# Bolded lead-in on every list item 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: bold-lead-list-items - Category: Formatting (formatting) - Subcategory: list-shape - Also known as: bold lead-ins, inline-header vertical lists, key takeaways bullets, WP:AILIST, inline-header vertical list, bolded run-in head, key-takeaways list, key takeaways list, label-and-explain bullets, bold-first bullets - Status: Contested. Credible people dispute that this signals anything at all. - Severity: low - Evidence grade: corroborated (named independently by multiple credible secondary sources) - Languages: en - Added: 2026-08-14 - Updated: 2026-08-15 - Page: https://feedsquad.com/ai-tells/bold-lead-list-items ## Description Every item in a vertical list opens with a bolded phrase followed by a colon and an explanation. The bolded fragment acts as an inline header, so the list reads as a glossary rather than as a sequence of points. ## Why it reads as machine-written The shape is a default in assistant output: models bold the lead term of each enumerated item unless told otherwise, and Anthropic and OpenAI both publish guidance instructing their models to stop doing it in prose contexts. Three or more consecutive items in that shape is the signal, not any single bolded phrase. ## Detection Type: deterministic (a regular expression, run as written) Scope: document ECMAScript regular expression. The build validator compiles it and tests it against this entry's own specimens. ```regex (?:^[ \t]*(?:[-*+•]|\d{1,2}[.)])[ \t]+\*\*[^*\n]{1,60}?(?:[:：]\*\*|\*\*[ \t]*[:：—–-])[^\n]*\n){2}^[ \t]*(?:[-*+•]|\d{1,2}[.)])[ \t]+\*\*[^*\n]{1,60}?(?:[:：]\*\*|\*\*[ \t]*[:：—–-]) ``` Flags: gm ## Examples Constructed specimens. Written for this index. Never quoted from anyone's posts. ### Specimen 1 Before, exhibiting the tell: > **Consistency:** Post regularly so your audience knows when to expect you. **Authenticity:** Share real experiences rather than generic advice. **Engagement:** Reply to comments to build community. After, repaired: > We post on Tuesdays and Thursdays because that is when the support inbox is quietest and someone can actually answer the replies. When we tried five days a week, comments sat unanswered for two days and the threads died. Note: The three labels in the before block are interchangeable with any other three. The after paragraph names the constraint that produced the schedule, which no label could carry. ### Specimen 2 Before, exhibiting the tell: > **Step 1:** Open the settings panel. **Step 2:** Choose your export format. **Step 3:** Click Export. After, repaired: > Open the settings panel and choose an export format. CSV keeps the raw timestamps; the PDF option rounds them to the nearest minute, which matters if you are reconciling against billing. Note: This pair is here as a warning. Numbered procedure steps are legitimate technical writing, and the repair is only an improvement because the original steps carried no information the buttons did not already carry. ### Specimen 3 Before, exhibiting the tell: > - **Clarity:** say what you mean. > - **Brevity:** cut what you can. > - **Consistency:** keep the same voice throughout. After, repaired: > We cut the opening paragraph from every post after watching where readers stopped. The target length went from 400 words to 180. Nothing in the shorter version needed the paragraph we removed. Note: The specimen has three headings and no facts. The repair drops the headings and keeps what happened. ## False positives Who legitimately writes this way. Technical writers use definition lists, and the major documentation style guides sanction a bolded term followed by its description. Release notes and API references are built from the shape. Recipe writers use it, and anyone drafting in a Markdown editor with a formatting toolbar produces it without deciding to. Wikipedia's list of ineffective indicators cuts the other way too: when a writer's pre-2022 work already shows the same formatting habits, that continuity is evidence against AI rather than for it. ## Model attribution ChatGPT surface, hedged. The idiosyncrasies study describes bolding inside enumerations for ChatGPT and minimal formatting for Claude, measured on dated API snapshots that no longer correspond to shipping models. OpenAI documents no Markdown by default in the GPT-5 API, so the same weights behave differently through a different door. ## Platform notes - linkedin: LinkedIn's stated target is posts that restate without adding. Formatting is not mentioned in the policy, and the penalty is reduced distribution outside a person's network rather than removal. - x: The published X ranking code carries a post-level slop label and an account-level spam label, neither tied to formatting. Author diversity decay, which reduces the score of each additional post from the same account, is the closer analogue to a volume penalty. - wikipedia: Shortcut WP:AILIST. Wikipedia notes the list markers are often literal characters rather than wiki markup, which is a stronger signal than the bolding itself. ## Sources 1. Sun, Yin, Xu, Kolter, Liu: Idiosyncrasies in Large Language Models (ICML 2025) https://arxiv.org/abs/2502.12150 (tier: peer-reviewed; accessed 2026-08-14) 2. Anthropic: published Claude system prompts, release notes (formatting instructions by model and date) https://platform.claude.com/docs/en/release-notes/system-prompts (tier: primary-doc; accessed 2026-08-14) 3. OpenAI Cookbook: GPT-5 prompting guide (API does not format final answers in Markdown by default) https://developers.openai.com/cookbook/examples/gpt-5/gpt-5_prompting_guide (tier: vendor; accessed 2026-08-14) 4. Wikipedia: Signs of AI writing, sections on inline-header vertical lists and ineffective indicators https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing (tier: community; accessed 2026-08-14) 5. Lorenzetti: Keeping conversations real on LinkedIn (LinkedIn policy announcement, 2026-05-20) https://www.linkedin.com/pulse/keeping-conversations-real-linkedin-laura-lorenzetti-9821e (tier: primary-doc; accessed 2026-08-14) 6. xai-org/x-algorithm: published ranking and abuse-enforcement code (llm_slop labels, author diversity) https://github.com/xai-org/x-algorithm (tier: primary-doc; accessed 2026-08-14) 7. tropes.fyi pattern directory https://tropes.fyi/directory (tier: community; accessed 2026-08-14) 8. vale-ai-tells, 111 machine-checkable rules https://github.com/tbhb/vale-ai-tells (tier: community; accessed 2026-08-14) ## Status history Ids are permanent. A retired tell keeps its id and its page. - 2026-08-14, Contested: Supported at model level by the ICML 2025 idiosyncrasies study and by Anthropic's dated system prompts, which instruct against exactly this shape. Undercut by OpenAI's documented API default of no Markdown, which relocates the tell from a model family to a product surface. Technical documentation has used definition-style bullets since long before 2022, so the base rate in the affected genres is high. ## 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.