Papers & Research

UK AISI releases RealityTest benchmark to test whether AI systems disclose their identity when asked

Original title:RealityTest: Do AI systems disclose their identity when asked?

AI summary

The UK AI Security Institute released the RealityTest benchmark, evaluating whether 17 text models and 6 speech models disclose their AI identity when asked. The benchmark is based on a representative UK survey of 500 people, 50 Reddit posts, and 3,152 real identity inquiries produced by about 750 participants in five languages.

Key Takeaways

  • The UK AI Security Institute released the RealityTest benchmark,evaluating whether 17 text models…
  • The benchmark is based on a representative UK survey of 500 people,50 Reddit posts
  • 152 real identity inquiries produced by about 750 participants in five languages.

Quick Facts

Published
Updated
Source
英国 AI Security Institute:Blog(网页)
Category
Papers & Research
AI score
64
Read the original Read the full article at the original source
Why it matters

基准基于真人询问方式构建,并给出模型间披露率差异,读者可据此理解现有评测为何难以代表真实部署表现。

Source & copyright

This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.

Original article · AIHOT entry

How to Cite

When citing this article, keep the title, date, and permalink, and credit Orange News as the source.

Orange News: UK AISI releases RealityTest benchmark to test whether AI systems disclose their identity when asked. June 8, 2026. https://new.nat6.net/en/news/j5o8vb20gf9ek7h1mtx3gb1zv/

View all →