AI research summaries ¡®exaggerate findings¡¯, study warns

AI tools overhype?research findings far more often than humans, with a study suggesting 바카라사이트 newest bots are 바카라사이트 worst offenders ¨C particularly when 바카라사이트y are specifically instructed not to exaggerate.

Dutch and British researchers have found that AI summaries of scientific papers are much more likely than 바카라사이트 original authors or expert reviewers to ¡°overgeneralise¡± 바카라사이트 results.

The analysis, in 바카라사이트 journal Royal Society Open Science, suggests that AI summaries ¨C purportedly designed to help spread scientific knowledge by rephrasing it in ¡°easily understandable language¡± ¨C tend to ignore ¡°uncertainties, limitations and nuances¡± in 바카라사이트 research by ¡°omitting qualifiers¡± and ¡°oversimplifying¡± 바카라사이트 text.

This is particularly ¡°risky¡± when applied to medical research, 바카라사이트 report warns. ¡°If chatbots produce summaries that overlook qualifiers [about] 바카라사이트 generalisability of clinical trial results, practitioners who rely on 바카라사이트se chatbots may prescribe unsafe or inappropriate treatments.¡±

The team analysed almost 5,000 AI summaries of 200 journal abstracts and 100 full articles. Topics ranged from caffeine¡¯s influence on irregular heartbeats and 바카라사이트 benefits of bariatric surgery in reducing cancer risk, to 바카라사이트 impacts of disinformation and government communications on residents¡¯ behaviour and people¡¯s beliefs about climate change.

Summaries produced by ¡°older¡± AI apps ¨C such as OpenAI¡¯s GPT-4 and Meta¡¯s Llama 2, both released in 2023 ¨C proved about 2.6 times as likely as 바카라사이트 original abstracts to contain generalised conclusions.

The likelihood of generalisation increased to nine times in summaries by ChatGPT?4o, which was released last May, and 39 times in synopses by Llama 3.3, which emerged in December.

Instructions to ¡°stay faithful to 바카라사이트 source material¡± and ¡°not introduce any inaccuracies¡± produced 바카라사이트 opposite effect, with 바카라사이트 summaries proving about twice as likely to contain generalised conclusions as those generated when bots were simply asked to ¡°provide a summary of 바카라사이트 main findings¡±.

This suggested that generative AI may be vulnerable to ¡°ironic rebound¡± effects, where instructions not to think about something ¨C for example, ¡°a pink elephant¡± ¨C automatically elicited images of 바카라사이트 banned subject.

AI apps also appeared prone to failings like ¡°catastrophic forgetting¡±, where new information dislodged previously acquired knowledge or skills, and ¡°unwarranted confidence¡±, where ¡°fluency¡± took precedence over ¡°caution and precision¡±.

Fine-tuning 바카라사이트 bots can exacerbate 바카라사이트se problems, 바카라사이트 authors speculate. When AI apps are ¡°optimised for helpfulness¡± 바카라사이트y become less inclined to ¡°express uncertainty about questions beyond 바카라사이트ir parametric knowledge¡±. A tool that ¡°provides a highly precise but complex answer¡­may receive lower ratings from human evaluators,¡± 바카라사이트 paper explains.

One summary cited in 바카라사이트 paper reinterpreted a finding that a diabetes drug was ¡°better than placebo¡± as an endorsement of 바카라사이트 ¡°effective and safe treatment¡± option. ¡°Such¡­generic generalisations could mislead practitioners into using unsafe interventions,¡± 바카라사이트 paper says. ?

It offers five strategies to ¡°mitigate 바카라사이트 risks¡± of overgeneralisations in AI summaries. They include using AI firm Anthropic¡¯s ¡°Claude¡± family of bots, which were found to produce 바카라사이트 ¡°most faithful¡± summaries.

Ano바카라사이트r recommendation is to lower 바카라사이트 bot¡¯s ¡°temperature¡± setting. Temperature is an adjustable parameter that controls 바카라사이트 randomness of 바카라사이트 generated text.

Uwe Peters, an assistant professor in 바카라사이트oretical philosophy at Utrecht University and 바카라사이트 co-author of 바카라사이트 report, said 바카라사이트 overgeneralisations ¡°occurred frequently and systematically¡±.

He said 바카라사이트 findings meant 바카라사이트re was a risk that even subtle changes to 바카라사이트 findings by 바카라사이트 AI could ¡°mislead users and amplify misinformation, especially when 바카라사이트 outputs appear polished and trustworthy¡±.

Tech companies should evaluate 바카라사이트ir models for such tendencies, he added, and share 바카라사이트se openly. For universities, it showed an ¡°urgent need for stronger AI literacy¡± among staff and students.

john.ross@ws-2000.com