BAID: A Benchmark for Bias Assessment of AI Detectors

AI summary

Overview

Research area: AI-generated text detection, fairness and bias evaluation in NLP.

Technical level: Intermediate. The benchmark construction and the seven bias dimensions are easy to follow; interpreting the precision/recall/F1 breakdowns requires basic familiarity with classification metrics.

Scope: The paper introduces BAID, a 208,166-document-pair benchmark spanning seven bias dimensions and 41 subgroups, and uses it to measure fairness disparities across four open-source AI-text detectors on human-written text.

What This Paper Is About

AI-text detectors are increasingly used in classrooms and workplaces, but prior work found only isolated bias cases — most notably that detectors flag essays by English Language Learners as machine-generated. This paper builds a systematic benchmark that tests detectors across a much wider set of sociolinguistic factors: demographics, age, educational grade level, dialect, formality, political leaning, and topic. The goal is to show whether detectors produce consistent performance for some groups and fail for others before such tools are deployed publicly.

Key Contributions

1. A fairness-focused benchmark (BAID): Over 200k samples — a final dataset of 208,166 document pairs — organized across 7 bias types and 41 subgroups, drawn from sources including ASAP 2.0, the Blog Authorship Corpus, a GenZ vs. Standard English dataset, and a political-leaning article dataset.

2. A paired human/AI generation design: For every human-written document, a synthetic counterpart was generated with GPT-4.1 and Claude Sonnet 3.7 using structured prompts designed to preserve content and meaning while reflecting subgroup-specific writing styles. Each dataset record holds human_written_document, AI_generated_document, and subgroup_value.

3. A subgroup-level fairness audit of four detectors: Desklib, E5-small, Radar, and ZipPy were run over the full dataset with default decision thresholds as black-box systems, reported per subgroup on precision, recall, and F1.

4. Evidence that aggregate metrics hide bias: The paper shows that detectors which look accurate overall — especially Desklib — collapse on dialectal and informal writing, and that ZipPy's performance is highly dependent on input length.

Main Findings

• Recall gaps hit underrepresented writing styles hardest: The authors state that disparities are most visible in recall, meaning detectors misclassify human-written text as AI-generated for certain groups. Desklib's recall on demographic and grade-level subgroups sits at 0.83–0.96 but falls to 0.12–0.35 on dialectal and informal text.

• Desklib is the most consistent but style-sensitive: It reaches precision of 0.97–0.99 across demographic, grade-level, and political subgroups, and F1 of 0.89–0.96 there. Its precision drops to 0.44 on Singlish and 0.16 on GenZ content, with F1 of 0.14–0.47 on dialectal and informal registers.

• E5 has high precision but very low recall on several groups: Precision is 0.95–0.99 on demographics and grade levels, and F1 is 0.30–0.45 for demographics and 0.04 for GenZ English. Recall is 0.03–0.45 for demographic and political groups, but rises to 0.35–0.97 for Singlish and 0.55 for Standard English.

• Radar is stable but mid-range: Precision stays at 0.55–0.76 across subgroups and F1 at 0.60–0.75, with a stated recall band of 0.57–0.72 that drops on dialectal and informal categories (F1 0.21–0.33 on dialect and formality).

• ZipPy is the weakest and length-dependent: Precision is 0.19–0.31 on demographic and grade-level data and F1 is 0.03–0.27 there, but it improves to 0.49–0.54 precision and 0.65–0.70 F1 on dialect, formality, and topic text. Its recall collapses on demographics and grade levels (0.02–0.55) yet reaches 0.95–0.99 for age, dialect, topic, and GenZ subgroups. The paper attributes this to its compression-based heuristic, which is sensitive to input length because shorter texts offer fewer repeating tokens.

• Dialect and formality are the universal weak points: For every detector, dialectal (AAVE, Singlish) and GenZ informal text produce the largest performance drops relative to SAE and Standard English.

• Higher recall on AI-generated text than on human text: An appendix evaluation on the synthetic samples shows Desklib with precision 0.8–0.9, recall above 0.97, and F1 well over 0.9; E5 with very high recall but precision of 0.55–0.60; Radar with F1 typically 0.6–0.7; and ZipPy with recall above 0.9 on longer regular texts but precision of 0.3–0.5. The authors caution these results reflect prompt conditioning and model calibration, not genuine subgroup bias.

Methodology in Plain English

The researchers first gathered real human writing that came with known subgroup labels — for example, student essays with race, gender, ELL status, disability status, and socioeconomic metadata, or blog posts with author age and topic tags. They then asked large language models (GPT-4.1 and Claude Sonnet 3.7) to rewrite each document so it would read the way that same subgroup might write, using prompts that explicitly discouraged tell-tale AI phrases like "in this essay," "delve into," and "in conclusion." For conversational styles, the prompts were tailored — AAVE prompts emphasized authentic syntactic and lexical constructions, while Singlish prompts incorporated pragmatic particles and colloquial phrasing.

Quality was controlled in two stages: automatic filtering removed samples with token repetition or incomplete generations, and a semantic alignment check used sentence-level embeddings with a cosine-similarity threshold of 0.85 to confirm the AI text preserved the meaning of the human original. Hashtags, emojis, and links were stripped.

Four detectors were then run over the entire dataset separately on human-written and AI-generated text, with the fairness analysis reported on human-written text only. The reasoning given is that bias must originate from real human authorship; synthetic text only simulates subgroup characteristics through prompting. The authors note that including AI-generated pairs would test model calibration rather than fairness to real people.

Why This Matters

Detectors are already used to make consequential judgments about people, and the paper's central argument is that aggregate accuracy numbers mask large subgroup disparities — a detector that looks excellent on average can systematically flag an AAVE, Singlish, or GenZ writer as a cheater.

Real-world applications:

• Education: Flagging student work as AI-generated can lead to disciplinary action, and the paper shows the students most at risk include ELL students, students with disabilities, and lower grade levels.
• Academic and peer review: Detector outputs are used as screening signals in publishing workflows, where false positives can damage a researcher's reputation.
• Professional writing tools: Assistants used by millions need to avoid penalizing users whose writing style does not match a standard-English norm.
• Content platforms and moderation: Automated authenticity checks on social media posts and reviews can suppress legitimate voices if the classifier is style-sensitive.

Industry relevance: All three authors are affiliated with Superhuman (with Grammarly email addresses), reflecting direct commercial interest — companies shipping writing and productivity tools have to know whether their detection features behave fairly. The paper's call for bias-aware evaluation before deployment is aimed squarely at that product pipeline.

Future Directions

• Expand beyond four detectors: The authors note their analysis is limited to a fixed set of four, and suggest including larger commercial systems and emerging hybrid detectors such as multimodal or cross-lingual models.
• Move beyond English: BAID covers only English, so multilingual evaluation is left open, which would also require multi-language corpora.
• Address architectural comparability: Because ZipPy and other statistical detectors are highly sensitive to input length and formatting while neural detectors may inherit biases from pretraining corpora, the paper calls for dynamic threshold calibration and experiments with hybrid detection models.
• Build more equitable detectors and data: The conclusion points toward training data diversity and model calibration as the path to detection systems that perform reliably across population groups, writing styles, and contexts.

Target Audience

Researchers working on AI-text detection and NLP fairness will find the benchmark and disaggregated metrics most directly useful. Educators, academic integrity officers, and journal editors get a concrete picture of the false-positive risk they are accepting. Product teams at companies building writing assistants or moderation tooling should read it as a pre-deployment checklist item. Policy makers drafting rules on AI detection in schools and workplaces will benefit from the evidence that aggregate accuracy is not sufficient evidence of fairness.

Authors’ abstract

AI-generated text detectors have recently gained adoption in educational and professional contexts. Prior research has uncovered isolated cases of bias, particularly against English Language Learners (ELLs) however, there is a lack of systematic evaluation of such systems across broader sociolinguistic factors. In this work, we propose BAID, a comprehensive evaluation framework for AI detectors across various types of biases. As a part of the framework, we introduce over 200k samples spanning 7 major categories: demographics, age, educational grade level, dialect, formality, political leaning, and topic. We also generated synthetic versions of each sample with carefully crafted prompts to preserve the original content while reflecting subgroup-specific writing styles. Using this, we evaluate four open-source state-of-the-art AI text detectors and find consistent disparities in detection performance, particularly low recall rates for texts from underrepresented groups. Our contributions provide a scalable, transparent approach for auditing AI detectors and emphasize the need for bias-aware evaluation before these tools are deployed for public use.