Skip to content

People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text

Conference: ACL 2025
arXiv: 2501.15654
Code: https://github.com/jenna-russell/human_detectors
Area: AI Safety
Keywords: AI text detection, human detectors, LLM writing, human-AI comparison, robustness

TL;DR

Through an experiment with 1,740 annotations, it was found that human annotators who frequently use LLMs for writing tasks can detect AI-generated text with extremely high accuracy (only 1/300 errors via a 5-person majority vote). Even when facing paraphrasing and humanization evasion strategies, they perform significantly better than most automated detectors.

Background & Motivation

Background: AI-generated text is proliferating, and automated detectors (such as Binoculars, Fast-DetectGPT, and GPTZero) are the primary countermeasures. However, they suffer from low detection rates, lack of robustness against evasion attacks, and a lack of interpretability.

Limitations of Prior Work: Previous studies on human detection of AI text were mostly conducted prior to ChatGPT, using models far weaker than current LLMs (e.g., GPT-4o, o1-Pro). These studies generally concluded that human performance is close to random guessing.

Key Challenge: Although automated detectors are effective in simple scenarios, their True Positive Rate (TPR) drops drastically when facing paraphrasing and "humanization" strategies (e.g., Binoculars drops from 100% to 6.7%). Yet, these evasion strategies are common in actual malicious use cases.

Goal: Can humans reliably detect text generated by modern LLMs? If so, what kind of people excel at this, and what cues do they rely on?

Key Insight: It is hypothesized that individuals who frequently use LLMs for writing tasks (e.g., editors, copywriters, creative writers) naturally become excellent detectors because daily exposure to large volumes of AI text allows them to internalize AI writing patterns.

Core Idea: Human "expert detectors" (frequent LLM writing users) can achieve near-perfect detection of AI text without any training. The cues they focus on (vocabulary, structure, originality) can be codified into a guidebook to enhance automated detection.

Method

Overall Architecture

Five progressive experiments are designed: (1) GPT-4o without evasion, (2) Claude-3.5-Sonnet, (3) GPT-4o + paraphrasing, (4) o1-Pro, and (5) o1-Pro + humanization. Each experiment contains 60 articles (30 human-written + 30 AI-generated). Annotators provide binary labels, confidence ratings, highlighted cues, and paragraph-level explanations for each article.

Key Designs

  1. Paired Article Design:

    • Function: Generates a corresponding AI article for each human article, varying only the authorship.
    • Mechanism: Prompts the LLM with the title, subtitle, target length, and publication source of a human article to generate a comparable AI article. This forms minimal pairs, eliminating confounding variables like topic and length.
    • Design Motivation: A controlled-variable approach—focusing the detection task on "differences in writing style" rather than "differences in content."

  2. Expert vs. Non-Expert Findings:

    • Function: Evaluates the relationship between LLM usage experience and detection capability.
    • Mechanism: Recruiting 9 annotators with diverse backgrounds for Experiment 1 revealed that individuals who rarely use LLMs achieved a TPR of only 56.7% (near random), whereas those who edit text with LLMs daily achieved a TPR of 92.7%. Subsequent experiments focused on 5 "expert" annotators.
    • Design Motivation: Non-experts not only perform poorly but are also overconfident (average confidence 4.03/5), demonstrating a Dunning-Kruger effect.

  3. Codification of Human Detection Guide and LLM Simulation:

    • Function: Extracts detection cues from expert explanations, codifies them into a guide, and attempts to have LLMs simulate human detection.
    • Mechanism: Performs qualitative analysis on all experts' free-text explanations to extract three key types of cues: (1) AI vocabulary (e.g., vibrant, crucial, significantly), (2) stylized structures (e.g., optimistically vague conclusions), and (3) lack of originality (creative or engaging elements). These cues are codified into a guidebook and provided to GPT-4o/o1 for prompt-based detection.
    • Design Motivation: Explores whether the judgment process of human experts can be automated to reduce the cost of human annotation.

Key Experimental Results

Main Results

Five experts' majority vote vs. automated detectors TPR% (FPR%):

Method GPT-4o Claude GPT-4o Paraphrased o1-Pro o1-Pro Humanized Overall
Expert Vote 100 (0) 100 (0) 100 (0) 96.7 (0) 100 (0) 99.3 (0)
Pangram 100 (0) 100 (3.3) 100 (0) 100 (0) 90.0 (6.7) 98.0 (2)
GPTZero 100 (0) 96.7 (0) 100 (0) 76.7 (0) 46.7 (3.3) 85.3 (0.7)
Fast-DetectGPT 100 (0) 96.7 (3.3) 56.7 (3.3) 86.7 (0) 23.3 (3.3) 80.0 (7.2)
Binoculars 100 (0) 93.3 (0) 60.0 (6.7) 73.3 (0) 6.7 (0) 66.7 (1.3)

Ablation Study (Prompt-based Detection with GPT-4o as Detector)

Configuration GPT-4o o1-Pro o1-Pro Humanized
Zero-shot 100 (10) 56.7 (3.3) 6.7 (3.3)
Zero-shot + Guide 100 (10) 80.0 (6.7) 3.3 (3.3)
Zero-shot + CoT + Guide 100 (10) 86.7 (6.7) 3.3 (3.3)

Key Findings

  • Expert annotators made only 1 error out of 300 articles overall: The majority vote achieved TPR=99.3% and FPR=0%, outperforming all other methods.
  • Automated detectors collapse on humanized text: Binoculars' TPR on o1-Pro humanized text is only 6.7% and RADAR's is 0%, whereas the expert vote maintains 100%.
  • Non-experts perform comparably to random guessing: Achieving TPR=56.7%, FPR=51.7%, while being overconfident (confidence rating of 4.03/5).
  • AI vocabulary is the most common cue: The most frequently mentioned cues in expert explanations are specific "AI words" (e.g., testament, crucial, vibrant), followed by stylized article structure and a lack of originality.
  • LLMs struggle to simulate human detectors: While adding the detection guidebook helps the prompt-based method reach 100% TPR in simple scenarios, it still drops to only 3.3% on humanized text.

Highlights & Insights

  • Overturning the consensus that "human detection of AI text is near-random": The key lies in finding the right people—professional users who use LLMs daily for writing represent natural "expert detectors" who require no prior training.
  • Interpretability is a core advantage of human detection: Paragraph-level explanations not only clarify the detection process but also, in turn, enhance annotation quality by forcing annotators to read the text thoroughly.
  • Discrepancy in robustness against evasion strategies: Paraphrasing and humanization completely break down most automated detectors, yet remain virtually ineffective against human experts. This indicates that humans capture deeper patterns (e.g., narrative structure, originality) rather than superficial statistical features.

Limitations & Future Work

  • Only English non-fiction articles (<1K words) were tested, which may not generalize to academic papers, social media context, or other languages.
  • The sample size is limited (300 articles, 5 experts); population-level conclusions need verification on a larger scale.
  • High cost of expert annotation (approx. $865 per person for 5 rounds) makes scaling difficult.
  • The possible degradation of expert detection capability as LLMs evolve remains unexplored—whether experts will remain effective as future LLM generation quality further improves is unknown.
  • vs. Binoculars/Fast-DetectGPT: These statistical methods are effective in simple settings but vulnerable to evasion strategies. The robustness of human experts stems from their perception of semantic-level patterns.
  • vs. Pangram (Commercial Detector): The only automated method that matched human experts, scoring a TPR of 96.7% vs. the experts' 100% in the o1-Pro humanized setting.
  • vs. Prior Human Detection Studies (Ippolito 2020, Clark 2021): Previous conclusions stating that humans perform near-randomly was due to incorrect participant targeting—frequent LLM users represent an entirely new demographic in the ChatGPT era.

Rating

  • Novelty: ⭐⭐⭐⭐ Finding the counter-intuitive conclusion that "heavy LLM users make the best detectors" is highly valuable.
  • Experimental Thoroughness: ⭐⭐⭐⭐ Five progressive experiments are elegantly designed with multiple baseline comparisons and qualitative analysis, though the sample size is relatively small.
  • Writing Quality: ⭐⭐⭐⭐⭐ Clear logical arguments, excellent figure/table designs, and thorough experimental details.
  • Value: ⭐⭐⭐⭐ Directly guides AI text detection strategies in high-stakes scenarios.