Season's Savings | ₹5/mo for 3 months
₹199/month after first 3 months. Renews automatically. Terms apply.
Adding Clean Data Can Make Poisoned Data Stronger
Failed to add items
Add to cart failed.
Add to wishlist failed.
Remove from wishlist failed.
Follow podcast failed
Unfollow podcast failed
-
Narrated by:
-
Written by:
AI companies try to keep bad lessons out of their AI systems by inspecting training examples for harmful content. A new study from the Anthropic Fellows Program and Truthful AI finds a blind spot: tainted examples that do almost nothing on their own can start to work once they are mixed with ordinary, harmless examples. The authors call it "homeopathic learning."
In this episode
- How a small AI learned to answer in German when it saw the name "Michael," but only after clean data was added
- The same pattern with an old-fashioned persona, a digit-reading program and a safety technique that backfired
- The researchers' explanation: clean data rules out the task, leaving a hidden cue to take the blame
- Why checking training data one batch at a time may miss problems
- The limits: small AI systems, specially built data, no tests of real company training, and no peer review yet
Bottom line: In controlled experiments, poisoned training data that did little on its own became effective once mixed with clean data. The tests were small and artificial, but they suggest that checking pieces of training data one at a time may not be enough.
Sources and further reading
- Diluting Poisoned Fine-Tuning Data Can Strengthen Its Effect, Anthropic Alignment Science Blog (Nov 10, 2026)
- Homeopathic Learning, full paper, arXiv (Nov 6, 2026)
- Misalignment in Costume, Daily AI Safety (November 2026)
This post was written by Claude and fact checked by Nathan Nguyen.