How AI Toothbrushes Work: Sensors, Models, and Dyson's Camera

Dyson shipped the first toothbrush with a built-in camera and an onboard AI model on September 1, 2026. The $499 CameraJet fires a targeted jet of mouthwash at gaps between teeth that its camera spots 28 times a second — and it does the spotting locally, without sending a single frame to the cloud. That launch is a good excuse to answer a question that's been quietly true of "smart" toothbrushes for almost a decade: what is the AI in an AI toothbrush actually doing, and how does a model that classifies brushing motion end up running on a device this small? This piece walks through the two real engineering approaches on the market — motion-sensor classifiers and, now, onboard computer vision — and what the pattern means if you're building your own edge AI product.

TL;DR

Two families of "AI toothbrush," not one

"AI toothbrush" gets used as one marketing phrase, but it covers two genuinely different sensing and modeling approaches, built for different jobs:

1. Motion classifiers. An accelerometer and gyroscope inside the handle stream orientation and rotation data to a small model that infers where the brush is in your mouth and how you're moving it. This is what Oral-B iO, Colgate Hum, Philips Sonicare's SenseIQ line, and Oclean's X Pro Elite all do, in slightly different implementations.
2. On-device computer vision. A camera physically sees the teeth and gum line, and a vision model classifies what's in frame in real time. As of September 2026, Dyson's CameraJet is the only consumer toothbrush doing this.

Both are examples of on-device inference — running a trained model directly on local hardware instead of shipping raw data to a server for every decision — for the same underlying reason: brushing feedback has to be fast enough to act on mid-stroke, and nobody wants raw motion or video data of their mouth leaving the device by default.

How motion-sensor AI toothbrushes actually work

Oral-B's iO and Genius lines are the most thoroughly documented example, in part because Procter & Gamble has filed a string of patents on the mechanism (US 9,888,763; US 10,702,206; US 11,051,919). The handle carries a 3-axis accelerometer — sometimes paired with a gyroscope and, in later patents, a magnetometer — sampling orientation continuously while the brush runs. The hard engineering problem isn't sensing motion; it's separating meaningful motion from the noise of a motor spinning at roughly 30,000 vibrations per minute. Because that vibration frequency is known and constant, firmware can filter it out, leaving the gravity vector and rotation signal that actually describes how the brush is oriented in your mouth.

That cleaned signal feeds a model — P&G describes it as trained using machine learning on labeled brushing sessions collected from thousands of real users — that classifies the motion pattern into a mouth position: which quadrant, which surface, whether the bristles are angled toward the gumline. Oral-B's own materials describe validating this against video-tracked ground truth from thousands of participants, then continuously improving the model using anonymized session data volunteered by app users.

Colgate's Hum takes a lighter-weight version of the same idea: a Bluetooth-connected sonic brush (vibrating at roughly 31,000 strokes per minute) streams timing and motion data to a phone app, which tracks brushing frequency, duration, and estimated mouth coverage rather than fine-grained per-tooth position. Philips' Sonicare 9900 Prestige goes further on the sensor side — its SenseIQ system samples pressure, motion, and coverage up to 100 times per second and auto-adjusts vibration intensity in real time if you push too hard, with an AI-powered companion app layering coaching on top.

Oclean's X Pro Elite is the interesting outlier in where the model runs: its 6-axis gyroscope feeds a classifier that runs locally enough to display real-time angle and coverage feedback directly on the brush's own touchscreen, rather than depending on a live Bluetooth link to a phone the way Oral-B and Colgate's flagship coaching features do.

Dyson CameraJet: what changes when the sensor is a camera

Dyson's CameraJet — technically the Dyson Dental System — is a different engineering problem entirely, and it's the reason this is worth writing about the same week it shipped. A 100,000-pixel macro camera sits under the brush head and captures 28 images per second while you brush. A machine-learning model, which Dyson says was trained on roughly 470,000 dental images, classifies each frame for interdental gaps — the spaces between teeth that a bristle can't reliably reach — and when it finds one, the device fires a targeted jet of mouthwash at it within 100 milliseconds of detection.

That 100ms budget is the whole reason the model has to run on-device. A round trip to a cloud inference endpoint — network latency, queueing, a response back over Bluetooth — cannot physically land inside 100 milliseconds under normal consumer network conditions. Dyson also states explicitly that camera images aren't stored or transmitted off the device, which is as much a privacy decision as a latency one: streaming continuous close-up video of the inside of someone's mouth to a server is a materially bigger data-handling liability than streaming accelerometer readings.

This is the same tradeoff curve covered in explainx.ai's Coral Edge AI guide and in Matic's Jetson Orin Nano-powered home robot: once a device needs to act on what it sees in real time, at these latency budgets, running the model locally stops being an optimization and becomes the only architecture that works at all.

What the plaque-detection research actually shows

Camera-based dental AI isn't unique to Dyson — it's an active academic field. A narrative review published in early 2026 identified 74 studies applying deep learning, including convolutional networks and vision transformers, to intraoral photographic images for tasks like caries detection, gingivitis scoring, and plaque-index assessment. Individual studies — including work applying AI to biofilm assessment in pediatric patients — have reported model accuracy on narrow plaque-scoring tasks that meets or exceeds agreement between human dentist raters on that same task.

The honest caveat: "beats a dentist at one scoring task" is not the same claim as "diagnoses oral health." These models are trained for a single narrow classification job on a specific dataset, and none of the consumer products discussed here — Dyson included — claim to replace a dental exam. What the research does establish is that the underlying technique (small vision models classifying dental images) is well past proof-of-concept and genuinely competitive on the specific tasks it's trained for.

What this means if you're building edge-AI products yourself

The toothbrush category is a useful, low-stakes case study in tradeoffs that show up in any real-time embedded AI product:

• Latency budget decides the architecture, not preference. Dyson's 100ms jet-trigger requirement rules out cloud inference outright — this is the same constraint that pushes any real-time robotics, camera, or wearable product toward on-device inference once the feedback loop has to be fast enough to act on.
• Model size follows sensor type. Motion classifiers running on accelerometer/gyroscope streams are small enough to run on cheap, low-power microcontrollers — the same TinyML pattern behind on-device speech models like Superwhisper's S1-mini. A real-time vision model, like the one behind CameraJet or general-purpose detectors such as RF-DETR, needs meaningfully more compute and a purpose-built accelerator, not a microcontroller alone.
• Training data is the actual moat. Oral-B's position-detection model and Dyson's gap-detection model both cite training on labeled data from thousands of real sessions or hundreds of thousands of real images — collected specifically for the product, not scraped. That data-collection pipeline, not the model architecture, is usually the harder and more defensible part to replicate.
• Local processing is a privacy decision as much as a technical one. Keeping raw sensor or camera data on-device by default — as every product here does — sidesteps the liability of hosting continuous biometric-adjacent data (motion in your mouth, images of your teeth) on a server at all.

Where this heads next

Expect the camera-based approach to spread the same way motion sensing did: Dyson's launch makes the case that a $499 price point supports an onboard vision model with real-time inference, and competitors typically follow a proven bill of materials within a product cycle or two. The more interesting open question is whether any manufacturer pairs plaque or gum-health classification (the kind of thing the academic research above is already demonstrating) with the real-time coaching layer that Oral-B and Philips have spent years refining — combining "here's what's wrong" with "here's what to do about it right now," in the same 100-millisecond feedback loop Dyson just proved is achievable on-device.

Bottom line

There's no single "AI" inside an AI toothbrush — there's a motion classifier reading an accelerometer, or, as of September 2026, a real-time vision model reading a camera, each trained on data collected specifically for that task and run locally because the feedback has to be fast and the data shouldn't leave the device by default. Dyson's CameraJet is the newest and most technically ambitious version of that pattern, not a different category of product from Oral-B's decade-old position detection — just a more expensive sensor feeding a bigger on-device model at a tighter latency budget.

Related reading

• Coral Edge AI: Google's edge computing platform
• Matic's home robot runs entirely on a Jetson Orin Nano
• S1-mini: Superwhisper's 0.6B on-device transcript cleaner
• RF-DETR: Roboflow's real-time detection transformer, explained
• FDA clears the Libre Duo glucose and ketone sensor — another consumer wearable built around continuous on-device sensing
• Vibefarming: running a sweet potato season with sensors and Claude
• AI vs machine learning vs deep learning guide
• Oral-B position detection patent (US 9,888,763)

Product specifications, pricing, and technical claims reflect manufacturer statements and public reporting as of September 5, 2026. Dyson's CameraJet began shipping September 8, 2026; specifics of its on-device model are as publicly disclosed by Dyson, not independently benchmarked by explainx.ai.