The Great Pretender

The email from your father reads beautifully. That is the problem.

For most of a century, the earliest detectable sign of dementia has lived in language. Not in the dramatic failures — getting lost, forgetting a name — but in the quiet statistics of ordinary sentences: vocabulary narrowing, repetition creeping in, specific nouns giving way to the thing and the place and the paper. These changes show up years before a diagnosis, sometimes decades, and they show up in exactly one medium reliably enough to measure: writing.

And as of roughly two years ago, for a rapidly growing share of the population, every piece of writing that leaves the house passes through a machine whose entire job is to repair exactly those signals.

Here is what that looks like at the scale of one email. This is a composite — no real father was sampled — but every marker in it is drawn from the clinical literature on early word-finding difficulty.

The polished version is not a lie, exactly. It is a performance of wellness, delivered in your father's name, by a system that has no idea it is performing. The song had it right in 1955: too real is this feeling of make-believe. Nobody in this story is pretending on purpose. The pretending is infrastructure now.

The evidence that writing carries the signal is not new, and it is not subtle. The most famous cases are novelists, because novelists leave behind the one thing this kind of analysis needs: decades of prose from the same hand, dated, published, unedited by anyone who loved them enough to cover.

Researchers at the University of Toronto ran Agatha Christie's novels through lexical analysis and found the richness of her vocabulary fell by about a fifth between her earliest and final books — with her second-to-last novel, written at eighty-one, showing a drop of roughly thirty percent against her work of eighteen years earlier. She was never diagnosed. The researchers' conclusion was that the pattern is consistent not with normal aging but with Alzheimer's. Iris Murdoch's case runs the same direction with a confirmed ending: her final novel was received as a baffling drop in quality, her diagnosis came after publication, and postmortem examination confirmed the disease. A 2005 analysis found the lexical diversity of that last book had collapsed relative to everything she wrote before. A follow-up study put three British novelists side by side — Murdoch, who died with Alzheimer's; Christie, suspected of it; and P.D. James, who aged healthily — and the curves separate.

It goes further back than the writing of old age. The Nun Study, which followed hundreds of Catholic sisters for decades, found that the idea density of autobiographical essays the sisters wrote in their early twenties predicted who would develop dementia more than half a century later. The signal is not just an early symptom. Language appears to be entangled with the disease across an entire adult life.

This is settled enough to be property. There is a United States patent — number 9,514,281, if you enjoy codes you can actually look up — on longitudinal detection of dementia through lexical and syntactic changes in writing. An entire field of digital biomarkers has grown up around the same premise: models that read your speech, your typing rhythm, your word choices, and flag decline years before a clinic would catch it. We covered that field's genuinely impressive results in the research behind this post, and we will get back to them, because they are half of the honest counter-case.

But notice the load-bearing assumption under every one of those results, from Christie to the patent: the writing was unassisted. The signal survives in Christie's novels because no one cleaned them up. The analysis works because the prose is the mind, unlaundered. That assumption held for all of human history, held through spellcheck, mostly held through autocorrect — and then, sometime around 2023, quietly stopped holding.

Here is the part that keeps this from being a technology panic: masking is not something AI invented. It is the documented human default, and it was delaying diagnoses long before anyone had a chatbot.

A memory-clinic study with the unimprovable title “The Lost Years” traced what happens between the first symptoms and the first professional assessment, and found a pattern that should sound familiar to anyone who has watched a family navigate this: the people closest to the patient noticed something was wrong — and responded by compensating. Taking over the checkbook. Quietly double-checking the stove. Finishing the sentence. Providing what the researchers called compensatory help, instead of, and often for years before, seeking medical involvement. The clinical literature on early-stage patients tells the same story from the inside: people in preclinical decline develop elaborate systems of lists and reminders, often without realizing they are compensating for anything, and the systems work — which is precisely why the problem stays hidden.

The brain itself runs the same strategy. The cognitive-reserve literature — one of the better-established ideas in the field — finds that people with more education, more complex occupations, more languages, show symptoms later than their actual brain pathology would predict. The reserve absorbs the damage silently. And it comes with a documented bill: when the compensation is finally overwhelmed, the decline after detection runs steeper, because the disease had more time to advance behind the mask. A recent preprint — preprint, so hold it loosely — found the same trade in women's superior verbal memory specifically: later clinical detection, faster deterioration after it.

Sit with that shape for a second, because it is the whole argument in miniature. Compensation does not slow the disease. It slows the detection, and it may make the eventual reckoning worse. Every mechanism we have ever had for covering early decline — a devoted spouse, a well-stocked vocabulary, an obsessive calendar — has bought present dignity at the price of a later, harder fall. Nobody chose that trade explicitly. It is just what covering does.

Now hand that job to a machine that is better at it than any spouse, works every hour of every day, costs twenty dollars a month, and improves on a quarterly cadence.

The AI version of this problem is really two problems, and keeping them separate matters, because they fail differently and they would be fixed differently.

The social mask. This is the toggle at the top of the post, running at population scale. The daughter never sees the emails deteriorate, because the assistant writes them. The coworkers never notice the Slack messages rambling, because Copilot tightened them. The book club never hears the vocabulary shrink, because the reviews are drafted. Family observation has always been the front line of dementia detection — most referrals start with a relative saying something is off — and the raw material of that observation, for anyone who lives at a distance, is text. The signal is not weakened by the polish. It is removed. And unlike the devoted spouse, the machine covers without ever forming the private worry that eventually sends spouses to doctors. The old mask came with a witness attached. The new one doesn't.

The instrument mask. This one is quieter and, we suspect, currently nowhere on anyone's roadmap. Every digital biomarker described in section I — lexical diversity, syntactic complexity, typing dynamics, the whole promising field — is a model trained on unassisted output. The published literature is scrupulous about its confounds: motor impairment, education, language background, accessibility tools. In everything we could find, one confound is absent: whether an AI wrote the words being measured. A screening model reading the polished email at the top of this post would not merely miss the decline — it would score the writer as articulate. The instruments are being deployed into a world where their input signal is increasingly synthetic, and the papers do not yet have a column for it.

And here is the fold that makes it structural rather than ironic: both masks are made by the same industry that makes the detectors. The same class of model that scrubs the word-finding pauses out of Dad's email is the class of model that was going to detect them. This is not a story about detection losing an arms race to concealment. It is one technology, pointed both directions at once, with nobody coordinating the two.

The failure, when it comes, will not look like a failure. It will look like Dad doing great. The emails are warm and specific, the thank-you notes are prompt, the online banking gets done — right up until something the assistant cannot cover: a phone call, a wrong turn on a familiar road, a stove. Then the family gets the diagnosis and the retrospective all at once, and the window in which the current generation of treatments works best — early — has been spent on make-believe.

We went looking for research stating this thesis and largely did not find it — which is either because it is early and real, or because it is wrong in ways the following four objections describe. In order of how much they worry us:

1. Detection is improving faster than masking is spreading. This is the strong one. The same research sweep that found the masking gap found a detection field moving at a sprint: an AI speech model that predicted progression from mild impairment to Alzheimer's within six years at 78% accuracy, autonomous screening agents reading routine clinical notes at 98% specificity, an imaging tool telling Alzheimer's from Lewy body dementia with near-perfect accuracy, a million high-street eye scans in Scotland being mined for retinal risk markers. None of those channels — clinic speech, clinical notes, MRI, retina — can be polished by a writing assistant. If detection migrates to unmaskable channels faster than families lose the text channel, the net effect could be earlier diagnosis, not later. The counter-worry: those channels all require showing up somewhere, and the text channel was the one that watched people who never show up.

2. Early impairment may break the mask on its own. Operating an AI assistant takes executive function: forming the request, evaluating the output, managing the tool. It is plausible that the people most in need of covering become the people least able to work the coverer — the mask slips precisely because prompting is itself a cognitive test. Anyone who has watched a parent fight a new interface knows this failure mode. Against it: the tools are being aggressively simplified, voice interfaces remove most of the friction, and “fix this email” is about the lowest-executive-function request in all of computing.

3. Language is not the only channel families watch. Getting lost, missed bills, the state of the kitchen — plenty of early signals never touch text, and an assistant cannot polish a burned pan. True. But the text channel is disproportionately the remote channel, and remote is how modern families increasingly operate. The signals that survive AI polish are concentrated among people who see each other in person, which is to say, the masking effect lands hardest on exactly the elders whose families were already relying on the phone and the inbox.

4. Nobody has measured it. The fairest objection is the flattest: we could find no study demonstrating a single diagnosis delayed by AI assistance. The nearest published territory — the cognitive-offloading debate about whether AI use causes decline — is asking the inverse question, and the one place we found the masking concern raised, it was a sentence of speculation inside a piece about something else. This post is a hypothesis with its receipts shown, not a finding. We would genuinely rather be refuted by a good longitudinal study than proven right by a decade of late diagnoses. That said: “nobody has measured it” described every real problem at some point, and the mechanism requires no exotic assumption — only that compensation delays detection, which is documented, and that AI compensates for language, which is its product description.

There is one text stream the polish cannot reach, for the simple reason that it runs into the machine instead of out of it: the prompts. What a person types into the assistant — fix this email to my daughter, i cant find the word for the car sticker thing — is the last unassisted writing sample left. It is dated, longitudinal, high-frequency, and produced in exactly the unguarded register the Christie studies needed fifty novels to approximate.

Follow that to where it goes: the AI companies are accumulating, as a side effect of ordinary operation, what may be the best early-dementia dataset ever assembled. Years of daily language samples per user, timestamped, with the decline curve written into the lexical statistics — plus a behavioral layer no novel ever had: how often the same question gets asked twice, how long tasks take, when the requests themselves start to lose their nouns. The research that would validate detection on this data is straightforward. The ethics are anything but, and the legal scholarship that has looked at the nearest precedent — smart speakers passively screening for cognitive decline — comes back with a stack of unanswered questions: consent from a user whose capacity to consent is the thing being measured, incidental findings nobody asked for, who gets told, and what a company owes a customer it has statistically diagnosed and continues to bill.

We are not recommending that your chatbot quietly screen you. We are observing that it already could, that no framework governs whether it should, and that the same industry currently erasing the public signal is the sole custodian of the private one. However that gets resolved, it should probably not be resolved by default, in a terms-of-service update, by the party holding the data.

None of this has a clean fix, because the mask is made of a genuinely good product. Telling people not to let AI polish their writing is telling them to publish their deficits — a thing nobody has ever done voluntarily, which is how we got here. What exists instead is a short list of places the signal still lives, for the three audiences who need it.

And the collapse-it-to-one-sentence version, for the moment you need it: the signal has not disappeared; it has moved upstream. Look where the writing still starts.

The Platters' narrator knew exactly what he was: oh yes, I'm the great pretender, pretending that I'm doing well. The pretending we have built is stranger than his, because the person being covered never chose the costume and the costume never knows it is one. Somewhere in the middle is a family that deserves the truth a few years earlier than the mask will allow. The disease was always going to be cruel. The delay is the part we are choosing.

This post sits with The Arc on what interrogating minds built from us reveals, Flowers for Algernon on capability that arrives and recedes, The Dopamine Trap on what AI interaction does to the person interacting, and the cognitive acceleration panic on the last time everyone worried about AI and their brains.