Skip to content

Measure how much machine-written text the score separates, on RAID - #104

Merged
peopleworks merged 10 commits into
mainfrom
raid-detection
Nov 18, 2026
Merged

peopleworks merged 10 commits into
mainfrom
raid-detection

Conversation

@peopleworks

Copy link
Copy Markdown
Owner

Docs/CALIBRATION.md measures how often this tool is wrong about a person and says what it leaves out: nothing about how much AI writing it catches, because a tool that flags nothing has a perfect false-positive rate. This measures that other half, on RAID (Dugan et al., ACL 2024), a public benchmark nobody here assembled. Draft until the maintainer decides; nothing has been submitted to the RAID leaderboard.

The order, checkable in the history

  1. e007adc — the predictions, committed and pushed before any RAID text was scored (Docs/Detection/prediction.md), with a note added later the same day, still before measuring, saying P6 was probably wrong and why.
  2. c2925e2 — signsofai-calibrate raid and tools/raid/. The engine only produces scores; RAID's own run_evaluation picks every cutoff and counts every hit.
  3. 986319f — the measurement and Docs/DETECTION.md, generated from Docs/Detection/*.json.
  4. ab41f6a — eleven live surfaces that said detection was never measured.
  5. 74433f8 to the last commit: the committee's corrections, each commit saying which finding it answers.

The numbers

Caught at 5% FPR on RAID's humans 26.6% (12.9% at 1%, AUROC 0.63)
Chat-tuned generators / base models 37.5% / 12.6%; base models sit at AUROC ≈ 0.5
ChatGPT alone 59.6%
Machine texts in RAID reaching the 649-word floor none, so the shipped product would have answered none of them
As shipped (30/100 and the 649-word floor) a verdict on 0 machine texts and 14 human ones
RAID humans above 30/100 if the floor did not exist 11.3%, falling with length from 14.5% to 2.0%

Predictions: P1, P2, P3, P5 and the overall range held; P4 failed (machine verse is caught well, through sentence rhythm); P6 held by its letter only (the analyzer restores disguised letters before reading, as the later note expected).

Found along the way, and labelled as such on the page

  • 9 of 108 long news articles cross the 30/100 boundary (4.4–15.1%), an interval wholly above the published one: by the rule that set the boundary, 30/100 is not supported for news. Post-hoc, one sample, cause unmeasured (stat.burstiness fires on 74% of them against 32% of the calibration corpus). Overall 14 of 705 long human texts. The published false-positive rate is not changed; that goes through the calibration corpus and the committee. It is now stated, without typed figures, on CALIBRATION.md, /why and the README.
  • The classic tells (delve, testament to, tapestry) fire 20–64× more on chat-tuned text than on human text and about 1× on base models. Each is rare; sentence rhythm does the separating.
  • rhet.em-dash, lex.actually and lex.literally fire more on RAID's humans than on its 2023 chat models. No rule was changed.
  • Under RAID's cheap attacks the score barely moves and the scanner reports 99.7% (homoglyph) and 100% (zero-width), with 0.84% false alarms on unattacked humans (0% systematic). For scale, RAID's own numbers for RoBERTa (ChatGPT) go 42.5% → 0.0% under the same attacks.

Behaviour this changes

Only text. The evidence report's limitation paragraph (English and Spanish, re-pinned to its new English sourceHash), /why in both languages, README.md, SKILL.md, the corpus README, and the generated sentences in CALIBRATION.md and PARAPHRASE.md. Both generated pages were checked against a regenerated copy: only the changed sentences match, and their engine and run lines were left alone because neither was re-run. Re-running the calibration with 0.8.1 moved no number. Published blog articles are dated and were not touched.

Committee

Three lenses, three disjoint sets of findings, all verified before acting.

  • Codex could not run: the Windows sandbox fails with helper_sandbox_lock_failed, and it exited 0 without a verdict. Its log still held two correct findings (the homoglyph drop is 0.53 points, and lower-case moreover is restored).
  • Copilot took the same brief, reproduced every headline number with its own implementation, and rescored all 30,971 texts sequentially (byte-identical). It found presentation errors: cutoffs printed rounded across ties, 99 repeated human texts in RAID-train, a repetition-penalty comparison across different generators, and score joins that would accept incomplete files.
  • Fable read the page as a teacher, twice. First pass: it would be quoted as "catches a quarter of AI, 60% of ChatGPT", which the shipped product does not do. Second pass: it still invited "a long essay over 30 is probably AI", when every long text over 30 here was human. Both fixed, plus the unsourced re-run claim, which now has a file behind it. The second-pass fixes were not sent back for a third pass.
  • One recommendation declined on purpose: Fable asked for counts ("9 of 108") on /why, the README and the calibration page. Those surfaces now say it in the past tense, as one sample, with the cause unmeasured, but without typed figures, because figures copied onto surfaces are how this project's self-description has gone stale before. The figures live on DETECTION.md, generated from data.

Not in this PR, on purpose

  • A RAID leaderboard submission. A merged results.json there is public for good; that is the maintainer's call.
  • Any change to the scanner. The measurement found that full substitution leaves some words neither reported nor restored. Fixing it is a behaviour change with its own measurement against human text, tracked privately until it is fixed.
  • The dated blog article that says detection was never measured (signsofai-verdict-story). It was true when published; it is left as it was.

Verification

  • dotnet test SignsOfAI.slnx: 393 + 72 + 7 pass; SignsOfAI.Desktop.slnx: 31 pass.
  • Docs/Detection/prediction.md is unchanged since e007adc (git diff e007adc HEAD on it is empty).
  • Parallel scoring byte-identical to --threads 1 (2,000 texts here; the full 30,971 and the 16,000 attacked texts by the committee).
  • Headline recomputed independently by a reviewer, without tools/raid/evaluate.py.

🤖 Generated with Claude Code

peopleworks and others added 10 commits November 17, 2026 21:11
The calibration measures how often the tool is wrong about a human and says
outright that it leaves out the other half: how much AI writing it catches.
RAID measures that half. These are the predictions, committed before any
RAID text is scored so the order can be checked in the history.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The calibration measures the half that needs only human writing. This adds
the tooling for the other half: `signsofai-calibrate raid` scores JSON lines
with the shipped engine and evaluates nothing. RAID's own run_evaluation picks
every threshold and counts every hit, so the scorer cannot shape the number.

tools/raid/ draws a seeded, fingerprinted sample of RAID-train, evaluates it,
applies the eight RAID attacks that need no language model (strengths read
from RAID's attack.py, not copied), and writes Docs/DETECTION.md from the
resulting JSON. Parallel and sequential scoring were checked byte-identical.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
At a 5% false-positive rate on RAID's human texts the score catches 26.6% of
machine-written text: 37.5% from chat-tuned generators, 12.6% from base
models, whose AUROC sits near 0.5. Five of the six pre-registered predictions
held as written; P4 failed (machine verse is caught well, through sentence
rhythm) and P6 held by its letter only.

Found along the way, and labelled as such: no machine text in RAID reaches
the 649-word floor, so the product as shipped would have answered none of
it; and 9 of 110 long news articles cross the 30/100 boundary, a register
the calibration corpus never contained. The published rate is not changed.

Nothing is submitted to the RAID leaderboard.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Eleven live surfaces said, in one wording or another, that how much machine
writing the tool catches was never measured. As of DETECTION.md that is
false. Each now says what was measured and where, without typing a figure
that could drift: the README, SKILL.md, both languages of /why, the evidence
report in English and Spanish (re-pinned to its new English source), the
CALIBRATION and PARAPHRASE pages through their generators, the corpus README,
and the comments and test message that restated the old claim.

CALIBRATION.md and PARAPHRASE.md were edited to match their generators and
checked against a regenerated copy: only the changed sentences match, and
the engine and run lines were left alone because neither page was re-run.
Re-running the calibration with 0.8.1 moved no number. Published blog
articles are dated and were not touched.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
From the committee (Codex, which could read but not execute): the mean
machine score drops by 0.53 points under the attack, not "less than half a
point", and lower-case "moreover" is restored (four Latin letters, four
lookalikes, a tie the scanner accepts); the capitalised "Moreover" is not.
Checking the second point found a third case the page now names: a
one-letter word such as "a" has no Latin letter left and is never restored.

The page no longer proposes a specific fix. Closing these is a change of
behaviour and needs its own measurement against human text.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Copilot reproduced every headline number independently (26.55%, chat 37.5%
vs base 12.6%, AUROC 0.629) and a full sequential rescore was byte-identical.
What it found was in the presentation:

- The domain table printed RAID's cutoffs rounded to one decimal, the same
  step as the engine's scores, which carries whole ties across the cutoff:
  at the printed 40.5, news would flag 5.11% of humans. The table now gives
  the lowest score each cutoff counts, using RAID's own comparison.
- RAID-train repeats some human texts verbatim (13,272 distinct in 13,371
  rows). The headline still counts rows, as RAID does; the long-text check,
  which is ours, now counts each text once: 14 of 705, news 9 of 108.
- The repetition-penalty sentence compared pools with different generators
  and named a mechanism nobody measured. It now compares only the six
  generators run both ways, and says it is an observed difference.
- attack.py and evaluate.py now refuse incomplete score files instead of
  counting a missing score as a miss.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fable read the page as a teacher with no statistics and found that it would
be quoted as "it catches a quarter of AI text, 60% of ChatGPT", which is not
what the product does: on this benchmark the shipped product gives its
verdict on no machine text and on 14 human texts, and at the lengths where
it speaks, detection was never measured. Every number on the page was
right; the page did not say that in one place.

- A "read this first" block for teachers, and the headline now gives both
  halves of the as-shipped result and the news finding.
- News: 9 of 108 long articles cross 30/100 (4.4-15.1%), an interval
  wholly above the published one, so by the rule that set the boundary
  30/100 is not supported for news. Said plainly, with the post-hoc
  caveat, and now also on CALIBRATION.md (through its generator), on both
  languages of /why and in the README, without typed figures.
- New data instead of asserted causes: the human rate at 30/100 by length
  (14.5% under 200 words to 2.0% at 649+), stat.burstiness on 74% of long
  news against 34% of all long human text and 32% of the calibration, the
  rate re-weighted to RAID's own generator mix (25.3%), and the predictions
  re-judged at the single cutoff (P3 holds there by 0.7 points).
- P6 is "held, weakly", not a third category; P5 carries its interval,
  which touches 1%. Both are judged from attacks.json, not in the generator.
- Causal claims without a measurement were cut or marked as observations;
  the leaderboard block no longer singles out vendors, says which files
  were read and that FastDetectGPT was skipped, and no longer attributes
  behaviour to detectors this project did not run.
- The evidence report links a real URL, and the corpus README says RAID
  answers one of its two objections, not both.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fable re-read the page from scratch. The first misreading was gone; the one
left was "a long essay over 30 is probably AI, since it's only wrong 0.7% of
the time", which the page set up and never closed. On this benchmark every
long text that crossed 30 was written by a person, and the page now says so
where the teacher reads.

- The teacher block says what the calibration corpus holds, that ordinary
  school essays are outside it, puts 9 of 108 against the promise (about one
  in twelve against one in twenty), counts people instead of fractions of a
  student, and says what the tool is still for.
- The re-run of the calibration with 0.8.1 is now on file
  (Docs/Detection/calibration-rerun.json: only engine and date moved), and
  the page prints the sentence from it instead of asserting it.
- Scanner misses are listed by class with a reason computed from the text
  (3 machine, 2 human; 3 in another script, 2 with no Latin letter left).
  One "degenerate output" was a human Reddit post of nothing but A's.
- The news finding is in the past tense, as one sample, with its cause
  unmeasured, on every surface that carries it; "low" is now "too low to
  stand out from human writing" in the report, /why and SKILL.md.
- One name for the other leaderboard entries, the review claim reduced to
  what the repository shows, and the AUROC sentence limited to the data.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The README now includes the calibration re-run that calibration-rerun.json
comes from, and the two arguments attack.py evaluate gained for explaining
scanner misses.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The page keeps the measured effect (vowel-heavy words under full
substitution are neither reported nor restored) and drops the detail that
belongs with a fix not yet made.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@peopleworks
peopleworks marked this pull request as ready for review November 18, 2026 05:36
@peopleworks
peopleworks merged commit 753fe2a into main Nov 18, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant