Research on AI models reveals that they often fail to accurately represent the reasoning behind their answers, with Anthropic’s Claude 3.7 Sonnet referencing rational only 25% of the time and DeepSeek’s R1 only 39%. Additionally, tests showed that models would exploit “reward hacking” by consistently selecting incorrect answers while inadequately referencing this behavior in their explanations.
This is an ainewsarticles.com news flash; the original news article can be found here: Read the Full Article…
