# Measured model idiolect Part of [The AI Tells Index](https://feedsquad.com/ai-tells). A tell signals low effort. It does not identify an author. Skilled writers produce every shape listed here, some of them daily, and automated detectors misread those writers for it at rates measured above 60 percent on non-native English prose. Nothing in this index proves that a machine wrote anything. Read an entry as one piece of evidence to weigh against the false-positive notes printed beside it. ## Facts - Id: measured-model-idiolect - Category: Model register (model-register) - Subcategory: idiolect - Also known as: five-way model classification, product fingerprint, per-model style signature - Status: Contested. Credible people dispute that this signals anything at all. - Severity: low - Evidence grade: peer-reviewed (a published study measures the pattern) - Languages: en - Added: 2026-08-15 - Updated: 2026-08-15 - Page: https://feedsquad.com/ai-tells/measured-model-idiolect ## Description Model families are separable from their output at corpus scale. A classifier trained on text alone told five chat products apart at 97.1 percent in an ICML 2025 study, and the signal survived rewriting and translation. That result is about distributions over many samples. It licenses nothing about the document in front of you, and this entry is in the directory to say so. ## Why it reads as machine-written Nothing here reads as a named product to a human reader. The finding is real and the popular version of it is not. Three limits travel with the measurement. It is aggregate, so a per-document call has no support in it. It is dated to the snapshot the authors collected, and vendor prompts move between releases. What it separates is products. The chat surface adds formatting that the same weights do not produce through an API: the OpenAI cookbook states that GPT-5 emits no Markdown by default there. A bullet-heavy layout therefore records where a text was generated rather than what generated it. ## Detection Type: statistical (a measured metric against a threshold with a stated basis) Metric: model-family-classifier-accuracy-corpus-level Threshold basis: > No document-level threshold exists, and this entry sets none on purpose. The published number is 97.1 percent five-way accuracy in Sun and colleagues, ICML 2025, from a machine classifier over a dated corpus with many samples per product. Turning a corpus-level classifier score into a per-document cutoff would invert what the study measured. Treat this entry as a boundary marker on the index rather than as a rule to run over a text. ## Examples Constructed specimens. Written for this index. Never quoted from anyone's posts. ### Specimen 1 Before, exhibiting the tell: > Overall, there are several factors to consider when choosing a project tracker. First, consider your team size. Second, consider your reporting needs. Overall, the right choice depends on your workflow. After, repaired: > We switched trackers in March 2026, when the sprint board passed 400 open cards and the weekly export started taking nine minutes. Card count decided it. Reporting was the tie-breaker and took ten minutes to check. Note: The specimen shows the register that classifiers pick up in aggregate. It is not an identification of any product, and no reader should treat it as one. Figures in this repair are invented for the specimen. ### Specimen 2 Before, exhibiting the tell: > This response is concise and direct, and it avoids unnecessary formatting while covering the key considerations you raised about hiring. After, repaired: > Hire the second candidate. She has shipped the exact integration twice, and our last two hires without that experience each took about five months to get productive. Note: A different surface register, repaired the same way, by stating a decision and the evidence behind it. Figures in this repair are invented for the specimen. ## False positives Who legitimately writes this way. Plain-prose human writers land inside any of these profiles by coincidence, and the study measured a machine classifier, not a reader. A classifier score built over thousands of samples per product does not transfer to one paragraph, and a person has neither the corpus nor the training that the classifier had. Sparse formatting, short openings and analytic vocabulary are also what a careful human editor produces on a deadline. Treat any claim that a given text came from a named product as unsupported, including a claim made with this entry open. The honest use is the reverse of the popular one: it explains why the per-document version fails. ## Model attribution The ICML 2025 study classified output from five chat products: ChatGPT, Claude, Grok, Gemini and DeepSeek. Reported tendencies differ by product, including sparser formatting in some and heavier emphasis markup in others, all snapshot-dated and all a property of the deployed surface as much as of the weights behind it. Anthropic dated system prompts and the OpenAI cookbook both document instructions that move these surfaces between releases, which is why per-product detail belongs in one dated entry instead of a page per vendor. ## Sources 1. Sun et al., Idiosyncrasies in Large Language Models, ICML 2025 (arXiv:2502.12150) https://arxiv.org/abs/2502.12150 (tier: peer-reviewed; accessed 2026-08-14) 2. Anthropic dated system prompts, release notes https://platform.claude.com/docs/en/release-notes/system-prompts (tier: vendor; accessed 2026-08-14) 3. OpenAI GPT-5 prompting guide, cookbook https://developers.openai.com/cookbook/examples/gpt-5/gpt-5_prompting_guide (tier: vendor; accessed 2026-08-14) 4. Wikipedia: Signs of AI writing https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing (tier: community; accessed 2026-08-14) ## Status history Ids are permanent. A retired tell keeps its id and its page. - 2026-08-15, Contested: Shipped as contested on 2026-08-15. The underlying measurement is peer-reviewed and strong at corpus level. What is disputed is the inference people draw from it, that one document can be traced to a named product, and that inference has no support in the paper it cites. The entry exists to state the boundary. Snapshot-dated to the ICML 2025 collection, and the two vendor documents cited show the chat surface changing release by release. ## License CC BY 4.0. https://creativecommons.org/licenses/by/4.0/ Attribution: The AI Tells Index, feedsquad.com/ai-tells Reuse the data, including commercially. Keep the attribution line. --- Dataset version 1.0.0. Schema version 1. Part of [FeedSquad](https://feedsquad.com). Built by [Herman Foundry](https://hermanfoundry.com) from Levi, Finnish Lapland.