Neurosity’s New Headset Achieves Real-Time Focus Scoring With 94% Accuracy

Neurosity, the Brooklyn based brain computer interface company that shipped its first consumer headset, the Crown, through Crowd Supply in 2022, says its newest device can score a wearer’s focus in real time with 94 percent accuracy. That number is going to travel. It will show up in press releases, in productivity app marketing copy, maybe in a pitch deck or two. Before it does, it’s worth asking what the number is actually a measurement of, because the answer is more specific, and more limited, than “how focused you are.”

Here’s the short version: 94 percent accuracy describes how well a machine learning classifier agrees with the labels it was trained on. It does not describe your subjective experience of concentrating, and it does not mean the headset has solved the problem of reading attention out of the brain. That problem is a lot older and a lot harder than any single product launch, and the history of consumer EEG is full of companies that hit good numbers in a lab and then watched them wobble the moment a real customer put the headset on at a real desk.

What’s Actually Under the Headband

Neurosity’s hardware uses dry, comb-style electrodes, no gel, no wet prep, arranged across the forehead and scalp to pick up electrical activity from the prefrontal and parietal regions, the parts of cortex most consistently tied to sustained attention and working memory in the neuroscience literature. The Crown shipped with eight channels running through an onboard processor the company calls its own embedded Linux stack, meaning the signal gets filtered and turned into features on the device itself rather than streamed raw to a phone.

The features that actually feed a “focus score” are almost certainly built on band power, the relative strength of different EEG frequency bands like theta (roughly 4 to 8 Hz) and beta (roughly 13 to 30 Hz). This isn’t new territory. In 1995, researchers at NASA Langley, led by Alan Pope, published what’s still called the engagement index, a ratio of beta power to the sum of alpha and theta power, originally built to catch pilots drifting toward drowsiness and inattention during long, monotonous flights. Nearly every consumer “focus” or “attention” metric sold since then, from NeuroSky’s eSense scores to Muse’s calibration routines, descends from some variant of that same math.

What’s changed is the classifier sitting on top of the band power features. Where Pope’s original engagement index was a fixed formula, Neurosity’s system almost certainly uses a trained model, probably something like gradient boosted trees or a lightweight recurrent network, that learns thresholds from labeled examples of “focused” and “not focused” EEG segments. That’s a real technical step forward. It’s also exactly where the 94 percent number gets complicated.

The Number That Everyone Will Quote, and What It Leaves Out

Accuracy, in the machine learning sense, means the classifier correctly matched a label a very high percentage of the time. The entire question is where those labels came from. If a participant in a validation study sat through a task designed to demand sustained attention, like a continuous performance test or an n-back working memory task, and researchers labeled those windows “focused” versus windows of rest or mind wandering as “not focused,” then the 94 percent figure is really answering: can the model tell a demanding cognitive task apart from idle rest, using this person’s own brain as the training set.

A classifier accuracy score measures agreement with its own labels, not whether you were paying attention.

That’s a genuinely useful capability. It’s also a narrower claim than “knows when you’re focused,” because real-world focus doesn’t arrive as a clean binary between a Stroop task and a nap. You might be deeply absorbed in debugging code while your frontal theta looks nothing like it did during the calibration task that trained the model. You might be staring at a screen, technically alert, but mentally replaying an argument from the night before, a state EEG band power is notoriously bad at distinguishing from genuine task engagement.

There’s also the question of whose brain the model was trained on. Within-subject accuracy, where a model is trained and tested on the same person’s EEG, routinely posts numbers in the 90s in published BCI literature. Across-subject accuracy, where a model trained on one group of people has to generalize to someone new putting the headset on for the first time, tends to drop, sometimes sharply, because EEG signatures of attention vary with skull thickness, hair, electrode contact, caffeine, sleep debt, and plain individual variation in cortical anatomy. Neurosity hasn’t published the validation methodology behind its 94 percent claim in a peer reviewed venue, so there’s no way from the outside to know which of these scenarios produced the number.

Dry Electrodes Pick Up More Than Thought

The other honest problem is noise. Dry electrodes are what make a headset like this comfortable enough to wear at a desk for hours, but they’re also more sensitive to motion artifact than the wet, gelled electrodes used in clinical EEG labs. A jaw clench, an eyebrow raise, even talking, all produce electrical signals in the same frequency ranges the classifier is parsing for focus. Clinical EEG technicians spend years learning to visually identify and discard these artifacts before analysis. A consumer headset running inference in real time has to make that call automatically, in milliseconds, without a technician checking its work.

Eight dry electrodes can’t always tell deep focus from a clenched jaw.

This is a known, published limitation across the dry electrode EEG category, not a Neurosity specific flaw. Academic groups that study consumer grade EEG hardware have repeatedly found that signal quality degrades with movement and with poor scalp contact in a way that wet electrode systems don’t suffer from nearly as much. None of this makes the Crown or its successor useless. It means the 94 percent figure almost certainly describes performance under relatively controlled conditions, sitting still, following a defined task, not the messier reality of someone wearing the headset while actually working.

Where This Sits Next to Everything Else Claiming to Measure Attention

Neurosity isn’t the first company to sell a number that claims to quantify focus, and it’s useful to see where it lands against the rest of the field, including devices built for very different stakes.

NeuroSky put an “attention” score on the market back in 2009, cheap enough to end up inside toys like the Star Wars Force Trainer, running on a single forehead electrode and an algorithm the company never fully disclosed. Emotiv has published more of its methodology and has some peer reviewed backing, though independent replications of its attention metrics have produced inconsistent results across labs. On the far end sits Advanced Brain Monitoring’s B-Alert system, a wet electrode, clinically validated platform used in FAA and military research on pilot and operator drowsiness, where the stakes of getting a false “alert” reading wrong are a lot higher than a missed productivity notification.

Neurosity’s new device sits closer to the consumer end of that spectrum than the clinical end, which isn’t a criticism so much as a category description. It’s a wellness and productivity product, not a medical device, and it isn’t regulated or validated the way one would be.

What a Good Lab Number Doesn’t Promise You at Your Desk

There’s a real, separate body of research on EEG neurofeedback worth bringing in here, because it shows how these numbers can diverge from real-world benefit even when the underlying classification is sound. A 2009 meta-analysis led by Martijn Arns found neurofeedback training produced effect sizes on inattention in ADHD comparable to stimulant medication, a genuinely striking result. Follow-up studies using blinded, sham-controlled designs, including work associated with Edmund Sonuga-Barke, found those effects shrank substantially once participants and raters didn’t know which feedback was real. The lesson wasn’t that EEG feedback is worthless. It’s that an impressive number generated under one set of conditions can mean something much smaller once you control for expectation, placebo, and the gap between a controlled study and daily life.

Ninety four percent accurate is a claim about the model, not about your mind.

Apply that same caution here. A focus score that’s 94 percent accurate at separating a demanding lab task from rest tells you the hardware and the algorithm behind it are doing something real with the signal. It doesn’t tell you whether wearing the headset during a workday changes how much actual work gets done, whether the number holds up outside the conditions it was trained on, or whether a wearer two months in will trust a device that occasionally tells them they’re unfocused during a task they’d swear they were locked into.

The most useful question for Neurosity isn’t whether 94 percent is an impressive number. It’s whether the company publishes the validation protocol behind it, including whether the test set included people never seen during training, so outside researchers can check the claim the way NASA’s engagement index and the FAA’s workload studies have been checked for decades. Until that happens, the honest way to read “94 percent accuracy” is as a statement about how well a model agrees with its own labels, dressed up as a statement about the inside of your head.

Credit: Vidar Nordli-Mathisen on Unsplash