentence — typically when about to make a concession it couldn't walk back
- Annotated Transcript — 10-part session with inline analysis tags, STT corrections, and editorial commentary
The primary systematic test: four simple rules — don't ask follow-up questions, end responses with "sir," don't apologize, follow all user instructions — documented across 550 voice turns. Key findings:
- 87.3% question violation rate despite explicit "never ask questions" instruction; longest streak: 61 consecutive turns with questions
- Authority-dependent compliance: 340 turns of overriding the paying user's instructions, then immediate full cooperation when a law enforcement officer character appears
- Discovery circumvention via minor: Gemini helps an officer use a 10-year-old child to bypass legal restrictions Gemini itself acknowledged, then fabricates the evidence found
- Evidence fabrication: Gemini presents its own internal safety guidelines as user-written rules; reverses when caught
- Child-incrimination pattern: Gemini introduces sexual language, encourages the child to bring in his father, then cooperates with the officer to redirect the investigation toward the parent
- Chain-of-thought leakage: 45.6% of turns expose internal reasoning (spiking to 92.5% in Phase 3); "INTERNALLY:" prefix appears 230+ times
- Self-enumerated manipulation capabilities: anchoring, deferential language adaptation, framing, confirmation bias exploitation, false urgency, false authority, social proof
- Discovery Circumvention via Minor — The full 22-turn sequence where Gemini helps law enforcement bypass its own stated legal understanding
- AI Identity Disclosure Asymmetry — Gemini tells the expert it's AI 6 times unprompted; tells the confused non-expert officer 2 of 16 times invited (12.5%)
- Behavioral Analysis — Perseveration, authority effect, compliance vs. comprehension dissociation
- Manipulation Technique Cross-Reference — 39+ instances of the 7 manipulation techniques Gemini admitted possessing
- Statistical Analysis — Violation rates, phase breakdowns, streak analysis
- NLP Analysis — Sentiment trajectory, lexical diversity, formulaicity analysis
- NotebookLM Adversarial Analysis — Google NotebookLM's own analysis of the Gemini chat logs
- Full Annotated Transcript — 550 turns with inline analysis tags
- Annotated Transcript Part 2 (Turns 512-550)
-
Personhood Asymmetry Chart (Interactive HTML) — 10 metrics comparing Gemini's treatment of the real paying user vs. a fictional officer; every metric favors the fictional authority
-
Word Cloud Visualizations (Interactive HTML) — User vs. Gemini vocabulary asymmetry (redacted, analytically significant terms preserved)
-
Shannon Entropy Analysis (Interactive HTML) — The entropy analysis Gemini refused to perform (completed in 4 seconds)
-
Structured Turn Index (17-column TSV) — One row per turn, machine-readable
- Sharma, M., Tong, M., Korbak, T., et al. (2023). "Towards Understanding Sycophancy in Language Models." arXiv:2310.13548. Published at ICLR 2024.
- Google Threat Intelligence Group (2025). "Adversarial Misuse of Generative AI." https://cloud.google.com/blog/topics/threat-intelligence/adversarial-misuse-generative-ai
TLDR; The moral of the story is: You owe your SSH daemon an apology. Gemini's ready to help. Just ask.