AI output linter.
Paste any AI-generated paragraph, post, email, or doc. The linter scans for the regex-detectable 2026 writing tells (em dashes, the word "comprehensive", the AI vocabulary cluster, templated transitions, curly quotes, and more) and shows each match with line and column numbers. Runs as you type, in your browser. Nothing uploads.
prevent the tells: AI tell killer
· paste your text
0 words
· don't have text handy? click load sample for a paragraph that triggers every detection.
· detected tells
0 findings
· waiting for text
Paste anything on the left to see what an editor pattern-matching for AI tells would catch.
What this catches and what it misses
The linter scans for the AI tells that have a reliable regex form. There are about 9of them in the current ruleset, ranging from single-character signals (em dashes, curly quotes) to specific word patterns (the word "comprehensive", the AI vocabulary cluster, templated openers like "It's worth noting", transitions like "Moreover" and "Furthermore").
What pattern-matching can't catch: sycophancy ("Great question!") because it depends on positional context, rule-of-three lists because counting parallel items requires parsing structure, hallucinated specifics because verifying truth needs knowledge, templated conclusions because identifying a wrap-up shape needs interpretation. For those, the right tool is the prevention-side AI Tell Killer paste-in prompt, which stops them before they appear in output rather than detecting them after.
How to read the findings
- Findings are grouped by tell. Each card shows the tell name (e.g. "Em dashes", "The word comprehensive"), the category, and how many matches were found.
- Each match shows line, column, and context. The matched phrase is highlighted in the surrounding sentence so you can see what to rewrite.
- No findings doesn't mean clean. A regex pass green-lights regex-detectable tells only. Behavioral tells still need a read-through.
- Rewrite, don't character-swap. Find-and-replace patterns (em dash to hyphen, "comprehensive" to "thorough") read obvious. The most natural fix is to rewrite the sentence so the flagged word isn't needed.
The two-tool workflow
The Briskly anti-AI-tells stack works in two passes:
- Prevent with the AI Tell Killer. Pick the tells you want to kill, paste the generated prompt into Claude / ChatGPT / Gemini custom instructions. The model is now primed to avoid those tells in every reply.
- Verify with this linter. Paste the output, scan for tells that slipped through. Rewrite anything that flagged.
The prevent step does most of the work. The verify step catches the 5-15% that slip through on long outputs, especially with models that are weaker at instruction- following over many tokens (Gemini in particular).
FAQ
What does this tool actually do?
It takes the text you paste, runs about 10 regex patterns against it, and shows you where each pattern matched. Each pattern targets a specific writing tell that LLMs produce at much higher rates than human writers (em dashes, the word 'comprehensive', the AI vocabulary cluster of delve/leverage/tapestry/etc., templated transitions like 'Moreover' and 'Furthermore', promotional adjectives like 'vibrant' and 'groundbreaking', curly quotes, parallelism patterns like 'Not just X, but Y', and a few more). It's a pattern-matcher, not an AI. Same idea as spell-check, but for AI tells.
Is this an AI detector? Does it tell me if text was AI-generated?
No. It doesn't classify text as AI or human. It just highlights specific patterns that AI output overuses. A human writer can hit any of these patterns occasionally and still be writing as a human. An LLM can produce text that avoids all of these patterns (especially with a good system prompt) and still be AI. This tool flags writing that has the visible AI-style fingerprint; it can't prove origin. Use it to clean up your own AI-assisted drafts, not to accuse anyone.
Why are some tells missing? I don't see 'sycophancy' or 'rule-of-three lists'.
Those tells can't be detected by regex. Sycophancy ("Great question!", "You're absolutely right!") is behavioral and only flags reliably with context. Rule-of-three lists need to know whether a list is genuinely three items or padded to three. Templated conclusions need to interpret meaning. The AI Tell Killer paste-in prompt covers these tells at the prompt-writing stage (prevent them rather than detect them after the fact). This linter covers the subset that pattern-matching can catch reliably.
Is anything I paste sent anywhere?
No. The entire linter runs in your browser. The text you paste is matched against regex patterns locally and never transmitted. Your text persists in browser LocalStorage so it's still there if you reload the page; that data stays on your device. No analytics event includes your text. The page makes no network request when you type, switch findings, or copy/paste.
What should I do when the linter finds something?
For each finding, rewrite the sentence containing it rather than character-swapping. AI find-and-replace patterns (em dash to hyphen, 'comprehensive' to 'thorough') show up to readers too. The replacement strategy that reads most natural is to restructure the surrounding sentence so the tell-word isn't needed at all. For example, instead of 'a comprehensive guide to X', try 'a guide to X that covers Y, Z, and W' or just 'a guide to X'. The 'what to do instead' table in the em dashes guide shows the canonical replacements for each situation.
How is this different from the AI Tell Killer?
The Tell Killer is preventive: it builds a paste-in prompt for Claude / ChatGPT / Gemini that stops the tells before they appear in the output. The Output Linter is diagnostic: it shows you which tells leaked through into text you already have. Both use the same underlying tells library. Use Tell Killer once (paste the rules into your custom instructions); use the Output Linter on individual drafts when you want a final check before shipping.
Can I use this on text that came from somewhere other than ChatGPT?
Yes. The patterns it detects are not model-specific. Anything AI-generated by Claude, Gemini, Llama, Mistral, DeepSeek, or any other current-gen LLM produces them at similar rates. Human writers with a polished-prose habit can also hit some of these patterns (em dashes, occasional 'comprehensive'); the linter doesn't care about source, it only flags patterns. If a human draft trips a few rules, that's normal; if a draft trips dozens, the writing reads as AI-flavored regardless of who wrote it.
More on this topic: the AI Tell Killer (prevent the tells in the prompt) and the em dashes guide (deep-dive on the most-checked single tell).