AI writing tells: drift

This is the seventh piece in the AI writing tells series, and it covers a tell that only appears at length. The other six describe patterns you can find in a paragraph. This one you can only see across a whole book, because it is about what happens to machine writing as it runs: it drifts. It loses track of itself, and it loses quality, and by the end it is a different and worse thing than it was at the start, and the author, who generated it in pieces, often cannot see that it happened.

I am putting it last because it is the tell that does the most damage to the longest works, which is to say to novels, which is to say to the exact thing most people are using these tools to produce. A short piece does not drift much. A novel drifts enormously, and the drift is a large part of why machine-written novels fall apart in the back half in ways their authors never intended and often never notice. This piece also covers drift’s close cousin, hallucination, the machine inventing facts that were never there, because the two failures travel together and you have to guard against both.

Two kinds of drift

There are two and they compound. The first is continuity drift, where the machine loses track of the facts of its own story. The second is quality drift, where the writing itself degrades the longer it runs. They have different causes and they happen at the same time, and together they mean that the last chapter of a machine novel is both less accurate and less well written than the first, while looking, sentence by sentence, about the same.

That last part is what makes drift so dangerous. It is gradual. No single chapter is dramatically worse than the one before it. The decline is slow enough that if you are generating and reading in order, chapter by chapter, you acclimate to it, the way you do not notice a room getting dark until someone turns on a light. Only when you set the last chapter beside the first do you see how far the thing slid, and most authors never do that, because by the time they finish they are sick of the book and they never read it whole.

Continuity drift: the story forgets itself

The machine does not hold the whole story in mind. It holds a window, and as the book runs, the early parts slide out of that window, and once they are gone the machine no longer knows what it established. So it re-establishes, and the re-established version does not match.

A character’s eyes are grey in chapter two and brown in chapter nine. A house has three storeys early and four late. A character who was an only child mentions a brother. The protagonist’s dead father is alive in a flashback that contradicts the death established two hundred pages back. A minor character’s name shifts by a letter, then by a syllable, then becomes a different name entirely while remaining the same person. None of these are dramatic. Each is the kind of small factual slip a human writer catches on a reread and the machine never catches because the machine never rereads, it only ever generates forward from a window that has already forgotten chapter two.

Worse than the contradicted details are the dropped threads. The machine introduces a mystery, a tension, a planted object, a relationship, and then, when the thread slides out of the window, it simply forgets to pay it off. The locked drawer in chapter three is never opened. The ominous stranger in chapter five never returns. The illness the protagonist was diagnosed with stops being mentioned and is, apparently, cured by being forgotten. A human writer plants in order to pay off, and tracks the debts. The machine plants because planting is what stories do at that moment, and then loses the debt when the window moves on, and the back half of a machine novel is littered with these unpaid setups, these promises the book made and forgot it made.

The deepest continuity drift is structural. You can give the machine a careful outline, a nine-act structure, a specific plan, and it will follow it for a while, and then the plan slides out of the window and the machine reverts to the only structure it truly knows, which is the three-act shape covered in the spine article. By the end it has quietly abandoned your structure and collapsed back into its default, and unless you are holding the outline beside the draft, you will not catch the moment it happened, only the result, which is a book that started as the thing you planned and ended as the thing the machine always makes.

Quality drift: the writing wears down

At the same time as the story is forgetting itself, the prose is wearing down, and this is the subtler and in some ways sadder degradation.

Early in a generation, the machine is at its best. It has the most room, the most attention, the freshest sense of the voice you asked for. The opening chapters of a machine novel are often genuinely decent, and this is a trap, because they are the chapters the author judges the tool by, and they are not representative. As the book runs, the quality erodes. The voice you specified at the start fades, and the machine slides back toward its own default register, the smooth average tone it produces when no voice is actively held. The vocabulary narrows. The same constructions start to repeat more often, the same handful of gestures, the same sentence shapes, because the machine is falling back on its highest-probability moves and the highest-probability moves are the most generic ones.

The tells get worse, in other words, the longer it runs. The held breaths multiply. The rule of three tightens its grip. The paragraphs get more uniform. Everything this series has described gets denser toward the back of a long generation, not because the machine is trying less but because sustained quality requires holding many things in mind at once, and the machine’s capacity to hold is finite and gets consumed, and what is left when it is consumed is the default, and the default is the soup of every tell at maximum concentration.

This is why the annotated example in this series, the machine-written story with every tell marked, gets denser toward the end. That is not an artifact of the marking. That is the drift, visible. The early chapters carry tells but breathe a little. The late chapters are nearly solid with them, because by then the machine has drifted fully into its default, and the default is where the tells live.

Drift’s cousin: hallucination

There is a second failure that travels with drift and is so closely related that they are easy to confuse, but they are not the same thing, and the difference is worth holding clearly. Drift is the machine losing track of a fact it established. Hallucination is the machine inventing a fact that was never established at all, and asserting it with total confidence. Drift forgets. Hallucination fabricates. One is amnesia, the other is confabulation, and a long machine draft is riddled with both.

The reason they travel together is that they share a root. The machine does not know things; it produces text that is locally plausible. When it cannot recall what it established, it does not say so, because it has no sense of the boundary between what it knows and what it is making up. It simply generates the most plausible-sounding continuation, and if that continuation invents a detail, the invention arrives wearing exactly the same confidence as a fact. There is no hedge, no flag, no tonal difference between the machine reporting something true and the machine making something up. That flatness of confidence is the thing to watch for, because it means you cannot trust the texture of authority. The machine sounds equally certain when it is right and when it is hallucinating.

Hallucination in fiction

In a novel, hallucination shows up as invented internal facts that the story never set up and often cannot support. A character suddenly knows a thing they were never told, because the machine needed them to know it and forgot they had no way to. A history gets invented for a place, confidently and in detail, that contradicts the history established three chapters back. A rule of the world, how the magic works, how the technology works, what the threat can and cannot do, gets asserted in chapter ten in a way that quietly breaks the rule asserted in chapter four, not because the machine is revising, but because it generated a plausible-sounding rule in the moment without checking. Characters reference shared events that never happened. Objects appear in a room that was described as empty. The machine produces the texture of a coherent invented world while failing to actually track that world, so the details accrete into something that sounds rich and falls apart the moment you hold two of them up together.

The particular danger in fiction is that invented texture is often good. The hallucinated detail can be vivid, specific, evocative, exactly the kind of thing a strong writer might invent on purpose. So you keep it, because it reads well, and you do not notice that it contradicts something or that it commits the story to a fact the rest of the book cannot honor. The machine’s confident fabrication seduces you into building on sand, and three chapters later the invented thing has to be paid off and cannot be, because it was never real, it was just plausible.

Hallucination in nonfiction

In nonfiction the stakes are higher and the failure is more concrete, because now the invented facts are claims about the real world, and they are wrong. This is the failure that has ended careers and produced sanctioned legal filings, and it deserves separate and serious attention.

The machine invents citations. It will produce a reference, a book, an author, a journal, a page number, a court case, formatted perfectly and entirely fictional, because it has learned what a citation looks like and will generate one whether or not the source exists. It invents quotations, attributing to a real person words they never said, often words that sound like something they might have said, which is worse, because the plausibility is what gets it past you. It invents statistics, confident numbers with no source, because a number makes a sentence land and the machine is optimizing for a sentence that lands. It invents historical detail, dates and events and names that are close to real and not real, woven seamlessly into accurate material so that the false thread is invisible inside the true cloth. It will tell you, with complete assurance, a thing that is simply not so.

For anyone writing nonfiction with the machine, this is the failure that matters most, and the rule it forces is absolute: every fact the machine gives you is unverified until you have checked it against a real source, and every citation is presumed fictional until you have personally found the source and confirmed it says what the machine claims. Not most citations. Every one. The machine cannot tell you which of its references are real, because it does not know, because to it they are all just plausible strings. The work of fact-checking a machine draft is not a light pass. It is the assumption that nothing is true until proven true, applied to every claim in the document, and skipping it is how invented sources end up in published books and filed briefs under a real person’s name.

Why the author cannot see it

Drift is uniquely hard for the author to catch, and the reason is built into how machine writing gets made.

You generate a novel in pieces, in order, over days or weeks. You read each piece as it comes, in the context of the piece before it. So you experience the book the way you experience a slow decline you are living inside: relatively, never absolutely. Chapter nine seems fine because it is about as good as chapter eight, which seemed fine because it was about as good as chapter seven. At no point do you feel a cliff, because there is no cliff, only a slope, and you are walking down it one acclimating step at a time. The continuity slips slide past for the same reason. By chapter nine you do not clearly remember what color you established the eyes in chapter two, because you read chapter two weeks ago, so when the machine says brown you do not flinch, because you have drifted right along with it.

The only defense is the thing almost no one does, which is to read the whole book straight through, cold, after it is finished, the way a reader will. Set chapter twenty beside chapter one. Track every established fact in a list and check it. Hold your outline against the finished draft and find the place the structure reverted. This is real work, the unglamorous continuity-editor work that a human first-drafter does half unconsciously and the machine cannot do at all, and skipping it is why machine novels reach readers full of contradictions and dropped threads and a back half visibly worse than the front.

The seductive opening

There is a specific trap inside quality drift worth naming on its own, because it fools people into trusting these tools more than they should. The opening is the best part, and the opening is what sells you.

When you first try a machine on a piece of writing, you give it a short prompt and it gives you a short output, and that output is the machine at full strength, all its attention on a small space. It is often genuinely impressive. So you conclude the tool is good, and you scale up, and you ask for a novel, and the first chapter arrives and it is still impressive, because the first chapter is still the machine at something near full strength. You are sold. You commit to the project. And then, chapter by chapter, the thing degrades, but you are already committed, already invested, already telling yourself the dip in chapter twelve is just that chapter, already acclimated to a slide you cannot feel because you are inside it.

The opening is not representative and it never was. Judging a machine’s novel-writing by its opening chapter is like judging a runner’s marathon by their first hundred meters. Everyone looks fast at the start. The question fiction asks is whether you can hold it for twenty-six miles, and holding is precisely the thing the machine cannot do, and the opening is specifically the part of the race before holding matters. If you want to know what a machine novel will actually be, do not read its first chapter. Read its last one, and then read its first, and measure the distance between them. That distance is the drift, and the distance is the truth about the tool that the seductive opening was built to hide.

What drift looks like in practice

Let me make it concrete with the kind of thing that actually happens across a generated novel, so you know what to hunt for.

A character is introduced in chapter one as a cautious, soft-spoken accountant who has never left his home town. By chapter fourteen he is making confident speeches and is revealed to have traveled widely, not as a deliberate arc the book earned, but because the machine lost the original characterization when it slid out of the window and rebuilt him from scratch in the back half as a generic protagonist. No scene shows him changing. He is simply a different man, and the book does not notice.

A subplot about a missing sum of money drives chapters three through seven, generates real tension, and then vanishes. It is never resolved, never mentioned again, because by chapter nine the machine has forgotten the money existed. A reader reaches the end still waiting for it, and the wait is never paid, and the book does not know it owes anything.

The prose in chapter one varies its sentence length, holds a distinct dry voice, lets paragraphs run different sizes. The prose in chapter eighteen is smooth, mid-length, uniform, voiceless, every paragraph the same size, the dry voice entirely gone, replaced by the machine’s default register, because eighteen chapters in, the voice you asked for has long since slid out of the window and the machine is running on its defaults. Set the two chapters side by side and they read like different writers, because in the way that matters they are: chapter one is the machine holding your instruction, chapter eighteen is the machine that forgot it.

These are not exotic failures. They are what happens, reliably, to any long generation, and they are invisible to the author who read the book in pieces and never read it whole. They are the drift, and the drift is why the back half of a machine novel is where the dream of effortless authorship goes to die.

Drift and the one spine

There is a way these two ideas, drift and the single uniform spine from the spine article, are really the same idea seen from two angles, and seeing the connection is worth the trouble because it explains why long machine works fail so reliably.

The machine has one default spine, the tame, balanced, three-act, everything-resolves shape it produces when nothing is actively holding it to something else. A good prompt, a careful outline, a strong specified voice can hold it away from that default for a while. But holding requires the instruction to stay in the window, and over the length of a novel the instruction slides out, and once it is gone there is nothing holding the machine away from its default anymore, so it drifts back to it. Drift, in other words, is the machine relaxing toward its one spine over length. The structure reverts to three acts because three acts is the default and your nine-act plan drifted out of reach. The voice flattens to the average because the average is the default and your specified voice drifted out of reach. The chapters become uniform, the sentences cluster, the tameness sets in, all of it, because all of it is the default, and drift is just the name for the long slow fall back into default once the things holding the machine up have slid away.

This is why you cannot prompt your way out of it. People try. They write elaborate instructions, detailed outlines, careful voice guides, and these help, at the start, in the window. They cannot survive the length, because nothing the machine was told survives the length, and the back half always belongs to the default no matter how good the front-loaded instruction was. The drift is not a failure of your prompt. It is the gravity the prompt was briefly fighting, and gravity wins over distance, and a novel is a very long distance.

So the fix for drift is the fix for the spine is the fix for everything in this series, because they were always one fix. You, the human, hold the whole. You keep the outline in front of you and refuse the reversion. You track the facts the machine forgot. You vary what drifted into uniformity. You hold the voice the machine let go. You do, deliberately and consciously and with great effort, the holding that the machine cannot do at all, because you have the thing the machine lacks, which is not better technique but an actual continuous mind that contains the whole book at once instead of a window sliding over its surface. That continuous mind is what writes a novel. The machine does not have one, and over the length of a book, the lack of one is the whole story.

How drift makes the whack-a-mole worse

The spine piece describes the whack-a-mole feeling of fixing a machine book, where every layer you flatten reveals another underneath. Drift adds a second dimension to the game, because the moles are not evenly distributed. They multiply toward the back.

So you fix the early chapters, which were the best to begin with, and you feel good, and then you reach the back half and the density of problems climbs and keeps climbing, and the continuity errors start, and the structure has reverted, and you realize the work is not linear, it is getting harder as you go, exactly when you have the least energy left for it. The back half of a machine novel needs the most rebuilding and is the part the author is most exhausted by, and that mismatch is why so many machine-assisted novels are published with strong openings and collapsing endings. The author fixed what they had the energy to fix, which was the front, and let the drifted, degraded, contradiction-riddled back half through because they were done, and the back half is where the drift had piled all its moles.

The fix is the same as it always is

You cannot edit drift out any more than you can edit out the spine, because drift is not a set of local errors, it is the cumulative consequence of how the thing was made. You can fix individual continuity slips once you find them, and you should, but the quality drift, the slide into the default register, the densifying tells, the reverted structure, these are not slips. They are the machine being the machine at length, and the only real fix is the one every piece in this series arrives at: stop trying to repair the generated thing and rebuild it, holding the whole in mind the way the machine never could, paying the debts it dropped, varying what it made uniform, and refusing the default it drifted into.

That is the whole series in the end. Six layers of tells from the surface to the spine, and a seventh that runs through all of them and gets worse over length, and underneath every one of them the same fact: the machine writes from a window and an average, and a book is not a window or an average, a book is a held whole, and holding the whole is the work, and the work is yours, and it always was.