being-human

being-human

Your assistant writes like a press release because it doesn't know who you are. This tells it.

being-human reads the prompts you've already typed, measures how you actually write, and hands the model a description of your voice specific enough to follow. It also builds a list of words to avoid, computed by contrast against the model's own output, so it's personal to you rather than borrowed from a blog post about ChatGPT tells.

Local, stdlib Python 3, no dependencies, no network calls.

my own writing, measured by this tool. every bar is one real sentence. make your own with being-human card.

quickstart

pip install being-human

being-human learn --name "Your Name"     # harvest your writing, measure it
being-human check draft.md               # AI tells
being-human verify draft.md              # is it statistically you?
being-human export --target chatgpt      # or agents, cursor, system, json
being-human card                         # draw your voiceprint as a shareable svg

No Claude Code history? Point it at a ChatGPT export or your own files:

being-human harvest --source chatgpt --path conversations.json --out .being-human/ --contrast
being-human harvest --source files --path ~/Documents/writing --out .being-human/

what comes out

Instructions with numbers attached, because "be concise" is not actionable and "your median message is 12 words" is:

- Write "i" lowercase. They do, 83% of the time.
- Vary sentence length hard. Theirs average 9 words but deviate 8, up to 81.
- No exclamation marks. Rate is 0.28 per 1k.
- Don't hedge. Hedge rate is 1.1 per 1k -- they assert.

the contrast trick

Your transcripts contain both sides of every conversation. So every word can be scored by log-odds: how much more likely you are to use it than the model is, on the same topics, in the same threads. The top of that list is your voice. The bottom is the model's.

Mine came back actually, clean, rather, genuinely, exactly, roughly. None of those appear on any published list of AI words. They're the specific tics this model has when talking to me. Yours will differ, because the contrast is against your writing.

two checks, not one

They catch different failures, and passing one says nothing about the other:

check	is this generic AI writing? tells, rhythm, structure
verify	is this you? authorship distance, Burrows's Delta against a length-matched null resampled from your own corpus

Measured on my own corpus:

	authorship	slop
my own writing	6th pct, me	99/100
generic LinkedIn slop	19th pct, me?	0/100
a README by someone else	94th pct, not me	97/100

Row two is the catch: slop has unremarkable grammar, so authorship distance passes it. Row three is clean prose with no tells that is plainly a different hand. The second case is the one nobody else tests for, and it's the one that bites.

Held-out chunks of real writing sit at the 49th to 60th percentile, with a 1-6% false-positive rate above p97. Thresholds are personal. Em dashes get flagged against your measured rate rather than a universal rule, which is why this doesn't punish people who genuinely use them.

mcp server

claude mcp add being-human -- being-human-mcp

Seven tools, two resources, two prompts. Configs for Claude Desktop, Cursor, Codex and Zed are in mcp/README.md.

MCP tools are model-invoked and nothing forces a model to call one, which is awkward when the payload is instructional. So every tool returns the voice rules alongside its own result. There's no path through the server that hands back a score without the target attached.

There's a Claude Code plugin too, via /plugin marketplace add Syedomershah99/being-human. It adds /being-human:learn, :check, :write, :verify, :export, and a hook that grows the corpus as you type.

privacy

Everything is local. Grep for urllib, requests or socket and you'll find nothing. Secrets are redacted at harvest time before anything touches disk, and .being-human/ is gitignored. The voiceprint is safe to share. The corpus it was built from is not.

prior art

This space got crowded in 2026. inside-lago framed the thesis, slop-guard is a mature slop linter, idiolect is architecturally closest, and writer-persona had the best idea of the lot: validate the profile with a backtest instead of trusting it.

Two things here I couldn't find elsewhere. The user-vs-model contrast, which derives the avoid-list from both sides of your own transcripts rather than from a curated list. And a deterministic impostor test. writer-persona uses an LLM judge across 8 axes; this is arithmetic, so it costs nothing and returns the same answer every time.

Burrows's Delta and log-odds with a Dirichlet prior are both established methods. The claim is the application and the calibration, not the statistics.

license

MIT. The tell list is data/slop-lexicon.json. If you keep seeing something that isn't in there, send a PR.

_{mcp-name: io.github.Syedomershah99/being-human}