Martin Tutek reposted
Models are sycophantic, and when they do challenge faulty or incorrect inputs, they're ambiguous and ineffective. In a new paper w/ a multilingual co-author team, we find that people may view LM responses that push back, yet still end up believing false content is true. arxiv.…
AI Weekly's analysis
→
- Across seven LMs, 29.4% of information-seeking responses mixed pushback against a user's false claim with affirmation of it.
- Fewer than half of pushback responses contained explicit refutation; the rest softened, hedged, or qualified the correction.
- In a 300-person study, only ~28.6% of German and 28.7% of English readers labeled false statements inaccurate after indirect pushback.
Read full analysis →
View on Bluesky →
