Roughly half of 49 two-person teams, randomly assigned, were hit mid-task with an unannounced jump in difficulty. They felt it โ perceived difficulty went from 3.14 to 6.36 on a ten-point scale. Their behaviour changed โ speaking time tilted toward the partner holding the information (ฯยฒ(1) = 8.48, p = .004). Their brains registered nothing: no disruption-related change in interbrain synchrony at level, slope, or temporal recurrence, all Time ร Group interactions p > .05 (Rรฉveillรฉ et al., 2026).
If you are deciding what counts as a team-health signal in your operation, that contrast is the whole brief. The most expensive measurement in the room sat still while a free one moved.
What 49 Teams Did When the Task Broke
The design is why this result is worth a Head of Operations' attention rather than a neuroscientist's.
Ninety-eight healthy adults (mean age 22.5 ยฑ 3.22; 69 female) were assigned to 49 dyads, each completing a 20-minute computer-based navigation task under simultaneous fNIRS hyperscanning. One partner โ the Guide โ knew the route and instructed; the other moved the cursor. Half the dyads met an unannounced mid-task spike in difficulty; half did not. The protocol was preregistered on the Open Science Framework before data collection, which was run January to March 2025 (Rรฉveillรฉ et al., 2026).
Three things were established before the headline result, and they matter because they close off the easy dismissals.
The synchrony was real. Interbrain synchrony reliably emerged during the task: ฯยฒ(1) = 50.24, p < .001. The instrument was not measuring noise.
The disruption was real. Performance degraded sharply (ฯยฒ(1) = 24.99, p < .001) and perceived difficulty in the experimental group nearly doubled, 3.14 to 6.36 against a flat control (4.00 to 4.20).
And the teams did respond โ just not where the sensors were pointed. Total speaking time did not move at all (ฯยฒ(1) = 0.21, p = .644; control 453โ463 seconds, experimental 461โ477). What moved was the distribution. The speech ratio between Guide and Drawer shifted significantly (ฯยฒ(1) = 8.48, ฮฒ = 3.17, 95% CI [0.10, 6.25]), with the Guide taking a larger share as the task got harder. The volume of communication was a non-signal. The shape of it was the signal.
Why Interbrain Synchrony Looked Like a Team-Health Signal
This is not a case of a bad idea being debunked. It is a case of a good correlation being asked to do a job it was never tested for โ and that distinction is one you should be able to make about any metric on your dashboard.
The same lead author published a systematic review and meta-analysis in 2024: 41 studies drawn from an initial 229, representing 1,326 teams, with three multi-level meta-analyses finding statistically significant positive associations between interbrain synchrony and inputs, processes, and outputs of teamwork. Its conclusion argued for continued use of the measure (Rรฉveillรฉ et al., 2024).
So synchrony does track teamwork โ across teams. What the 2026 experiment tested is different: does it track teamwork within a team, over time, as conditions change? It did not. The authors' own framing is that interbrain synchrony "may reflect a stable collective state rather than a reactive one."
The gap between those two questions is the one that quietly breaks dashboards. A measure validated on between-group differences gets deployed to detect within-group change, and nobody re-runs the validation, because the number is already on the wall and it moves a little every quarter.
Read that as a measurement-class problem, not a neuroscience problem. A stable trait can discriminate between teams and still be useless as a live readout of whether a given team is coping right now. Engagement scores, culture indices, personality composites, and most annual survey instruments sit in exactly the same class. They tell you which teams differ. They do not tell you that something broke on Tuesday.
The Signal That Moved Was Free
Conversational airtime distribution is not a novel measure. Woolley, Chabris, Pentland, Hashmi and Malone, across two studies with 699 people working in groups of two to five, found a general collective intelligence factor that was not strongly correlated with members' average or maximum individual intelligence โ but was correlated with average social sensitivity and with the equality in distribution of conversational turn-taking (Woolley et al., 2010).
The 2026 experiment adds the dynamic half of that picture. Turn-taking distribution is not just a stable correlate of good teams; it is a measure that moves within minutes when a team hits something it cannot handle. That is the property a diagnostic needs and a trait does not have.
What "airtime shifted" actually means on the floor
In the experiment, the Guide โ the partner holding the information โ took a larger share of talk as difficulty rose. Translate the mechanism rather than the setting: when work gets hard, conversation collapses toward whoever holds the critical knowledge, and the rest of the team's contribution thins out. Total talk stays constant, so a meeting-minutes dashboard sees nothing. Only the ratio changes.
One caution before anyone builds a tracker. The finding is about a team's distribution against its own baseline, not about individuals against each other. The moment a concentration metric is attributed to a person, it stops measuring whether the team is coping and starts measuring who talks โ and people adjust to the metric rather than to the work.
Every operation already has this data in a form nobody looks at: who spoke in standup, who raised the blocker, who answered in the incident channel, which three names appear in every thread about a given system.
The Honest Counter
Four limits, stated plainly, because this is a null result and null results get over-read in both directions.
The effect is small. The speech-ratio shift explained 9% of the variance (marginal Rยฒ = 0.09). It is a detectable signal, not a dominant one.
The groups were not well matched at baseline on the measure that worked. The control group's starting speech ratio was 13.11 against the experimental group's 5.68, with very large standard deviations on both. Treat the absolute ratios as noisy; the defensible claim is the within-group direction of change, not a benchmark number you can port.
Dyads are not teams. Forty-nine pairs of young strangers (mean age 22.5) doing a lab navigation task is not a cross-functional team under a real deadline with history, status differences, and consequences. The mechanism travels; the parameters do not.
The authors offer a methodological escape hatch, and it is fair. They explicitly discuss the operationalisation of interbrain synchrony as an alternative explanation for the null. A different synchrony metric might have reacted. This is one well-run experiment against a measure, not a verdict on the measure.
Also worth noting, since precision matters in a paper this careful: the results text renders the speech-ratio p-value as ".04" in one clause, while the abstract, the fixed-effect estimate, and the table all report .004. The latter is correct.
More Sensors Is Not More Signal
The reflex when a measure disappoints is to buy a better one. The evidence on that reflex, in an adjacent domain, is not encouraging.
Ravid, White, Tomczak and colleagues meta-analysed 94 studies covering more than 23,000 participants on electronic performance monitoring and found no overall association with better performance across speed, quantity, helping behaviour, or counterproductive behaviour โ while finding consistent associations with higher stress, lower job satisfaction, and weaker fairness perceptions (Ravid et al., Personnel Psychology).
The pattern across both literatures is the same: instrumentation density is not the binding constraint. Choosing a measure that changes when the thing you care about changes is the binding constraint. Interbrain synchrony is an extreme illustration โ about as high-fidelity a sensor as workplace research has, and it sat flat through a disruption that a transcript would have caught.
What to Instrument Before Your Next Incident Review
Three moves, none of which needs a procurement cycle.
1. Make contribution distribution an observable, not an impression. For your three or four most load-bearing recurring meetings, record who contributed, not for how long in aggregate. You are watching for a shift in the ratio against that team's own recent baseline โ not against a cross-team benchmark, which this study specifically does not support.
2. Add a distribution question to incident reviews. "Did airtime concentrate on one person during this?" is a faster read on whether the team was coping than any sentiment question, and it is answerable from the channel history you already retain. Concentration is the finding; concentration is what you look for.
3. Stop treating trait-class measures as state-class alarms. Audit the measures you currently escalate on. For each one, ask a single question: has this number ever moved before someone told me there was a problem? If it has not, it is a segmentation tool, not a warning system โ keep it, but stop waiting for it to warn you. Structured, multi-signal competency measurement belongs in the trait class and earns its keep there: Scovai's performance and succession layer documents how per-cycle competency data and 360ยฐ divergence synthesis are structured, which is the right home for that question โ a different question from the one a live team-health signal answers.
The deeper point is about where the useful signal tends to live. The field spent a decade establishing that synchronised brains correlate with good teamwork, and it was right to. But the measurement that told you a team was in trouble, in real time, was already in the room โ unpriced, unlogged, and sitting in a transcript.
Your next incident review has a transcript. Read who stopped talking.
