AI Labels Cut True Info Trust, Boost False Claims: Study
Get the Health newsletter
Daily health & science — research, biotech, public health, the studies worth knowing. Free.
- Teng Lin and Yiqing Zhang at the University of Chinese Academy of Social Sciences (UCASS) in Beijing published a study in the Journal of Science Communication (JCOM) documenting what they call a "truth-falsity crossover effect" — the same AI label pushes credibility down for true messages and up for false ones.
- The experiment recruited 433 participants online through the Credamo platform between March and May 2024, who rated the perceived credibility of four types of Weibo-style posts (correct or misinformation, each with or without an AI label) on a 1–5 scale.
- Researchers generated the test posts using GPT-4, adapting items from China's Science Rumor Debunking Platform into both accurate and misleading versions that were independently checked before deployment.
- Algorithm aversion produced an asymmetric reaction: participants with more negative attitudes toward AI penalized correctly labeled AI content more strongly, yet the credibility boost for misinformation was only partially reduced and varied by topic rather than disappearing.
- The authors recommend a dual-labeling approach that pairs an AI-generated tag with an explicit disclaimer that the information has not been independently verified, or adds a risk warning rather than relying on a single disclosure line.
- They also propose a graded or categorical labeling system that escalates warnings by risk level — stronger labels for medical or health content, lighter ones for lower-risk topics like new technology news — though they note both proposals need further validation.
Why it matters: Regulators pushing AI transparency labels now have controlled experimental evidence the policy can backfire: across 433 participants, the labels redistributed credibility in the wrong direction. That gives policymakers concrete grounds to move beyond simple "AI-generated" tags toward the dual-labeling and risk-tiered systems the authors propose as alternatives.
