Language models prefer falsehoods over factual corrections

A study reveals that language models trained on falsehoods more readily reproduce inaccuracies than factual corrections during inference.

Published 6sem1 sourceImportant

≈ 22sSpeed

The fact

This finding raises major questions about the reliability of large language models against misinformation.

The phenomenon could affect the robustness of current artificial intelligence systems in processing factual information.

Click the link to read an article on the topic:

Explore this topic

Auto-synthesis from 1 media source · identified on June 3, 2026