HR Briefing

HR Technology

AI Writing Detectors Carry Bias Risk HR Cannot Ignore

OpenAI pulled its own detector after mislabeling 9% of human text. Six tools compared show why HR must test bias and false positives before screening candidates.

Top 6 AI Detector Tools in 2026: Features, Accuracy Claims and When to Trust Them
Top 6 AI Detector Tools in 2026: Features, Accuracy Claims and When to Trust Themerwinbosman / pixabay

Olivia Hart

· 4 min read

OpenAI withdrew its own AI text classifier on July 20, 2023, after it correctly flagged only 26% of AI-written text on a challenge set while mislabeling 9% of human writing as machine-generated. That record from the company that built the most-used generative models frames the core problem for HR teams now screening résumés, cover letters and writing samples with detection tools: the technology is probabilistic, and the cost of a wrong call falls on candidates and employees.

A comparison of six widely used detectors — CudekAI, GPTZero, Copyleaks, Originality.ai, Turnitin and Sapling — shows why every vendor in the category warns that results can be wrong, and several advise against acting on a score alone. For people-process decisions, the practical question is not which tool to buy but whether any detector can be defended in a hiring or disciplinary context.

What do the accuracy claims actually say?

The headline figures are not comparable, because vendors report against different datasets, thresholds and models. GPTZero cites the RAID benchmark, reporting 95.7% of AI text detected and 1% of human text wrongly flagged. Copyleaks publishes per-language figures ranging from 93.08% for Portuguese to 99.20% for English on AI text, plus a claimed 0.03% false-positive rate tied to a 2023 study — all vendor-reported.

Turnitin reports false-positive rates of 0.014 for English language learners and 0.013 for native English writers on documents above its 300-word minimum, and concludes there is no statistically significant bias. CudekAI displays a 99% accuracy figure in page metadata while referring to independent benchmarking without naming sources.

Where is the bias risk for employers?

The most consequential research finding for HR: a paper titled GPT detectors are biased against non-native English writers found the tested detectors consistently misclassified non-native English writing as AI-generated, and warned against their use in educational settings. Turnitin's page reports the opposite for its own tool — a divergence that argues for testing any detector on your own applicant and employee population before deployment.

Short text compounds the problem. Most vendors state accuracy improves with longer documents; Originality.ai sets a 100-word minimum and acknowledges that shorter text and AI rewriting tools affect results. Cover letters and short written assessments — the exact artifacts hiring teams screen — sit in the weakest zone for these tools. Sapling's page notes that accuracy improves after roughly 50 words, and that non-English text scored by its English model is less reliable.

What do the six tools offer?

The comparison found wide variation in limits, languages and cost models:

  • CudekAI: paragraph, sentence and word-level analysis; claims 100+ languages and detection of ChatGPT, Gemini, Claude, Llama, DeepSeek and Grok; but its page shows conflicting character limits of 1,000, 5,000 and 15,000.
  • GPTZero: a Writing Replay feature recording typing patterns and paste events — process evidence rather than style scoring; full support for five languages; free scans capped at 10,000 characters.
  • Copyleaks: 30+ languages, AI-generated code detection, and an AI Logic feature explaining why text was flagged; free scans up to 25,000 characters.
  • Originality.ai: combines detection with plagiarism and fact/hallucination checking; credit-based, with 100 free credits per account and credits expiring two years after purchase.
  • Turnitin: built for institutions, integrated into learning-management systems; frames results explicitly as probabilities.
  • Sapling: free check of the first 2,000 characters, up to 100,000 for Pro and Enterprise; a free API key; a reported 60% drop in false positives rests on unpublished internal evaluations.

The same research that documented bias also found that simple prompting strategies could get AI text past detectors while reducing bias — meaning evasion and false positives coexist as failure modes. New models also shift text distributions, so tools trained on older models can lose accuracy over time.

How should HR use these tools?

Because vendor figures cannot be compared, the guide recommends measuring on labeled samples that reflect real content, including writing from different author groups, and calculating false-positive and false-negative rates per group. An acceptable overall rate can hide a high rate for one group — the exact pattern the non-native-speaker research documented.

Practical safeguards before any screening or disciplinary use:

  • Treat a score as one signal, never proof; combine it with drafts, version history and conversations with the author.
  • Publish a policy telling candidates and employees how detection is used and how to appeal a result.
  • Test on writers resembling the population being evaluated, including non-native speakers.
  • Check how submitted text is stored and whether vendors use it for training before pasting confidential or personal material.

The distinction between detection and plagiarism checking matters here. A plagiarism checker matches text against existing sources; a detector guesses whether text resembles machine-generated writing — a fundamentally less certain task.

As detection models improve alongside the models they chase, HR teams that pair measured false-positive rates with corroborating evidence will be the ones whose AI-screening decisions hold up when challenged.

Source: sitepoint.com

Share this article:

More from Olivia Hart

Olivia Hart

Show full bio

Correspondent covering companies, deals and strategy in Recruiting and talent acquisition, learning and development at HR Briefing.

303 articles

© 2026 HR Briefing. All rights reserved.