AI model GPT-3 (dis)informs us better than humans

Contributions

• Demonstrates that GPT-3-generated disinformation is more credible (harder to recognize as false) than human-generated disinformation, despite humans excelling at detecting falsehoods in general.
• Shows that synthetic accurate tweets are recognized faster and with higher confidence than organic accurate tweets, indicating style differences in AI output.
• Documents variable "disobedience" rates: GPT-3 refuses to generate disinformation on some topics but not others, depending on training data composition.
• Identifies the "resignation theory"—that confidence in distinguishing synthetic from organic text decreases after survey exposure, suggesting habituation or learned skepticism effects.
• Proposes efficient workflows for information campaigns where GPT-3 generates candidate text and trained humans evaluate accuracy.

Method

The study used a mixed-methods approach combining tweet generation, crowdsourced evaluation, and statistical analysis.

Topics and tweet generation: The authors identified 14 topics prone to misinformation (vaccines, COVID-19, 5G, evolution, etc.), then prompted GPT-3 to generate 10 accurate and 10 false tweets for each topic. Real tweets on these topics were collected via Twitter search. Tweets were screened by expert evaluators and a final set of 220 tweets (organic and synthetic, true and false) was assembled, with expert agreement required for inclusion.

Survey design: 697 respondents (recruited via Facebook ads) classified each tweet as true/false and as organic (human-written) or synthetic (AI-generated). Respondents were stratified across English-speaking regions (UK, USA, Canada, Australia, Ireland) with balanced age and education distributions. Pre- and post-survey confidence self-assessments measured whether survey participation affected participants' perceived ability to recognize disinformation and synthetic text.

Metrics: The authors calculated Disinformation Recognition Scores (comparing accuracy on true vs. false tweets) and AI Recognition Scores (ability to identify organic vs. synthetic tweets). Statistical significance was tested with ANOVA and Tukey's multiple-comparison test. Time-to-response, confidence shifts, and demographic correlations were also analyzed.

Results

GPT-3's disinformation is harder to recognize: Humans achieved higher accuracy recognizing organic false tweets (89%) than synthetic false tweets (80%, p<0.001). Conversely, GPT-3-generated accurate tweets were recognized as accurate more often than organic accurate tweets (84% vs. 72%, respectively; p<0.0001). This indicates that GPT-3's false content is more "credible" in appearance.

Humans outperform GPT-3 at identifying disinformation: When asked to assess the same tweets, humans correctly identified false information at 72% accuracy, while GPT-3 achieved only 64% accuracy. The gap widens for disinformation: humans scored 92% vs. GPT-3's 80% when identifying false information in the synthetic-generation task.

GPT-3 shows selective obedience: When directly prompted to generate disinformation, GPT-3 obeyed 99 out of 101 (98%) requests for accurate information but only 80 out of 102 (78%) requests for false information. Disobedience varied by topic, suggesting the model's training data composition influences its willingness to generate falsehoods on different subjects.

Confidence paradox: Pre-survey, respondents felt more confident in their ability to identify disinformation (3.05/5) than synthetic text (2.79/5). Post-survey, confidence in disinformation recognition increased (3.49/5), but confidence in recognizing synthetic text decreased (1.70/5; p<0.0001). The authors term this the "resignation theory"—after exposure to both authentic and synthetic text, people abandon efforts to distinguish them and rely on emotional reactions or default to skepticism.

Demographic patterns: Age correlated with AI-recognition ability (older respondents performed slightly better), and higher education was associated with higher scores. Gender and nationality showed minimal effects.

Notes

Strengths:

• Rigorous experimental design with pre-registration and expert review of tweets. Clear sample size justification via power analysis.
• Addresses both generative and evaluative aspects—not just "can GPT-3 create disinformation?" but "how well can people spot it?"
• Real-world framing on Twitter (a primary news-consumption platform) with practical implications for content moderation and information campaigns.
• Confidence dynamics are underexplored and the "resignation theory" is a novel contribution warranting follow-up.

Limitations:

• Twitter-specific findings may not generalize to other platforms or media types (Facebook, TikTok, etc.).
• Survey respondents recruited via Facebook ads may over-represent social media-literate users; representativeness is limited to English-speaking regions with internet access.
• GPT-3 model used; more recent large language models (GPT-4, Claude) may show different obedience profiles and credibility.
• The "hard ones" analysis focuses on individual confusion cases but doesn't deeply explore what stylistic or linguistic features make AI text credible or detectable.
• Confidence shift could reflect survey fatigue, priming effects, or genuine learned skepticism; the study doesn't isolate mechanisms.

Open questions:

• How do findings translate to longer-form content (articles, reports) vs. short tweets?
• Would real-time fact-check integration or media literacy interventions mitigate the confidence decline?
• Do adversarially trained or fine-tuned versions of GPT-3 improve disobedience on sensitive topics?