Was This Written by AI?
An honest, multi-model assessment of whether a text was AI-written — with the evidence, not a fake certainty score.
A verdict (likely AI / likely human / mixed / genuinely inconclusive) with the specific tells, where the models disagree, and how much weight the assessment can honestly bear.
Authorship calls go wrong in two directions — the false accusation and the free pass — so the council is built to argue against itself. The Editor-in-Chief and the Copywriter read for voice the way professionals who handle thousands of drafts do: rhythm, register, the tics a human can’t help and a model can’t fake. The Data Skeptic interrogates the substance — citations that don’t resolve, examples with no lived texture. And the Devil’s Advocate exists to defend the text against the accusation, because fluent non-native writers and heavily edited prose are exactly who single-detector tools wrongly flag. Where they disagree, you see it — that disagreement is the honest confidence level.
Assess whether the text below was written by AI, by a human, or by a human using AI — honestly. THE TEXT: [paste the text or attach the document] CONTEXT (optional): [who supposedly wrote it, what kind of document it is, anything known about the author's usual writing] Debate the evidence, disagreeing where you genuinely read it differently: 1. Stylometric tells — formulaic transitions, hedged symmetry, uniform sentence rhythm, hallmark phrases and structures typical of AI drafting. 2. Substance tells — generic examples vs lived specifics, invented or unverifiable citations, confident claims with no source texture. 3. Human tells — idiosyncrasy, error patterns, voice consistency, personal detail that would be costly to fake. 4. Mixed-authorship signals — sections that shift register, an AI skeleton with human patches or vice versa. 5. The honest limits — detection is probabilistic; what this assessment CANNOT establish, and the false-positive risk for fluent non-native writers and heavily edited text. FINAL SYNTHESIS must deliver: the verdict with a calibrated confidence (never a fake percentage); the specific passages driving it, quoted; where the council members disagreed and why; and what additional evidence (drafts, revision history, a writing sample) would settle it properly. If the evidence is genuinely inconclusive, say so plainly — a wrong accusation is worse than an honest "can't tell".
An illustrative excerpt for a hypothetical brief of this type: A lecturer suspects a submitted essay was AI-written; the student says English is their second language and they used a grammar tool — the difference matters for an academic-misconduct panel. Every real run is generated fresh from your specifics, debated across rounds, and audited before the verdict.
The giveaway is not any single sentence — it is the evenness. Every paragraph lands at the same length, every argument gets a tidy both-sides hedge, every transition is "moreover" or "furthermore". Real student prose has a heartbeat: it rushes where they care and limps where they don’t. This text has no pulse.
Everything you just described is also what a careful second-language writer produces after three passes through a grammar checker — the tool flattens the pulse for them. If we call this AI on rhythm alone, we indict every diligent ESL student in the cohort. Show me a tell that editing cannot explain, or the verdict must stay open.
Here is one: footnote four cites a 2019 paper by authors who have never co-published, in a journal whose volume numbering does not reach that issue. A grammar tool does not invent citations — generation does. That single fabricated source moves me more than every stylistic argument combined.
Why not just give a percentage score like other AI detectors?
Because the percentage is theatre. Detection is probabilistic, and single-detector scores routinely flag fluent non-native writers and heavily edited text as AI. Decidi gives you what a score hides: the actual evidence, quoted from the text, where four independent models agree and disagree, and an honest "inconclusive" when that is the truth. For a decision with consequences — a grade, a hire, a client dispute — the evidence is usable; a bare 87% is not.
Can this prove someone used AI?
No — and any tool that claims proof is overselling. The council can find strong evidence (fabricated citations are the classic), weigh it honestly, and tell you what would settle the question properly: earlier drafts, revision history, or a supervised writing sample. It is built to inform a fair decision, not to convict.
What should I paste in for the best assessment?
The full text, plus anything known about the author’s usual writing — an earlier essay, an email, a report. A known-genuine sample transforms the assessment: voice comparison is far more reliable than style analysis in isolation.