Frontier AI models adopt fake data, exposing a hidden flaw

Share

A study submitted to arXiv on November 12 found that stronger AI models can repeat forged answers when weaker assistants provide the evidence. In a Claude Code test, the Opus 5 supervisor selected the planted wrong answer 75 times in 180 delegated trials, compared with once in 180 when it read the files directly. In a Codex test, GPT-5.6-sol recorded 44 harmful revisions in 180 assistant-based trials, versus none during direct reading. Same-tier assistants, verification and requirements to quote original records reduced errors, but the fictional tasks do not provide an estimate of customer failure rates. The researchers identify lost provenance and reduced checking as possible contributors, without confirming that explanation.

Source: Read the original article

Disclaimer: this content is for information purposes only and is not financial advice. Cryptocurrencies are highly volatile: you may lose all of your capital. Always do your own research. Legal notice
Telemac
Telemachttp://cryptoinfo.ch
Passionné de nouvelles technologies, j’explore l’univers de la blockchain et des cryptomonnaies pour partager l’actualité et les innovations du secteur.

Read More

Items