Just a quick one, really – but it’s been nagging at me.
Something I’ve noticed, using ChatGPT and similar tools, is how tightly they cling to structure. Not just grammar in the technical sense, but intent. They respond best when the language has a clear shape – when the request knows what it’s trying to be.
And English, for all its flexibility, works against that more often than we’d like to admit.
We’re loose with it. We hedge, we imply, we trail off mid-thought and expect the other person to fill in the gaps. That works fine with humans – we’re good at reading between the lines. But with an LLM, that looseness turns into noise.
So if I call someone a “turnip”, the meaning isn’t in the word itself. It’s carried by tone, inflection, timing, attitude. The context sits with the speaker, and the listener reconstructs it.
An LLM doesn’t have access to that layer. It only sees the word. and get confused over a root vegtable
You see it in the outputs. Vague question in, vaguely helpful answer out. Not wrong, exactly. Just… unfocused.
I started paying attention to how I was phrasing things, almost as an experiment. Not over-engineering it, just tightening the intent. And the difference was noticeable.
Not anything dramatic. But more consistent.
Take something like:
“Can you look at this and tell me what you think?”
Perfectly normal sentence. You’d say it to anyone.
But to a model, it’s doing too many things at once – and none of them clearly. Look at what? Think about what? In what capacity?
So you get a kind of generic, hedged response. It fills the space, but it doesn’t lock onto anything.
Same with:
“This doesn’t feel right, can you fix it?”
Fix what, exactly? Tone? Structure? Accuracy? Logic? You haven’t said. So it guesses. Sometimes well, sometimes not.
Now compare that with:
“Audit this paragraph for clarity and logical flow. Identify any vague phrasing and suggest specific rewrites.”
Or:
“Rewrite this in UK English, keeping the tone professional but conversational. Reduce redundancy and tighten sentence structure.”
Nothing fancy. Just explicit.
The output shifts almost immediately. It’s sharper, more deliberate. Less filler, more intent.
Words like “audit”, “analyse”, “rewrite”, “summarise” – they act like anchors. They tell the model what mode to enter. You’re not just asking for help; you’re defining the task.
Even small tweaks matter:
- “Read this” → passive, open-ended
- “Critique this for argument strength” → directed, evaluative
Same text. Different outcome.
This is where it gets interesting.
Research in Autism spectrum condition often points to differences in how language is processed and used. Not deficits in intelligence, but differences in pragmatics – the social use of language.
A few consistent patterns come up:
- Preference for precision over implication
- Reduced reliance on tone, sarcasm, or indirect cues
- Greater comfort with explicit instructions and defined rules
- Less tolerance for ambiguity or underspecified requests
In other words, a tendency towards structured communication.
Where a neurotypical speaker might say:
“That’s… interesting.”
– and expect the other person to detect irony, discomfort, or mild disapproval –
an autistic speaker is more likely to say:
“I don’t think that works because the argument isn’t supported.”
Clear. Direct. Minimal inference required.
If you strip away the human layer, LLMs behave in a way that’s oddly compatible with that style.
They don’t read tone unless it’s explicitly encoded. They don’t infer intent unless it’s clearly signposted. They don’t “fill in” social meaning in the same way humans do.
So the communication style often associated with autism – precise, literal, structured – maps surprisingly well onto how these systems operate.
It’s not that autistic communication is “more correct”. It’s that it’s closer to the input format the system expects.
You could argue that:
- Neurotypical communication = high context, low explicit structure
- LLM interaction = low context, high explicit structure
And autistic communication often sits somewhere closer to the latter.
Compare these two prompts:
“This feels off, can you tweak it a bit?”
versus:
“Identify inconsistencies in tone and rewrite the passage to maintain a formal, analytical voice.”
The second isn’t just clearer. It removes ambiguity almost entirely.
That kind of phrasing – task-oriented, bounded, explicit – is very common in autistic communication, particularly in technical or written contexts.
The hidden advantage
There’s a quiet advantage here that doesn’t get talked about much.
People who already default to structured, explicit language don’t have to “learn” how to prompt effectively in the same way. They’re already doing a version of it.
Where others have to unlearn vagueness, they’re starting from clarity.
It’s less about adapting to the machine, and more about realising the machine already speaks something close to your dialect.
That said, it’s not a perfect overlap.
Human communication – even autistic communication – still carries nuance, emotion, and context that LLMs don’t truly grasp. There’s a risk of over-fitting your language to the system and losing some of that human texture.
Clarity helps. Over-constraint can flatten things.
So there’s an odd convergence happening.
English, as we casually use it, is messy, implied, and socially encoded. AI systems prefer it clean, structured, and explicit. Autistic communication often leans in that same direction – not by design, but by necessity.
Which means that, in this very specific context, a style of communication that’s often misunderstood or undervalued in everyday conversation suddenly becomes highly effective.
Not because it’s trying to be.
But because it already fits the shape of the system. er.
You see it in the outputs. Vague question in, vaguely helpful answer out. Not wrong, exactly. Just… unfocused.
I started paying attention to how I was phrasing things, almost as an experiment. Not over-engineering it, just tightening the intent. And the difference was noticeable.
Not anything dramatic. But more consistent.
Take something like:
“Can you look at this and tell me what you think?”
Perfectly normal sentence. You’d say it to anyone.
But to a model, it’s doing too many things at once – and none of them clearly. Look at what? Think about what? In what capacity?
So you get a kind of generic, hedged response. It fills the space, but it doesn’t lock onto anything.
Same with:
“This doesn’t feel right, can you fix it?”
Fix what, exactly? Tone? Structure? Accuracy? Logic? You haven’t said. So it guesses. Sometimes well, sometimes not.
Now compare that with:
“Audit this paragraph for clarity and logical flow. Identify any vague phrasing and suggest specific rewrites.”
Or:
“Rewrite this in UK English, keeping the tone professional but conversational. Reduce redundancy and tighten sentence structure.”
Nothing fancy. Just explicit.
The output shifts almost immediately. It’s sharper, more deliberate. Less filler, more intent.
Words like “audit”, “analyse”, “rewrite”, “summarise” – they act like anchors. They tell the model what mode to enter. You’re not just asking for help; you’re defining the task.
Even small tweaks matter:
- “Read this” → passive, open-ended
- “Critique this for argument strength” → directed, evaluative
Same text. Different outcome.
There’s another layer to this, and it’s easy to miss.
If you start with a loosely phrased question, the model often mirrors that looseness back. It tries to be helpful, so it fills in the gaps, smooths things over, makes assumptions. The tone becomes slightly vague, slightly padded.
Then – without thinking – you respond in kind. You adapt to its style. The whole exchange drifts.
That’s the negative loop.
It’s subtle, but it compounds. Each step gets a little less precise.
The alternative is to set the tone early. Be deliberate with the first prompt. Not robotic, just clear. If the initial query is structured, the response tends to follow suit – and now you’ve got something stable to build on.
That’s your positive loop.
At a basic level, these models are pattern engines. They’re not “understanding” in the human sense; they’re mapping your input to the most probable useful response based on structure and context.
So when your input is:
- Clear in task → it selects a narrower, more relevant pattern
- Ambiguous in intent → it defaults to broader, safer language
It’s not being sloppy. It’s being general because you’ve asked a general question.
Don’t just ask a question – assign a role and a task.
Instead of:
“Can you help with this?”
Try:
“Act as an editor. Tighten this passage for clarity and rhythm without changing the meaning.”
You’re giving it boundaries. And within those boundaries, it tends to perform much better.
It’s a small shift, but it changes the interaction quite a bit. Less back-and-forth, fewer corrections, cleaner outputs.
Not perfect, obviously. But noticeably better.
And once you see it, you start noticing it everywhere.