All That's 'Human' Is Not Gold: Evaluating Human Evaluation of Generated\n Text

2021/06/30 by Elizabeth Clark, Elizabeth A. Clark, Clark, Elizabeth +10 · 57 citations
Computer Science · Mathematics · Psychology · #Artificial intelligence #Computer science #Explainable Artificial Intelligence (XAI) #Fluency #Gold standard (test) #Mathematics #Mathematics education #Natural Language Processing Techniques #Natural language #Natural language generation #Natural language processing #Psychology #Statistics #Text generation #Topic Modeling #cs.CL

paper · pdf · doi:10.48550/arxiv.2107.00061

published in arXiv (Cornell University) (Cornell University) · references added, corrected typo

openalex publication_date 2021/06/30 · arxiv created 2021/07/07 · arxiv updated 2021/07/08 · openalex created_date 2022/07/25 · openalex updated_date 2026/08/06

Abstract

Human evaluations are typically considered the gold standard in natural\nlanguage generation, but as models' fluency improves, how well can evaluators\ndetect and judge machine-generated text? We run a study assessing non-experts'\nability to distinguish between human- and machine-authored text (GPT2 and GPT3)\nin three domains (stories, news articles, and recipes). We find that, without\ntraining, evaluators distinguished between GPT3- and human-authored text at\nrandom chance level. We explore three approaches for quickly training\nevaluators to better identify GPT3-authored text (detailed instructions,\nannotated examples, and paired examples) and find that while evaluators'\naccuracy improved up to 55%, it did not significantly improve across the three\ndomains. Given the inconsistent results across text domains and the often\ncontradictory reasons evaluators gave for their judgments, we examine the role\nuntrained human evaluations play in NLG evaluation and provide recommendations\nto NLG researchers for improving human evaluations of text generated from\nstate-of-the-art models.\n

Cited by

• Scalable and Personalized Oral Assessments Using Voice AI
• From NLG Evaluation to Modern Student Assessment in the Era of ChatGPT: The Great Misalignment Problem and Pedagogical Multi-Factor Assessment (P-MFA)
• Security and Detectability Analysis of Unicode Text Watermarking Methods Against Large Language Models
• Identifying Bias in Machine-generated Text Detection
• Ideal Attribution and Faithful Watermarks for Language Models
• Idea-Gated Transformers: Enforcing Semantic Coherence via Differentiable Vocabulary Pruning
• ART: Adaptive Response Tuning Framework -- A Multi-Agent Tournament-Based Approach to LLM Response Optimization
• A Cross-Cultural Assessment of Human Ability to Detect LLM-Generated Fake News about South Africa
• Detecting LLM-Assisted Academic Dishonesty using Keystroke Dynamics
• Bytes of a Feather: Personality and Opinion Alignment Effects in Human-AI Interaction
• Advancing Machine-Generated Text Detection from an Easy to Hard Supervision Perspective
• A Principle-based Framework for the Development and Evaluation of Large Language Models for Health and Wellness
• On the Detectability of LLM-Generated Text: What Exactly Is LLM-Generated Text?
• How Sampling Affects the Detectability of Machine-written texts: A Comprehensive Study
• Every Step Counts: Decoding Trajectories as Authorship Fingerprints of dLLMs
• Promptimizer: User-Led Prompt Optimization for Personal Content Classification
• Lemma Dilemma: On Lemma Generation Without Domain- or Language-Specific Training Data
• OptAgent: Optimizing Query Rewriting for E-commerce via Multi-Agent Simulation
• Beyond a Single Judge: The Evidence-Grounded, Social-Weighted Persona Panel for Generative UI Evaluation
• Evaluating the Creativity of LLMs in Persian Literary Text Generation
• The AI Memory Gap: Users Misremember What They Created With AI or Without
• Paladin: Defending LLM-enabled Phishing Emails with a New Trigger-Tag Paradigm
• ABCD-LINK: Annotation Bootstrapping for Cross-Document Fine-Grained Links
• LLM Encoder vs. Decoder: Robust Detection of Chinese AI-Generated Text with LoRA
• The illusion of a perfect metric: Why evaluating AI's words is harder than it looks
• Do language models accommodate their users? A study of linguistic convergence
• Designing Effective Digital Literacy Interventions for Boosting Deepfake Discernment
• Multi-Agent-as-Judge: Aligning LLM-Agent-Based Automated Evaluation with Multi-Dimensional Human Evaluation
• Beyond Easy Wins: A Text Hardness-Aware Benchmark for LLM-generated Text Detection
• Psychology-Driven Enhancement of Humour Translation
• The Thin Line Between Comprehension and Persuasion in LLMs
• Position: Machine Learning Conferences Should Establish a "Refutations and Critiques" Track
• Revisiting Compositional Generalization Capability of Large Language Models Considering Instruction Following Ability
• Large Language Models as 'Hidden Persuaders': Fake Product Reviews are Indistinguishable to Humans and Machines
• Min-p, Max Exaggeration: A Critical Analysis of Min-p Sampling in Language Models
• Search Arena: Analyzing Search-Augmented LLMs
• Explainability-Based Token Replacement on LLM-Generated Text
• An evaluation of LLMs for generating movie reviews: GPT-4o, Gemini-2.0 and DeepSeek-V3
• Intuitionistic Fuzzy Sets for Large Language Model Data Annotation: A Novel Approach to Side-by-Side Preference Labeling
• Frankentext: Stitching random text fragments into long-form narratives
• Is Your LLM-Based Multi-Agent a Reliable Real-World Planner? Exploring Fraud Detection in Travel Planning
• Reassessing Collaborative Writing Theories and Frameworks in the Age of LLMs: What Still Applies and What We Must Leave Behind
• Domain Gating Ensemble Networks for AI-Generated Text Detection
• Margin-Adaptive Confidence Ranking for Reliable LLM Judgement
• LLM Output Detectability and Task Performance Can be Jointly Optimized
• When LLM Essays Outscore Student Essays: What a Korean Writing Rubric Rewards and Where Readers Disagree
• Humans can learn to detect AI-generated texts, or at least learn when they can't
• JudgmentBench: Comparing Rubric and Preference Evaluation for Quality Assessment
• Measuring and Mitigating Persona Distortions from AI Writing Assistance
• Eroding the Truth-Default: A Causal Analysis of Human Susceptibility to Foundation Model Hallucinations and Disinformation in the Wild
• Gradual Cognitive Externalization: From Modeling Cognition to Constituting It
• MEQA: A Meta-Evaluation Framework for Question & Answer LLM Benchmarks
• Once a Response, Always a Response: Detecting LLM-generated Text via Latent Prompt Restoration
• Labeling Messages as AI-Generated Does Not Reduce Their Persuasive Effects
• Summarizing Speech: A Comprehensive Survey
• Toward Holistic Evaluation of Recommender Systems Powered by Generative Models
• Identifying Aspects in Peer Reviews

Related

• Human or Machine: Automating Human Likeliness Evaluation of NLG Texts
• Can Large Language Models Be an Alternative to Human Evaluations?
• When Automated Assessment Meets Automated Content Generation: Examining Text Quality in the Era of GPTs
• GENIE: Toward Reproducible and Standardized Human Evaluation for Text Generation
• The Authenticity Gap in Human Evaluation