Shyamnath Gollakota's lab at the University of Washington spent the last two years teaching an earpiece to whisper in your ear at the right moment (opens in a new tab). Its newest paper teaches the same kind of agent to shut up. Jarvis joins a group conversation, checks what it hears against a document the group agreed on beforehand, and speaks only when someone states a fact wrong and nobody fixes it. The headline result is a silence: it says nothing 97 percent of the time the group corrects itself.
Jarvis, a proactive speech agent from the University of Washington, stays silent 97 percent of the time the group resolves its own mistake. Its behavior is a policy rather than a bigger model: flag a claim that contradicts a shared document, wait a few turns, and speak only if nobody corrects it. A separate Texas Tech driving study found that letting an agent defer a comment, instead of only choosing between speaking and silence, cut the sense of interruption while trust held flat. For an ambient coach the takeaway is that proactivity is a scheduling problem, and the default slot is empty.
Why would a proactive agent stay silent 97 percent of the time?
Because in most moments the useful thing is to let the humans finish. Jarvis (opens in a new tab) (arXiv:2610.07506) grounds every claim it hears in a sentence from the shared document, so "wrong" means a contradiction it can point at, not a vibe. When a statement contradicts the document or leaves a known gap, Jarvis flags it and holds. If the group self-corrects within a few turns, it never speaks. If they do not, it has two ways in: it raises a hand to ask for the floor, or it waits until nobody holds it, and when it finally talks it cites the page it is correcting. Get talked over on the way in, and it yields, then decides whether to replay, resume, or drop what it had.
The backbone is a small open-weight model wrapped in deterministic checks, not a frontier system reasoning about social dynamics. What makes the agent well-behaved is the rule and the document. That runs against how most proactive assistants are being built, where the pitch is a smarter model that will somehow learn tact on its own. The lab's bet is that tact is a policy over a shared memory, and the model is almost a detail.
How much of this survives a real conversation?
Less than the 97 percent suggests, and the paper's own setup is why. The evaluation runs on CHI-180-proactive, a synthetic dataset seeded with known gaps, errors, and self-corrections. The agent is scored in a world where mistakes are planted on purpose and a ground-truth document exists to check them against. An always-on wearable has neither. Most of what it hears is casual talk with no agreed document behind it, and nothing hands it an oracle of what counts as an error worth flagging. The live test is 23 people in a lab, which is a pilot, not a deployment.
I could not read past the abstract to grade the rest. Every arXiv host and mirror was blocked from the sandbox this week, so the per-event precision, the latency to a raised hand, and the live-study breakdown are the authors' to report, and I am scoping my read to the silence rate and the design. The catch the framing underplays: the hard problem for an ambient coach is not the waiting rule, which is simple once you have it. It is deciding what even counts as a correctable error when there is no document to point at. Jarvis assumes that part away by construction, and that is precisely the part an ear-worn coach would have to invent.
Which is why the shared document matters more than it looks. It is the piece that makes the whole thing work and the piece a wearable does not get for free. The two-brain split I argued for in July put the "when to speak" decision in a cheap always-on layer; Jarvis says that layer needs something to be right about, and a transcript of your morning is not it.
The second data point, from a car
Texas Tech ran the other half of this independently (opens in a new tab) (arXiv:2609.12503). Forty-one people rode as passengers in a simulated self-driving car while an in-cabin agent either spoke at every driving event or weighed the event and what the passenger was doing and chose to speak now, defer, or stay quiet. Two findings travel. Deferring is a real option in its own right: holding a comment kept its usefulness and cut the felt interruption. And trust did not move with the policy at all. A better interruption rule is felt as the absence of annoyance rather than as more trust, and the people who already trusted the agent just wanted it to talk more.
Put the two papers together and the intervention layer turns into plumbing. It is a queue with expiry and a handful of inputs - event consequence, what the user is doing, whether the information still matters in a minute, whether a confirmation is needed - and none of those inputs is the model's confidence in its own answer. An assistant that will not interrupt to ask has the opposite failure, but the same root: nobody has given it a scheduler.
Who this is bad news for
Anyone building proactivity as a notification engine. If your coaching app measures engagement by messages sent, every number in both papers points the wrong way for you: the better product talks less, and talking less does not even register as trust, so you cannot A/B your way to it on the metric you are watching. The teams that win the ambient-coach slot will ship a policy that defaults to nothing and can defend every time it chose to speak.
Two dated calls. By the end of Q2 2027, at least one shipping voice assistant - Alexa+, Gemini Live, or Muse - documents a hold-and-wait behavior, some version of "does not interrupt to correct until the user has had a chance to." And the first ambient coaching wearable that ships with a genuine proactivity layer will act on fewer than one in twenty of the openings it detects; the ones that fire more often will be the ones with the churn.
A coach that waits for you to fail before it says anything is harder to build than one that talks, and much harder to sell, because the demo is a device doing nothing. It is also the only kind anyone keeps wearing past the second week.