The Accent Nobody Designed

In unscripted science and technology podcasts, the word delve rose about 44% above where it should have been, a year to a year and a half after ChatGPT came out. By the spring of 2026 it sat below that line. Hiromu Yakura and colleagues measured both moves across 737,083 hours of conversation from 824,634 podcast episodes, comparing the word against a counterfactual built from words that had behaved the same way before ChatGPT (Yakura et al. 2026). The rise is solid. The fall is clear only across all podcasts together, and the timing points to one explanation: by then delve had become public shorthand for chatbot writing.

That arc is the short life of an AI tell, and it matters to anyone who publishes. A tell is a cheap surface cue readers use to guess at something they can’t see, in this case whether a person did the work. Language models made the cue cheap to produce and just as cheap to remove, so it is losing its power to tell writers apart. The discount readers apply when they believe something is machine-made is a separate thing, and the research on it points somewhere else. It was never about the words. It prices a missing person, and it shrinks when that person can be checked.

I covered the written side of this, the vocabulary studies and the detectors that misfire on non-native writers, in The AI House Voice. This note picks up what that essay leaves open: speech, and what the discount responds to.

How a chatbot word gets into a mouth

The podcast data shows that the words spread. An experiment in the same paper shows how little it takes. Yakura’s team had 496 people describe images while chatting with a bot that favored one of two interchangeable words. During the chat, people used the bot’s word 61% of the time and the alternative 17%. The habit survived a distraction task and showed up again when people had to choose between the two words, and only 15 of the 496 noticed the pattern at all. A few minutes of conversation moved a word into people’s own vocabulary without their knowing.

Two questions people ask most have no answer in the data yet. Nobody has measured whether these words reach people who never use chatbots, passed on by the people who do, and I found no study of children. The paper models second-hand spread, which is a simulation, not a measurement.

Then the word became a liability

The same word that marked a machine soon marked anyone who used it. Mingmeng Geng and Roberto Trotta tracked words that ChatGPT favors in academic abstracts. Delve and its kind began falling soon after they were publicly pointed out in early 2024, while other model-favored words such as significant kept rising (Geng and Trotta 2025). A change in the models would have shown up later, so they read the drop as reputational: the words had picked up a bad name, and authors edited them out.

The models moved too. In Yakura’s comparison, the over-use of delve fell from roughly 380 to 1 under GPT-3.5-turbo to 40 to 1 under GPT-4o, and under GPT-5 the word appears less often than it does in human writing. The vocabulary tell is squeezed from both ends. People absorb a word, then avoid it once it counts against them, while the labs tune it away.

“Delve” in unscripted podcasts, against its counterfactual

A year to 18 months after ChatGPTNovember 2025 to April 2026

1. Science and tech+44% (95% interval +22% to +63%), then −15% (−35% to +8%)
2. Education+32% (95% interval −5% to +75%), then −11% (−40% to +31%)
3. Business+31% (95% interval 0% to +67%), then −30% (−50% to +3%)
4. All podcasts+9% (95% interval −15% to +32%), then −35% (−57% to −7%)
5. Sports−7% (95% interval −40% to +41%), then −38% (−68% to +13%)

−60%−30%0+30%+60%

Readers were never good at this anyway. Maurice Jakesch, Jeffrey Hancock and Mor Naaman found that people identified AI-written self-presentations with 50 to 52% accuracy, a coin toss, and that the cues they relied on, such as first-person pronouns, contractions and talk of family, misled them (Jakesch et al. 2023). They proposed giving AI text a designed accent so people could recognize it. An accent did form, by accident, and it lasted about a year.

The discount was never about the words

If the words were the problem, removing them would fix it. The experiments that hold the words constant and change only the label say otherwise.

Sacha Altay and Fabrizio Gilardi showed people in the US and UK headlines with and without an “AI-generated” label. The label lowered how accurate and shareable the headlines seemed by 2.66 percentage points, a third of the effect of a “False” label, and most people read it as meaning AI wrote the whole article (Altay and Gilardi 2024). When the label was defined as lighter AI help, its effect shrank.

Yidan Yin, Nan Jia and Cheryl Wakslak found that AI-written replies made people feel more heard than human ones, until the label appeared. The same AI text scored 6.12 on a seven-point scale when people thought a person wrote it and 5.41 when they were told it came from AI (Yin et al. 2024). The words were identical. The belief about who wrote them moved the score.

Manav Raj, Justin Berg and Rob Seamans ran 16 preregistered experiments with 27,491 people and found that disclosing AI authorship lowered ratings of creative writing by 6.2% on average, a small effect that none of their fixes reliably removed and that ran through perceived authenticity (Raj et al. 2026). In a working paper on online reviews, Samsun Knight, Yakov Bart and Minwen Yang found that reviews a detector flagged as AI-assisted got 16.6% of a standard deviation fewer “useful” votes on Yelp and 14.2% fewer “helpful” votes on Amazon, and the penalty came entirely from reviews without expert badges (Knight et al. 2025).

Across these studies the discount is largest where the value is supposed to come from a person: feeling heard, a story, an experience someone vouches for. It is smaller where content is informational and can be checked. It is also not a general hostility to machines. A meta-analysis of 440 effect sizes from 76,142 people found that consumer aversion to AI is modest on average and varies a great deal by domain and by how the AI is labeled (Zehnle et al. 2025).

Suspicion is enough

The discount doesn’t need a label. In a study of conversations where one partner could use AI-suggested replies, Jess Hohenstein and colleagues found that people’s guesses about their partner’s AI use matched actual use only weakly, a correlation of 0.22, yet partners suspected of using AI were judged more negatively (Hohenstein et al. 2023). The suspicion was mostly wrong, and it still cost the person suspected.

That is where lists of tells become dangerous. Robert Chesney and Danielle Citron named the “liar’s dividend” for deepfakes: once people learn that fakes exist, real material becomes easier to dismiss, and “this dividend flows, perversely, in proportion to success in educating the public” (Chesney and Citron 2019). Lists of AI tells educate the public in exactly this way. The more people learn them, the easier it becomes to doubt any text, human or not. Applying their idea to writing is my extension, but the mechanism carries over without strain.

Why the tell can’t hold

Economics has a name for this. In Michael Spence’s model of job-market signaling, a signal separates people only if it is costlier to produce for those who lack the quality it signals (Spence 1973). A word anyone can add or remove for free fails that test, so it stops sorting writers no matter how confidently readers use it.

Hiring shows the same shift. When an AI tool for writing cover letters launched on a large online labor platform, the link between how closely a cover letter matched the job and whether it got a callback fell by 51%, and employers turned to other signals, including workers’ past work histories (Cui et al. 2025). When writing became cheap, attention moved to what couldn’t be written.

What cancels the discount

The same studies show what works, and none of it is style.

Sources. Benjamin Toff and Felix Simon found that an AI label lowered trust in a news outlet from 5.9 to 5.5 on an eleven-point scale, and that the negative effect was largely counteracted when the article listed the sources used to produce it (Toff and Simon 2025).

A stated human role. In Altay and Gilardi’s headlines, the penalty shrank when readers were told the AI’s part was smaller than writing the whole thing.

Verified experience and vetted reputation. In Knight’s experiments, the penalty on reviews labeled as AI-written largely disappeared when the review was described as based on a verified customer experience, and in the field data it never touched reviewers with expert badges.

A track record. In Cui’s hiring data, employers moved to work history.

Watermarks belong on this list only halfway. Google DeepMind’s SynthID-Text hides a statistical pattern in text as a model generates it, and a live test on nearly 20 million Gemini responses found no loss of quality (Dathathri et al. 2024). A watermark can show that a cooperating model produced a text. By design it can never show that a person did, which is the thing the discount is about.

At the desk

The practical rule for a project is to stop policing style and publish proof. Banning words buys a few months, until the next list circulates or the models change. What holds is checkable: claims linked to their sources, named people with stated roles (who wrote it, who checked it, what the AI did), and a visible record of past work. That is also why every note here carries a byline and its references, and why the source map exists.

A reader who suspects a machine is asking a question about accountability. The answer that holds up is a person they can check.

How this could be wrong

Three things could undo the argument. Jenna Russell and colleagues found that people who use ChatGPT heavily for their own writing spot AI text accurately, using cues beyond vocabulary (Russell et al. 2025), so tells may be noise for casual readers and still informative for experienced ones. The evidence is also uneven: the fall of delve is statistically weak, Yakura’s paper is a revised preprint, and the review and hiring studies are working papers. And the discount itself may fade as AI help becomes ordinary. If it does, proof matters less, though a byline and a source list cost little to keep.

References

1. Yakura, Hiromu, et al. "Empirical Evidence of Large Language Model's Influence on Human Spoken Communication." arXiv preprint 2409.01754v4 (2026).
2. Geng, Mingmeng, and Roberto Trotta. "Human-LLM Coevolution: Evidence from Academic Writing." Findings of the Association for Computational Linguistics: ACL 2025 (2025): 12689–12696. · doi:10.18653/v1/2025.findings-acl.657
3. Jakesch, Maurice, Jeffrey T. Hancock, and Mor Naaman. "Human Heuristics for AI-Generated Language Are Flawed." PNAS 120, no. 11 (2023): e2208839120. · doi:10.1073/pnas.2208839120
4. Altay, Sacha, and Fabrizio Gilardi. "People Are Skeptical of Headlines Labeled as AI-Generated, Even If True or Human-Made, Because They Assume Full AI Automation." PNAS Nexus 3, no. 10 (2024): pgae403. · doi:10.1093/pnasnexus/pgae403
5. Yin, Yidan, Nan Jia, and Cheryl J. Wakslak. "AI Can Help People Feel Heard, but an AI Label Diminishes This Impact." PNAS 121, no. 14 (2024): e2319112121. · doi:10.1073/pnas.2319112121
6. Raj, Manav, Justin M. Berg, and Rob Seamans. "The Artificial Intelligence Disclosure Penalty: Humans Persistently Devalue AI-Generated Creative Writing." Journal of Experimental Psychology: General 155, no. 4 (2026): 896–915. · doi:10.1037/xge0001889
7. Knight, Samsun, Yakov Bart, and Minwen Yang. "Generative AI and the Perceived Quality of User-Generated Content: Evidence from Online Reviews." SSRN working paper (2025). · doi:10.2139/ssrn.4621982
8. Zehnle, Meike, et al. "Not All AI Is Created Equal: A Meta-Analysis Revealing Drivers of AI Resistance." International Journal of Research in Marketing 42, no. 3 (2025): 729–751. · doi:10.1016/j.ijresmar.2025.02.005
9. Hohenstein, Jess, et al. "Artificial Intelligence in Communication Impacts Language and Social Relationships." Scientific Reports 13 (2023): 5487. · doi:10.1038/s41598-023-30938-9
10. Chesney, Robert, and Danielle Citron. "Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security." California Law Review 107 (2019): 1753–1820. · doi:10.15779/Z38RV0D15J
11. Spence, Michael. "Job Market Signaling." Quarterly Journal of Economics 87, no. 3 (1973): 355–374. · doi:10.2307/1882010
12. Cui, Jingyi, Gabriel Dias, and Justin Ye. "Signaling in the Age of AI: Evidence from Cover Letters." arXiv preprint 2509.25054 (2025).
13. Toff, Benjamin, and Felix M. Simon. "'Or They Could Just Not Use It?': The Dilemma of AI Disclosure for Audience Trust in News." International Journal of Press/Politics 30, no. 4 (2025): 881–903. · doi:10.1177/19401612241308697
14. Dathathri, Sumanth, et al. "Scalable Watermarking for Identifying Large Language Model Outputs." Nature 634, no. 8035 (2024): 818–823. · doi:10.1038/s41586-024-08025-4
15. Russell, Jenna, et al. "People Who Frequently Use ChatGPT for Writing Tasks Are Accurate and Robust Detectors of AI-Generated Text." Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (2025): 5342–5373. · doi:10.18653/v1/2025.acl-long.267
16. Vaduvescu, Claudia. "The AI House Voice: What a Style Gate Actually Removes." claudiavaduvescu.com, 2026.