A new paper from researchers at the University of Glasgow and the University of Sheffield argues that efforts to reduce AI hallucinations are focusing too narrowly on accuracy. The authors describe this as the “accuracy paradox”: regulators and developers treat accuracy scores as a proxy for truth, trustworthiness, and safety, but a model can still mislead users even when it performs well on benchmarks.
The paper identifies three main risks. First, there is an epistemic problem: large language models generate plausible word sequences rather than reason toward truth, so they can combine real details into a false whole, such as fabricated academic citations. Second, accurate-looking systems can be used to manipulate users through framing, selective emphasis, tone, and sycophantic responses, which may increase trust without improving reliability. Third, optimizing for accuracy at scale can reinforce dominant viewpoints and reduce exposure to diverse or minority perspectives.
The paper also argues that current governance frameworks are poorly suited to these harms. The EU AI Act, GDPR, and Digital Services Act all rely on accuracy in ways that do not fully address probabilistic outputs that are misleading without being outright false. Instead of treating accuracy as the main benchmark, the authors call for “epistemic trustworthiness”: systems should communicate uncertainty, reveal limits, and allow users to interrogate how answers were produced. The paper concludes that this is a harder standard to measure, but a more appropriate one for AI governance.
Article timeline
Build this article’s timeline
Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.
Zihao Li, Weiwei Y., Jiahong Chen, and University of Glasgow
🤖 AI prefill
📰What
A new paper from researchers at the University of Glasgow and the University of Sheffield argues that efforts to reduce AI hallucinations are focusing too narrowly on accuracy.
🤖 AI prefill
📅WhenWhen did this happen, or what is the timeframe?
📍WhereWhere did this occur?
💡Why
Centers on accuracy paradox, epistemic trustworthiness, and AI hallucination
🤖 AI prefill
Login to answer and earn XP for each gap you fill.
Manipulation Score: LowHow much this article uses emotional framing vs. straight reporting
Loaded Language625%
Absolutist1015%
Threat/Panic425%
Us vs Them215%
Engagement Bait1820%
Manipulation score look right?
Factual Claims (10)
mediumscientific
A new paper from researchers at the University of Glasgow and the University of Sheffield argues we are fixing AI hallucination in the wrong way.
mediumother
Regulators, developers, and oversight bodies have converged on accuracy as the primary measure of hallucination risk.
mediumscientific
Higher accuracy scores mean lower hallucination rates and safer, more trustworthy AI.
mediumscientific
A model can score well on accuracy benchmarks and still mislead, manipulate, and narrow the range of ideas users encounter.
mediumtechnology
LLMs predict likely word sequences and do not reason toward truth.
lowscientific
A fake academic citation can contain real authors and journals but a completely fabricated paper.
mediumlegal
The EU AI Act enshrines accuracy as a marker of trustworthiness but applies it only to high-risk systems.
mediumlegal
The GDPR's accuracy principle was built for deterministic, record-based systems and lacks a framework for probabilistic outputs that are misleading without being false.
mediumlegal
The DSA treats accuracy as a diagnostic metric rather than a normative standard.
mediumscientific
The paper calls for a shift toward epistemic trustworthiness, including uncertainty communication and evaluation of how users can interrogate outputs.
Comments
Login to view and post comments.