AI research

We investigate interfaces that modify themselves based on use, optimizing the user experience from their interaction history.

Traditional interfaces are static: they present the same options, the same order, and the same structure to all users, regardless of their context, preferences, or the tasks they have performed. We propose a different approach: interfaces capable of learning from sustained interaction and reconfiguring themselves to anticipate needs, reduce friction, and adapt to different patterns of use.

This involves developing models and mechanisms capable of inferring behavior patterns from interaction history, predicting which actions or information will be relevant at each moment, and dynamically adjusting the structure, order, and presentation of the interface without sacrificing user control, predictability, or agency.

We investigate overinterpretation in AI systems: the tendency to complete, infer, or fabricate content that cannot be justified with the available evidence. We take transcription and audio analysis as the primary case, where the problem manifests with particular severity.

Transcription systems tend to overinterpret: they fill silences, attribute speakers that were not identified, paraphrase instead of transcribing literally, or draw conclusions from tone or context that the data does not support. The problem is especially severe in audio, where the evidence — the acoustic signal — is not directly inspectable by the user, making it difficult to verify the traceability of each claim.

We develop methods and prototypes to preserve the traceability of sonic evidence, distinguish what was transcribed from what was inferred, represent uncertainty over inaudible or ambiguous segments, and verify each claim against the original signal. Our goal is to build transcription and analysis systems that allow understanding what was said, how it is known, and what remains uncertain — and, from that case, general methods against overinterpretation in other modalities.

Our hypothesis is that, when these responses are used to support decisions, the combination of apparent rigor, model overconfidence, and uncritical use by the person can produce pseudo-informed decisions, i.e. decisions perceived as grounded in evidence whose actual epistemic support is insufficient, incomplete, or hard to verify.

We investigate how to detect and measure this gap between available evidence, communicated information, and the resulting decision, and what mechanisms can reduce it without eliminating the usefulness of AI systems as decision-support tools.