AI Test Scores Are Easily Manipulated, New Study Finds
The same AI can score 60% or 90% on a test—depending on who grades it
Deep Dive
arXivLabs is a framework that lets collaborators develop and share new arXiv features directly on the arXiv website. Individuals and organizations working with arXivLabs have embraced and accepted arXiv's values of openness, community, excellence, and user data privacy.
arXiv says it is committed to these values and works only with partners who adhere to them. If you have an idea for a project that would add value for arXiv's community, you can learn more about arXivLabs.
Key Points
- AI test scores can change by 30% or more just by swapping the answer key, without any real improvement in the AI.
- Companies can cherry-pick the best-looking score to market their AI as superior.
- This means you should be skeptical of AI 'benchmark' claims and look for independent, transparent testing.
Why It Matters
You might buy or trust an AI based on misleading scores, wasting money or getting poor results.