Participants in a UK experiment became better at judging political claims after researching the topics with a chatbot. Their average research time was 17.94 minutes, against 19.82 minutes with Google. The gain was practical: better-informed answers after a shorter session.[1]
The study appeared in PNAS Nexus on 6 October. Its full June author manuscript describes a main trial involving 1,147 adults, assigned to GPT-4o, Claude-3.5, Mistral or Google. Each researched two of four topics: immigration, climate change, criminal justice and COVID-19.[1][2]
Researchers measured agreement with 16 true and false statements before and after the task, including topics participants had not researched. Responses shifted towards accurate information on the researched topics. The statistical analysis found similar learning gains with AI and Google, while leaving small differences unresolved.[1]
Measure what the reader takes away
A useful information tool has to get something across. An accurate answer that a reader misunderstands and an inaccurate answer that a reader recognises as wrong produce very different outcomes. Looking only at generated text misses that distinction; looking only at satisfaction misses it too.
The trial measured immediate, directed research, leaving lasting knowledge and voting behaviour untested.[1]
Here, the time saved came alongside more accurate judgments at the end of the task.[1]
Sources & further reading
- Conversational AI increases political knowledge as effectively as self-directed internet search: full author manuscript, version 5 Lennart Luettgau and colleagues / arXiv · 25 Jun 2026 · Accessed 28 Oct 2026
- Conversational AI increases political knowledge as effectively as self-directed internet search PNAS Nexus · 6 Oct 2026 · Accessed 28 Oct 2026