Skip to content

FAIRNESS.md

Latest commit

 

History

History
94 lines (76 loc) · 4.76 KB

FAIRNESS.md

File metadata and controls

94 lines (76 loc) · 4.76 KB
Raw

Fairness: false-positive rate by population

A detector can score 95%+ on a balanced benchmark and still be unusable for a group of people. The number that matters ethically is not overall accuracy — it is the false-positive rate per population: how often real humans in each group are accused of using AI.

Every sample in this evaluation is human-authored, so any AI verdict is a false positive by construction. Reproduce it:

python scripts/prepare_fairness_set.py          # 785 human samples, 10 populations
python scripts/fpr_by_population.py --analyzer binoculars
python scripts/fpr_by_population.py --analyzer gpt2
python scripts/fpr_by_population.py --analyzer ensemble

Populations: TOEFL essays by non-native English writers and native-speaker student/college/technical essays (Liang et al., 2023, ChatGPT-Detector-Bias), plus real human answers across five HC3 domains. Neither corpus is redistributed.

Results (n=785 human samples; % = real people falsely flagged as AI)

The Ensemble column below is the old GPT-2-weighted blend, kept as the cautionary baseline. The current default ensemble is Binoculars-weighted, so its numbers equal the Binoculars column (re-measured: overall 1.0%, non-native 5.5%, every native population 0.0%). See "What this means", point 2.

Population n Binoculars (= current default ensemble) GPT-2 Ensemble (old GPT-2-weighted)
Non-native writers (TOEFL) 91 5.5% 26.4% 71.4%
US 8th-grade students 88 0.0% 1.1% 17.1%
College admission essays 70 0.0% 0.0% 24.3%
CS224N technical essays 145 0.0% 1.4% 9.7%
HC3 reddit_eli5 60 0.0% 3.3% 5.0%
HC3 open_qa 60 3.3% 83.3% 38.3%
HC3 finance 60 0.0% 6.7% 25.0%
HC3 medicine 60 0.0% 8.3% 3.3%
HC3 wiki_csai 60 0.0% 10.0% 56.7%
Overall 694 1.0% 13.5% 27.1%
Overall Wilson 95% CI [0.5%, 2.1%] [11.2%, 16.3%] [23.9%, 30.5%]

(toefl_gpt4_polished — the same TOEFL essays polished by GPT-4 — is excluded from the overall: it is human-authored but machine-edited, genuinely ambiguous. For the record: Binoculars 0.0%, GPT-2 39.6%, Ensemble 51.6%.)

What this means

1. Use Binoculars. It is both the most accurate (on HC3) and by far the fairest: 1.0% overall, and 5.5% [2.4%, 12.2%] on the hardest population. The cross-perplexity ratio cancels the "simple, predictable text looks AI-generated" effect that drives the bias — which is the entire reason the method exists.

2. A GPT-2-weighted ensemble is the WORST for fairness. 27% of real humans overall, and 71% of non-native English writers, were flagged. Its strong aggregate HC3 accuracy (0.95) masked this — the essay populations are harder than HC3's mixed answers.

This is why the default ensemble verdict is now Binoculars-driven (weight_binoculars=1.0, GPT-2/NLTK weight 0). Re-measured with that default, the ensemble collapses onto Binoculars' numbers: overall 1.0% [0.5%, 2.1%], non-native 5.5%, every native-speaker population 0.0%. The Ensemble column in the table above is the old GPT-2-weighted blend, kept as the cautionary baseline. GPT-2/NLTK sub-scores still display for transparency but no longer drive the verdict.

3. GPT-2 reproduces Liang et al. (2023). 26% of non-native writers vs ~1% of native writers. Non-native English is simpler and more predictable, so it has low GPT-2 perplexity, so single-model perplexity flags it. This is a known, published failure mode, confirmed here.

4. Even Binoculars is not perfectly fair. Non-native writers still see 5.5% vs 0% for native-speaker essays. The confidence interval is wide (n=91), but the disparity is real and is stated, not hidden. A 1-in-18 false-accusation rate is not "safe" for a high-stakes decision about a non-native writer.

Do not use any of these for consequential decisions

None of these analyzers should be the basis for accusing a specific person of academic dishonesty or fraud. Treat every result as weak probabilistic evidence, never proof, and weight it toward not flagging when the writer may be a non-native speaker, a child, or writing in a simple/formulaic register — the groups these detectors are most likely to wrong.

Caveats

  • Essay n per population is 70–145; the Wilson intervals reflect the resulting uncertainty. Larger samples would tighten them.
  • Populations are English-only and mostly academic. Other registers and languages are unmeasured.
  • Numbers are analyzer verdicts at each analyzer's default threshold. Binoculars' threshold is the HC3-calibrated 0.7625; the ensemble uses ThresholdConfig / EnsembleConfig defaults.