Confidence labels and source links changed how people trusted an AI answer, but only until they knew whether it was right
AspireAI community · arXiv preprint
Chatbot interfaces increasingly decorate answers with a confidence label or a citation. A mixed-method study - 8 interviews, a 1,000-person survey and an experiment with 5,151 participants on gardening, math and medical questions - asks whether those decorations actually calibrate trust.
The answer is partly. People read confidence indicators in wildly different ways, which the authors treat as a design problem rather than a user error, and sources were viewed as crucial even when users worried about both source reliability and the effort of checking them. In the experiment, confidence and source indicators did shift initial perceptions of trust and accuracy. But once participants learned whether the response was correct, those effects disappeared: F(5, 5115) = 1 for both trust and the indicator comparison. Medical questions drew skepticism regardless of which indicator was shown. Survey ratings put the highest stakes on health and medical advice at 79%, legal matters at 77% and financial planning at 72%.
The design lesson: indicators can set an expectation, but they cannot substitute for correctness. Caveat: the study measures perceived trust, not answer quality, and covers three topic types.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.