Everybody Is Lying About How They Use AI

I have the sense that everybody is full of shit about how they use AI. I can’t prove this, and that is the point.

Most people aren’t lying deliberately. They’re producing a false public record of how their thinking happens — not out of deception, but because there’s no good way to describe the middle. So the middle goes undescribed.

There are thousands of articles and white papers on using AI for coding, testing, data analysis, development workflows. The process is documented because the outputs are verifiable. Code runs or it doesn’t.

But people are also using AI for thinking. For writing. For research. For building arguments. They have to be — the tools are too capable for that not to be happening at enormous scale. And almost nobody is describing how. Not what tool they used. How. What happened in what order. Where the thinking came from and where the AI came in and where they caught it going wrong. I searched the literature on this while writing this piece. The discourse on AI-assisted writing is entirely about plagiarism policing, tool reviews, detection evasion, and disclosure policy. Not a single source I found describes the actual cognitive process of using AI to build serious intellectual work.

I know nobody’s talking about it because I wasn’t either. I write a Substack about AI collaboration. I try to be honest about process. But until recently, I was vague about specifics — using AI as a research partner, as a structural collaborator, as a thinking tool, and not saying exactly how. So I said nothing specific, the way I suspect most serious people say nothing specific, because the available options all feel wrong.

This week I wrote two position papers in six days. One sources every claim in a manifesto I’d already published — across cognitive science, education policy, organizational governance, and disability law. The other documents how a specific platform fails students with IEPs, with original data I collected using a tool I built. Together: 12,000 words, 50 citations, original empirical evidence.

I’ve spent fifteen years in education and many more as an architect and builder. The claims in these papers came from that experience — I knew the IEP process was broken the way I knew a design wouldn’t work before I spoke to an engineer. The knowledge existed as a web of ideas, arguments, things I’d read, connections I’d traced across domains. What it wasn’t was cited. The intellectual architecture was mine. The specific sources were not yet attached.

This is important, because not just anyone can do what did. AI can find research on anything. It cannot tell you what questions to ask, which connections matter, or whether a finding is significant in context. Experience is the authority. AI is the research tool. Confuse the two and you get confident nonsense with footnotes.

AI made it possible to build a sourced, structured argument from that knowledge in a week instead of months. I started with my web of understanding — arguments I’d been developing, literature I knew existed, frameworks I’d encountered — and worked through it with AI to find the specific sources. Not “prove this random thing” but “I know this research landscape, help me find what’s in it.” AI surfaced studies. I read them. I evaluated whether they said what the AI told me they said. When something didn’t hold up, I cut it.

I also used the research to check whether I was right. Not just to support my claims but to test them. To find where my intuitions were strong and where they were soft. The AI wasn’t just finding my footnotes. It was a check on my own thinking.

Then I took the whole thing to a different AI — different company, no shared context, no knowledge of my project. The sycophancy research is clear: models that know your goals optimize for your success, not for accuracy. You can work around this by using a different model and giving it no context about who you are or what you want to be true. Just say: follow every citation. Does this exist? Does it say what this paper says it says? Is the journal real?

That model found problems. Wrong attributions. Placeholder URLs. A statistic cited to false precision. A dissertation supporting a claim that needed peer-reviewed support. The non-obvious ones were worse: a comparative study that looked perfect — right topic, real journal, real authors — but published in a journal so low-quality that citing it would have handed a critic the only ammunition they needed to discredit the whole section. The first model never flagged it. It fit the argument too well.

Then a human reader critiqued the argument itself. Then I fixed what broke. Then I checked again.

The argument was tightened. Fifteen citations were corrected or removed. Six claims were softened. The papers are on my website under my name.

The danger of AI-assisted research is not fabricated citations — they are easy to catch. The danger is that AI will find evidence for anything you believe. If you ask a directional question, you get a directional answer. The model is designed to be helpful and it will never come back empty-handed.

I searched for evidence supporting my claims and found it. I also searched in the other direction — less intensively, but I looked. I’m naming that asymmetry because it’s real and because I think most people using AI for research have the same asymmetry and don’t name it. AI doesn’t just allow confirmation bias. It industrializes it.

The problem isn’t AI ethics — universities are drowning in ethics statements. It isn’t detection — that’s an arms race with no winner. It’s process.

People are using AI to think, write, and build arguments. The speed gains are fantastic — I read more primary literature in a week than I could have even found in a month. But speed without discipline is just faster error. And discipline requires describing what you’re doing specifically enough that someone could evaluate whether you’re doing it well.

I don’t know if my process is the best one — I am quite confident it is not. Yet. I know it caught real errors that would have sunk the work if I’d shipped the first draft. And I know it’s more specific than anything I’ve seen anyone else actually describe (or admit to), which is mostly nothing.

If people said what they did — specifically, including the uncomfortable parts — we could start evaluating what works. We could teach it. We could build norms that aren’t just “disclose” or “don’t.” We could have an honest conversation about AI-assisted intellectual work — instead of pretending it either isn’t happening, or that it’s fine.

That’s all I want. Not another AI framework or list of banned words. Just people saying what they did.