• Login
  • Register

Work for a Member organization and need a Member Portal account? Register here with your official email address.

Project

Feeling-the-Facts: AI-Powered Wearable Fact-Checker

ACM CHI 2026

Groups

The problem: misinformation you hear, not read

Most fact-checking tools were built for text, for example, a headline you can pause on, a post you can screenshot. But a lot of misinformation reaches us through speech: in conversations, meetings, lectures, podcasts, and videos. These claims are fleeting. By the time you think to question something, the conversation has already moved on, and pulling out your phone to Google it mid-chat is awkward, slow, and easy to forget.

Decades of misinformation research show why this matters: once a false claim lodges in memory, it keeps shaping our judgments even after we've seen a correction. Corrections work best when they arrive immediately, right as the claim is heard, before repetition makes it feel more true. That's a timing problem current tools don't solve.

Our question: could a smartwatch close that gap by listening in the background, quietly checking claims against the web, and nudging the wearer at exactly the right moment? 

Building a working prototype

We built FactNudger, an end-to-end Android + Wear OS system. It transcribes conversation in real time, interprets claims in their conversational context, retrieves web evidence, and then decides on one of three actions: wait for more speech, ignore the claim, or vibrate. When it vibrates, the watch shows a short glanceable explanation, and a companion phone provides the supporting sources.

The system is deliberately conservative about when to interrupt. Being wrong isn't just an accuracy problem, it's a trust problem. Vibration is reserved for claims with credible contradicting evidence and real consequences for the listener (health, finances, safety, etc.); everything else is quietly ignored or held for more context.

How well does it work? On a random sample of 125 real-world fact-checked claims, FactNudger got the correct true/false verdict 77% of the time. On a custom set of 50 tricky conversational edge cases, such as, incomplete sentences, corrections, repeated claims, jokes, location-dependent statements, it chose the right response (wait/ignore/vibrate) in 40 of 50 cases (80%), with a mean response latency of about 3.5 seconds after someone finished speaking.

Does it actually change what people believe?

Does a buzz on the wrist change how people process information. To test this rigorously, we built a Wizard-of-Oz version of FactNudger with a fixed, known accuracy (87.5%, in line with state-of-the-art video fact-checkers) so we could isolate the effect of real-time feedback itself from the technical performance of any one AI model.

We ran a within-subjects study with 34 participants, who each watched short video compilations of true and false claims: once while wearing FactNudger, and once with no support. Order was counterbalanced, and 1–2 deliberate AI errors were seeded into each video to see how people responded to an imperfect system.

We found that:

Belief in false claims dropped significantly when wearing the device — and belief in true claims was preserved, meaning the nudge didn't make people indiscriminately more skeptical.

Confidence in rejecting false claims went up, not just belief accuracy.

Verification activity roughly doubled (pauses, rewinds, web searches, watch glances) compared to the no-wearable condition, with no added subjective cognitive workload (NASA-TLX showed no significant difference).

The effect held up consistently across participants regardless of how much they personally cared about a topic, their prior exposure to the claim, or their general disposition toward open-minded thinking.

The catch: errors matter, and not symmetrically. A false positive, i.e. the watch flagging a true claim as false, made people significantly more skeptical of accurate information. A false negative, i.e. silently missing a false claim, had much less effect, since people mostly fell back on their own instincts. In other words, people leaned on the device more than they should have when it spoke up, but weren't badly hurt when it stayed quiet. That's a useful asymmetry for design, but it also means trust calibration (confidence scores, visible sources, honest uncertainty) should be central to using this kind of system responsibly.

What's next: taking it into the wild

The lab study told us the concept works: instant, body-integrated feedback measurably improves truth discernment and verification behavior without extra cognitive burden. But it was Wizard-of-Oz, using pre-scripted, pre-verified video content in a controlled setting, which likely inflated how often people actually engaged in verification compared to everyday life.

We plan to run an in-the-wild study with the fully functioning end-to-end FactNudger system, testing it in real, unscripted conversations, with all the messiness that entails: background noise, ambiguous speech, imperfect transcription, and realistic latency. The goal is to see whether the belief-change and verification effects we saw in the lab hold up when the system has to earn trust in the real world, not just a curated one.