Source-linked AI summary
How Close is ChatGPT to Human Experts? Comparison Corpus, Evaluation, and Detection
Biyang Guo, Xin Zhang, Ziyuan Wang, Minqi Jiang, Jinran Nie, Yuxuan Ding, Jianwei Yue, Yupeng Wu
TL;DR
The paper addresses how ChatGPT differs from human experts and how its generated content can be identified amid concerns about misuse. It constructs HC3, compares human and ChatGPT answers through evaluations and linguistic analyses, and develops detection systems across scenarios. The study reports distinctive response patterns, decent held-out detector performance, and publicly available data, code, and models.
Problem
The paper studies ChatGPT’s remaining differences from human experts and the need to identify generated content amid potential social risks from misuse.
Method
The authors construct HC3, conduct human evaluations and linguistic analyses, and develop detection models evaluated across training and testing scenarios.
Results
ChatGPT tends to provide long and detailed answers, while the proposed detectors show decent performance on held-out test sets.
Takeaways & Limitations
The findings provide insights into differences between humans and ChatGPT and guides for research and development of AI-generated-content detection tools.
Takeaways & Limitations
The collected data remain insufficient in amount and range, with imbalance across sources, and the analyses use ChatGPT answers generated without special prompts.
Abstract
from arXiv · showhide
The introduction of ChatGPT has garnered widespread attention in both academic and industrial communities. ChatGPT is able to respond effectively to a wide range of human questions, providing fluent and comprehensive answers that significantly surpass previous public chatbots in terms of security and usefulness. On one hand, people are curious about how ChatGPT is able to achieve such strength and how far it is from human experts. On the other hand, people are starting to worry about the potential negative impacts that large language models (LLMs) like ChatGPT could have on society, such as fake news, plagiarism, and social security issues. In this work, we collected tens of thousands of comparison responses from both human experts and ChatGPT, with questions ranging from open-domain, financial, medical, legal, and psychological areas. We call the collected dataset the Human ChatGPT Comparison Corpus (HC3). Based on the HC3 dataset, we study the characteristics of ChatGPT's responses, the differences and gaps from human experts, and future directions for LLMs. We conducted comprehensive human evaluations and linguistic analyses of ChatGPT-generated content compared with that of humans, where many interesting results are revealed. After that, we conduct extensive experiments on how to effectively detect whether a certain text is generated by ChatGPT or humans. We build three different detection systems, explore several key factors that influence their effectiveness, and evaluate them in different scenarios. The dataset, code, and models are all publicly available at https://github.com/Hello-SimpleAI/chatgpt-comparison-detection.
1 Introduction
The paper frames ChatGPT as both a strong conversational system whose distance from human experts merits study and a source of potential social risks. It responds by building a comparison corpus, analyzing human and ChatGPT answers, and developing detection models with public resources.
- ChatGPT’s capabilities raise questions about its remaining gaps from human experts and implicit linguistic differences.
- Potential misuse of ChatGPT-generated content threatens the quality and reliability of user-generated-content platforms and affects settings such as online exams and medical analysis.
- Nearly 40K questions and corresponding human and ChatGPT answers form the HC3 dataset across open-domain, computer science, finance, medicine, law, and psychology.
- Human evaluations and linguistic analyses reveal patterns in human and ChatGPT answers that inform detection and future directions for language models.
- Three ChatGPT detection models show decent performance on held-out test sets and expose factors essential to detector effectiveness.
- The corpus, evaluations, detection models, and code are publicly released to support research and regulation of AI-generated content.
2 Human ChatGPT Comparison Corpus (HC3)
HC3 pairs human and ChatGPT answers to shared questions across multiple domains and languages. The corpus combines existing question-answering resources and wiki-derived questions with manually collected ChatGPT responses.
- HC3 compares human and ChatGPT answers to the same questions to study their language and helpfulness, honesty, and harmlessness.
- The dataset draws mainly on public question-answering datasets with expert or highly voted answers and wiki text from Wikipedia and BaiduBaike.
- ChatGPT answers are generated for collected questions in refreshed threads, with additional dataset-specific instructions used to align some responses with human-answer contexts.
- ChatGPT can produce different answers to the same question across threads, but the observed differences were small, so most questions received one answer.
- The corpus records multiple human answers and ChatGPT answers for each question using separate human_answers and chatgpt_answers fields.
- The English version contains 24,322 questions, 58,546 human answers, and 26,903 ChatGPT answers; the Chinese version contains 12,853 questions, 22,259 human answers, and 17,522 ChatGPT answers.
3 Human Evaluation & Summarization
The study combines Turing and helpfulness evaluations with volunteer feedback to compare human and ChatGPT answers. ChatGPT was often judged more helpful, while its responses showed recognizable patterns in organization, neutrality, detail, and focus that distinguished them from human answers.
- Human Evaluation: The evaluation used Turing tests and a helpfulness test with expert and amateur volunteers.The Turing tests varied paired versus single answers, while helpfulness compared human and ChatGPT answers directly.
- Human Evaluation: ChatGPT answers were generally considered more helpful than human answers in over half of questions, especially in finance and psychology, but performed poorly in medical consultations.The paper associates the medical result partly with ChatGPT’s lengthy answers compared with humans’ more direct suggestions.
- Distinctive Patterns of ChatGPT: ChatGPT responses were organized, logically structured, detailed, and often defined core concepts before presenting step-by-step explanations and summaries.The paper links the tendency toward long answers partly to RLHF and notes that shorter prompts can change this behavior.
- Major Differences between Human and ChatGPT: Humans were more divergent, emotionally expressive, and stylistically varied, while ChatGPT stayed focused on literal question wording and used more logical connective language.Human answers more often included humor, irony, metaphors, examples, slang, punctuation, and grammar features conveying emotion.
- Major Differences between Human and ChatGPT: ChatGPT generally produced safer, more balanced, neutral, and informative answers, whereas humans more often supplied specific citations and subjective expressions.ChatGPT was described as relatively neutral on sensitive topics, while human answers included more individualized language and source-based detail.
- Major Differences between Human and ChatGPT: Overall, the authors characterize ChatGPT as a conservative team of experts: comprehensive and neutral, but lacking individuality.This characterization summarizes the reported contrast between ChatGPT’s broad, balanced responses and humans’ more distinctive communication.
4 Linguistic Analysis
The analysis compares human and ChatGPT answers through vocabulary, grammatical structure, sentiment, and perplexity. Across these dimensions, ChatGPT tends toward longer, less diverse, more neutral, and lower-perplexity language, while human writing shows greater lexical and emotional variation.
- 4.1 Vocabulary Features: Human answers are shorter but use larger vocabularies than ChatGPT, especially in Chinese open_qa and medical splits where ChatGPT answers are nearly twice as long.The comparison uses average length and vocabulary size across English and Chinese corpus splits.
- 4.1 Vocabulary Features: Human word density exceeds ChatGPT’s in every split, indicating more diverse vocabulary use at comparable text lengths.Density is defined from vocabulary size, answer length, and number of answers.
- 4.2 Linguistic Structure: In HC3-English, ChatGPT uses more NOUN, VERB, DET, ADJ, AUX, CCONJ, and PART words but fewer ADV and PUNCT words than humans.The POS comparison is shown in Figure 1 for the HC3-English dataset.
- 4.2 Linguistic Structure: ChatGPT shows more determination, conjunction, and auxiliary dependency relations, with longer punct and dep distances but shorter conj distances than humans.The dependency comparison is based on parsed answers and corresponding dependency distances.
- 4.3 Sentiment Analysis: ChatGPT expresses more neutral sentiment, while humans express significantly more negative sentiment and slightly more positive sentiment.Neutral sentiment is the largest category for both groups, and the authors characterize ChatGPT as less emotional overall.
- 4.4 Language Model Perplexity: ChatGPT-generated text has lower and more concentrated PPL distributions at both text and sentence levels, whereas human text has more high-PPL values and a long-tailed distribution.PPL distributions are reported for English and Chinese data at both text and sentence levels.
5 ChatGPT Content Detection
The paper develops multiple ChatGPT-content detectors and evaluates them across text granularities, data sources, and dataset variants. Its experiments specifically test lexical cues, sentence-level detection, question-answer inputs, and robustness to distribution shifts.
- 5.1 Methods: Three detector types are evaluated: GLTR Test-2 logistic regression, a RoBERTa single-text classifier, and a RoBERTa question-answer classifier.The systems combine classic machine learning and deep learning approaches for different detection scenarios.
- 5.2 Detection Systems: The question-answer detector jointly inputs a question and its corresponding answer, targeting question-answering scenarios rather than answer-only classification.RoBERTa’s text-pair format joins the question and answer with a separating token.
- 5.3 Experiment Design: The study frames detector evaluation around three questions: the influence of indicating words, the difficulty of sentence-level detection, and whether questions improve origin classification.These questions motivate the different filtering, splitting, and input-format experiments.
- 5.3 Experiment Design: The experiments construct six raw, filtered, sentence, full-text, and mixed dataset versions to test detector performance across granularities and distribution shifts.Cross-version evaluation is treated as out-of-distribution generalization because training and testing styles differ.
5.4 Results
The experiments compare detector robustness across text granularity, indicating-word filtering, training setup, distribution shift, and data sources. RoBERTa-based detectors generally outperform GLTR and benefit from fine-grained or QA-style training, but performance varies across languages and sources.
- Detector robustness: RoBERTa-based detectors lose only 1.5-2% F1 in English and 2-3% in Chinese after sentence splitting, whereas GLTR loses over 10% and above 15%, respectively.These results indicate greater robustness to interference for RoBERTa-based detection.
- Indicating words: Filtering indicating words improves full-text RoBERTa detection by more than 3% F1 on average but can slightly reduce sentence-level performance.The authors attribute this contrast to indicating words contributing more to sentence-level discrimination.
- Text granularity: Single-sentence ChatGPT detection is harder than full-text detection, with Chinese raw-mix F1 at 94.09% for raw sentences versus 97.43% for raw full answers.Fine-grained sentence data improves detection of subtle patterns, and mixed training can approach sentence-based performance.
- QA versus single-text detection: QA-style detectors generally outperform single-text detectors, averaging 97.48% F1 on English full test datasets and surpassing single models by 5.63%.The QA model also achieves the best reported Chinese evaluation at 94.22%.
- Data-source and language effects: Detection performance depends on data source and language: Chinese human and ChatGPT scores are comparable, while English ChatGPT scores are lower and transfer performance is more variable.The English wiki_csai source is challenging, whereas Chinese evaluation is more stable across transfer datasets.
6 Conclusion
The paper introduces HC3, a large comparison corpus pairing human and ChatGPT answers, and uses it for human evaluation, linguistic analysis, and detection experiments. The released data, code, and models support further research and applications in AI-generated-content detection.
- HC3 contains nearly 40K questions with corresponding human and ChatGPT answers for comparing their language and behavior.The corpus supports analyses of implicit differences between humans and ChatGPT.
- The paper combines human evaluations, linguistic analysis, and content-detection experiments to study human-ChatGPT differences and detector behavior.The authors present these studies as sources of guidance for future LLM and AIGC-detection research.
- All data, code, and models are publicly available to facilitate related research and applications.
7 Limitations
The paper identifies limitations in data coverage and balance, prompt coverage, and Chinese-language precision. These constraints bound how broadly its linguistic and detection conclusions should be interpreted.
- The collected data remain insufficient in amount and range, with imbalanced sources, styles, and languages.The authors call for more diverse data to improve linguistic analyses and content detection.
- All collected ChatGPT answers use no special prompts, so conclusions primarily reflect ChatGPT’s general style or state.Prompts such as role-playing Shakespeare may bypass detectors or make the conclusions untenable.
- Because ChatGPT may be trained mainly on English rather than Chinese data, conclusions from HC3-Chinese may not always be precise.
A.1 HC3 Dataset Splits Creation
HC3 combines question-answering datasets and wiki-derived concepts across English and Chinese domains. The construction uses source-specific splits, ChatGPT responses to collected questions, and additional instructions where needed to align answer styles.
- Data sources: The dataset is constructed mainly from public QA datasets with expert or high-voted answers and from wiki concepts and explanations.Wiki explanations serve as human answers, while concepts are used to construct questions.
- English splits: HC3-English contains five splits: reddit_eli5, open_qa, wiki_csai, medicine, and finance.
- Chinese splits: HC3-Chinese contains seven splits spanning open_qa, baike, nlpcc_dbqa, medicine, finance, psychology, and law.
A.2 Additional Results
The Chinese-corpus dependency-relation results are consistent with the main paper. Figure 6 compares the top-30 dependency relations and corresponding dependency distances between human and ChatGPT answers, while further detailed results are available online.
- Chinese dependency-relation results are broadly consistent with the main paper’s conclusions.
- Figure 6 compares the top-30 dependency relations and corresponding dependency distances in human and ChatGPT answers from HC3-Chinese.The upper panel shows dependency relations, and the lower panel shows corresponding dependency distances; results are sorted by human-answer relation proportions.
- Detailed vocabulary, sentiment, and dependency-parsing results for each data source are available in the project’s GitHub repository.
A.3 Human Evaluations Examples
Evaluation examples for the human evaluations are provided through the project’s GitHub repository.
- The repository provides examples illustrating the human evaluations.
- Readers can access the human-evaluation examples through the project GitHub repository.
- The evaluation examples are available online for further inspection.