TL;DR: On a long AI project, section one hits the ask, section three mostly hits it, and by section eight the output is covering something the prompt never asked for. By section twelve it has wandered into territory you wouldn’t have approved if you’d seen it framed plainly, and nobody catches it, because each section reads fine on its own.
The bend only shows up when you map the through-line end to end, and that mapping is how I catch AI drift on a book. Build the defense before you discover the drift the way most consultants do, when the client implements a recommendation and the supporting analysis turns out not to support it.
Drift worries me more than hallucination does. A made-up citation can be checked against a database. A document that slowly changes its mind leaves nothing to look up, and it gets past smart reviewers because every page looks fine on its own.
The symptom
You give the AI a clear prompt for a long document. Draft a 50-page strategic analysis. Write a chapter-by-chapter book outline. Generate a complete technical specification. The output arrives, you skim the first few sections, the structure looks right, the tone is consistent, and you ship it.
Drift is the silent killer of long AI projects because nobody catches it.Share on X
Three weeks later somebody who has to use the document points out that section 22 contradicts section 4.
Or that the conclusion of the executive summary doesn’t match the conclusion of the actual analysis. Or that the chapter outline you got was for a book about leadership and somehow ended up halfway through being a book about marketing. Or that the technical spec, which was supposed to describe a payment system, has gradually drifted into describing a CRM.
Nobody noticed because each individual section was internally coherent. The drift only became visible when somebody traced the through-line. The machine will tell you, though, if you ask. Ask Claude whether it’s drifting and it’ll say where and fix it, and ask whether it’s lying and it’ll often admit it made something up and go find facts that check out.
This is the failure mode I see more than any other in AI-generated long-form work. It’s silent, it’s hard to catch, and it produces the highest-cost failures because the documents that suffer from it are exactly the documents nobody reads end to end before publishing.
Watch: ML Drift: Identifying Issues Before You Have a Problem, Fiddler AI
What is AI drift?
Drift is the gradual divergence of an AI output from the original ask, accumulating across the length of a long generation. The system handles section one based on the prompt. It handles section two based on the prompt and section one. By section ten, the system is handling the next section based on the prompt and sections one through nine, and the prompt is increasingly far away in the context window. In plain words, hallucination is AI making up answers or changing them, and it happens a lot. Drift is where it starts off right and then slowly changes over time until it’s unrecognizable.
Each section is technically still on topic. It’s also responding to whatever was in the section just before it, more than to the prompt that set up the whole document. So if section three made a side observation that slightly redefined the scope, section four picks that up. By section eight, the side observation has become the implicit scope, and the original prompt’s framing has been demoted.
The mechanism is working as designed. Language models attend most strongly to recent context, because that’s how you produce coherent paragraph-to-paragraph prose. The same mechanism that keeps the writing coherent at the local level is what produces drift at the document level.
That’s why I get uneasy when someone says they had AI write a whole book over a weekend. The first chapter might be exactly what they asked for. By chapter nine the book can be arguing something they’d never sign off on, and they won’t know, because they read it the way it was written, one chapter at a time.
Why does nobody catch AI drift on review?
Three reasons.
The first is that drift doesn’t produce errors. It produces a different document than the one you asked for. Errors are easy to spot. A document that’s competent but on the wrong subject is harder, because nothing on the page is obviously wrong.
The second is that humans review long documents the way they were written, section by section. You read section one and check that it’s on topic.
You read section two and check that it follows from section one. You don’t, usually, read section twelve and ask whether it still answers the question you asked back in the prompt. The drift hides in the gap between local coherence and global coherence, and your review process is mostly local.
A document that’s competent but on the wrong subject is harder, because nothing on the page is obviously wrong.
The third is that AI prose is fluent. Every sentence reads like a sentence somebody would write. So as the document drifts, it sounds like a document that’s making subtle, perhaps even sophisticated, additions to the original scope. The reviewer’s instinct is to trust that the writer had a reason, because that’s the instinct calibrated against human writers. The AI didn’t have a reason. It was just responding to its own previous output.
That misplaced trust is the first thing I’d fix on any team using AI for long documents. Reviewers need to hear plainly that the machine had no plan, so a surprising turn in section eight is a warning sign and deserves a second look. I tell people to treat it like a three-year-old. It’s going to get things wrong and wander off into la-la land, especially on longer work, and you just have to know that.
Where does AI drift show up worst?
The longer the document, the worse the drift. Output under 500 words rarely drifts at all. Output between 500 and 2,000 words drifts occasionally, in ways that are usually catchable on review. Output above 2,000 words drifts often, and the drift grows with length.
Anywhere you ask the AI to generate something where the structure is sequential, drift becomes a structural risk. Book outlines, strategic analyses, technical specs, training curricula, multi-section reports, long-form research synthesis. The system handles the first section against the prompt and then drifts away from it across the remaining sections, and by the end you have a document that started in one place and ended in another.
The worst case I’ve seen, working with clients on long-form ghostwriting projects, is when a consultant uses AI to draft a strategic deliverable for a Fortune 500 client. Eighty pages. The executive summary lays out three recommendations. The body of the report contains the analysis supporting those recommendations. Somewhere in the middle of the analysis, drift sets in. By the time the analysis reaches its conclusion, it’s no longer fully supporting the recommendations in the executive summary.
The reader who reads the executive summary and the conclusion side by side notices the gap. The reader who reads the document section by section never sees it. The client implements a recommendation. The analysis turns out not to support it. The consultant doesn’t get hired again. That case is far too ordinary. The consultant read the parts he always reads, and the client paid for the parts he didn’t. I covered this scenario in Six Places AI Will Break Your Work. Drift is the most common reason long-form AI deliverables fail in professional services.
How do you spot drift in AI output?
Three checks. None of them are optional if you’re publishing the document.
The bookend check. Read the first 200 words and the last 200 words of the document side by side. Do they answer the same question? Do they assume the same scope? Do they treat the same subject as the central one? If the first 200 words say the document is about A and the last 200 words say it’s about B, the drift between them is what you need to investigate.
The section-title check. List the section titles in order. Do they tell a coherent story from start to finish? Does each section title follow logically from the section before? Or do you see a jump halfway through where the topics start drifting in a different direction than the early sections were promising?
The thesis check. Pick the document’s central claim. Read every section and ask whether that section supports the claim, contradicts it, or addresses something tangential. If you find sections that support a different claim than the document’s main one, you’ve found where drift took the document.
None of these checks takes more than 20 minutes on an 80-page document, and almost nobody’s running them on AI-generated long-form work right now. That drives me a little crazy. Those 20 minutes are the difference between a deliverable you can defend and one that surfaces three weeks after it ships with its argument pointing somewhere else.
How to defend against drift while generating
The fix is structural. Don’t generate the whole document in one pass.
Generate in short sections. Have the AI generate one section at a time, with the section’s specific objective in the prompt for each generation. Pass the previous section as context if you need continuity, but refocus the prompt on the document’s central question every time.
My own conversations with Claude tend to run very long, because it can hold a huge amount of context. Eventually the conversation starts slowing down, and that’s when I switch. I tell Claude to write up something suitable for itself so the next conversation can pick up where this one left off. I download that document, bring it into the new conversation, and it carries on from where I was.
Re-state the thesis in every section prompt. The AI’s attention bends toward recent context. If you re-state the thesis at the top of every section’s prompt, the recent context now includes the thesis, and the section is more likely to stay on it.
Generate the outline first, and lock it. Don’t let the AI generate the structure of the document at the same time as the content. Have it draft the outline. Edit the outline yourself, with your own judgment about what each section needs to cover. Then generate the content of each section against the locked outline, not against an open-ended “continue the document” prompt.
Review the through-line before you review the prose. Once you’ve assembled the full draft, do the three checks above before you do any other editing. If the through-line is broken, fixing the prose is wasted work because the document is going to need restructuring anyway.
One more habit I rely on. Whenever I have Claude write something of any length, like a long article or a series of posts, I ask it to check for drift. A single short article doesn’t need it. Ask for five posts on a topic, though, and the intent slowly shifts as it goes. By the fifth one, it might not have what you wanted.
| Step | What to do | When |
|---|---|---|
| Lock the outline | Have AI draft it, edit it yourself, then generate against it | Before generating |
| Generate in short sections | One section per prompt, each with its own objective | While generating |
| Restate the thesis | Put the thesis at the top of every section prompt | While generating |
| Bookend check | Read the first and last 200 words side by side | After the full draft |
| Section-title check | List titles in order and look for a mid-document jump | After the full draft |
| Thesis check | Ask whether each section supports the central claim | After the full draft |
| Review the through-line first | Run the checks before any prose editing | Before editing |
My take
Check the through line before you touch the prose on anything long the AI wrote. Each section reads fine alone, so nobody notices the argument bending. Lock the outline yourself, generate one section at a time, and restate the thesis in every prompt. If you ask for five posts on one topic, read the fifth against the first, because it can wander off. Twenty minutes of checking is cheap next to an argument pointing somewhere else after it ships.
The cost of skipping the defense
The cost of catching drift before publication is small. Twenty minutes of through-line review. Maybe a section that needs to be rewritten because it drifted somewhere unhelpful.
The cost of catching drift after publication scales with where the document ended up. A drift in an internal memo is a moment of confusion in a meeting. A drift in a strategic deliverable is a client implementing the wrong recommendation.
A drift in a published book is a chapter that contradicts another chapter and a reader who notices. A drift in a technical specification is an engineering team building the wrong thing.
Drift is the most preventable failure in AI workflows, and it’s the one most consistently shipping in production. The defense is cheap. Skipping it costs clients and credibility, and I think anyone who bills for a long AI-assisted deliverable without tracing the through-line is shortchanging the person who paid for it. Section-by-section review feels rigorous, and that feeling is exactly what lets the drift through.
The Death of Thinking is also about this broader failure mode, what happens to a profession or a culture when a system that stopped answering the original question several sections ago has generated the loudest output in the room, and nobody noticed.
Drift wins when nobody stands outside the work and checks it. The Death of Thinking covers that slow loss and what it costs the people who stop checking.
Frequently Asked Questions
- LongProc: Benchmarking Long-Context Language Models on Long Procedural Generation, arXiv (Princeton University and UT Austin researchers)
- Prompting best practices, Anthropic
