"Delve" is dead. The word that spent 2 years as shorthand for "a robot wrote this" peaked at 28x its human baseline in 2023-24 academic writing, then collapsed to roughly 1 appearance per 1,000 ChatGPT conversations by mid-2025. It got trained out, and the timing says the punchline is what killed it.
Most people never updated. The popular method for spotting AI text is still a word list absorbed from jokes in 2023: delve, tapestry, the em dash. Every entry on that list is now patched or dying, and some of it is migrating into human speech. Meanwhile the real fingerprint of machine writing sits in a layer no lab has patched, and probably none will.
A tell is any habit that gives text away as model output. I read AI-assisted writing all day, in cold emails, CRM notes, and drafts, and I keep a checklist of the tells that still hold up. It's on version 4.8 because the entries keep dying. Tells have shelf lives, and the pattern behind them is worth learning before you accuse a coworker over a dash.
01The vocabulary has vintages
Wikipedia's editors keep the best-maintained public list, Signs of AI writing, because they deal with machine-written submissions every day. Read it closely and you'll find something most tell lists skip: dates. The guide era-tags its vocabulary. "Delve," "tapestry," and "testament" read as 2023-24. "Fostering," "align with," and "showcasing" read as 2024-25. "Emphasizing," "enhance," and "highlighting" read as 2025 and later.
The dialects go deeper than the eras. Grok leans on "causal" and "empirical." GPT-4o's creative writing had a weakness for "camaraderie" and "palpable," and Claude opened a thousand code reviews with "You're absolutely right" until the tic was gone at the end of 2025.
Handed a paragraph of raw model output, you can estimate when it was generated, and often by which model family.
Put the eras and the dialects together and unedited AI text turns out to be datable. Handed a paragraph of raw model output, you can estimate when it was generated, and often by which model family, the way a geologist reads strata. Every word list is perishable from the day you learn it, and Wikipedia's editors are the only ones I've seen date theirs.
02Patches follow punchlines
Why did the words die? Embarrassment, mostly. "Delve" collapsed after it became a meme. The em dash is the cleaner case, because the fix shipped in public: on November 13, 2025, Sam Altman announced that ChatGPT would finally obey a standing instruction to skip em dashes, calling it a "small-but-happy win", while OpenAI's social accounts ran a joke apology. Look at what actually shipped, though. Obedience. The default never moved, and out of the box ChatGPT still writes em dashes today. The most famous tell in the world, after 2 years of screenshots, earned a settings toggle.
The rule underneath: a tell's lifespan is set by how easy it is to screenshot. Mockery is the only style feedback the labs reliably act on, a word fits in a screenshot, and nothing that can't be memed has ever been patched. By the time a tell is famous, its patch is usually scheduled.
The most famous tell in the world, after 2 years of screenshots, earned a settings toggle.
My own copy rules ban the em dash anyway, toggle or no toggle. An em dash in a draft here means a generated passage got through without editing, and since the default never moved, that check still works.
03The leak runs in both directions
While the models shed their vocabulary, people picked it up. A Max Planck group measured spoken academic English, lectures and podcasts, and found "delve" up 48% within 18 months of ChatGPT's launch, with "adept" up 51%. Nobody on those recordings was hiding machine text. Vocabulary spreads by exposure.
So the word list fails from both ends at once: the models retired the words while people were still adopting them. An accusation fired off a 2023 list in 2026 mostly hits people. The colleague who always liked long dashes. The non-native speaker whose careful formal register reads as "AI" for the same reason it trips automated detectors.
The output moved in the opposite direction from the stereotype, too. The 2023 giveaway was bloat: wordy and formal, hedged to death. Current models write shorter and more casual, with measurably narrower vocabulary than the generations before them, and the giveaway now is manufactured punchiness: short fragment runs, paragraphs that each land a tidy punch.
The stakes went up at the end of July, when LinkedIn added a report option for content that "seems like AI slop." Every reader just got a button, and the list most of them will use it with is 3 years old.
04The layer no patch reaches
The research on where tells come from keeps landing in the same place. A 2025 study (from COLING, a computational-linguistics conference) ruled out the training data as the source of the vocabulary; the words emerge in instruction tuning, the stage that turns a raw model into a helpful assistant. Reinhart and colleagues measured the same thing at the grammar level: base models, the raw thing before assistant training, write close to human rates, and the tuned versions come out the other side attaching the trailing "-ing" analysis clause ("...creating a lively community") at 5.3x the human rate. A 2024 study found models repeating grammatical templates at double the human rate even when every word changes.
Read those results together and the em dash patch looks like what it was: cosmetics. The vocabulary was never the fingerprint. The durable layer is structural, and it exists because instruction tuning optimizes for whatever human raters reward, which is text that feels confident and complete. The clearest product of that incentive is the contrast reframe, "it's not this, it's that," a shape that feels like insight while asserting nothing checkable. The Washington Post analyzed 328,744 ChatGPT conversations and found the "not just X, but Y" construction in 6% of them, across every kind of conversation.
Un-training the reframe means making the product feel worse to the raters who steer it.
No patch is coming for that layer, because the reframe is exactly what raters click thumbs-up on. Un-training the reframe means making the product feel worse to the raters who steer it. The same goes for the tidy punch landings and the "-ing" tails. They are the helpfulness, as the training defines it.
This is also why "humanizer" tools disappoint. They do the vocabulary half, swapping the 2024 words for quieter ones, and leave the skeleton intact. Turnitin, the plagiarism checker schools run essays through, formalized the distinction in early 2026 with a separate detection category: "AI-paraphrased."
05What still works
Detection moved up a layer instead of dying. The strongest result of 2025: researchers found that people who use ChatGPT heavily for writing are the best detectors anyone has measured. A 5-person majority vote of them misclassified exactly 1 article out of 300. That beat nearly every commercial detector, and the group's accuracy held against paraphrased and "humanized" text where the software collapsed. What those readers said they used was structure and originality more than any word.
So the practical update is small. Treat any word list as dated the day you learn it, and note the date. Read for the durable layer instead: the uniform rhythm, and the reframe that sounds like insight and asserts nothing. My checklist carries version numbers for exactly this reason, and the 2023-era entries are the ones I trust least.
The next model release will retire a few more words, and the jokes will find new ones. Before you call something AI off a single word, check the date on your evidence.