INMO Air3 Recall Shows the Gap Between Wearable Health Data and Wearable Safety

The US Consumer Product Safety Commission ordered the recall of the Air3 smart glasses, according to a notice relayed by the German tech outlet t3n on 27 September. Health Canada joined the action the same day. The left temple, the part that houses the electronics, can overheat during extended use. Regulators say that creates a risk of serious burns.

What the recall covers

The affected units shipped between December 2025 and August 2026. Regulators say 1,643 were sold in the United States through various distribution channels, and 120 in Canada. The devices are black and carry the labels "AIR3" and "INMO" on the inside of the right temple. The manufacturer is INMO International Technology Limited, based in Shenzhen, China.

INMO has logged ten official reports of overheating during use in the US, according to the recall documents. Neither US nor Canadian authorities have reported any actual injuries. The company is not pulling the hardware or swapping the battery. Owners are told to update the operating system to firmware version 3.16 before putting the glasses back on. The device runs an Android-based OS and displays content on a micro-OLED panel, so the fix appears to be a software throttle on the processor, t3n notes.

Support runs through a dedicated website, a hotline and email. Health Canada also warns that Canadian consumer protection law bans reselling or even giving away the recalled models. The Air3's suggested retail price runs between $900 and $1,300 depending on configuration.

"First hints of the hardware's temperature problems appeared in the trade press and among hardware testers months before regulators intervened," t3n writes.

That timeline matters more than the recall itself. According to Futurism, thermal problems showed up in early tests. The US hardware YouTuber "Non-Tech Tech Guy" said in a February 2026 video that the device became too hot to wear after just a few minutes, reaching temperatures that hurt at the ear. His video claims the manufacturer's first response was to ship a sleeve for the temple. The sleeve blocked direct skin contact with the plastic but did not remove the heat source inside.

So the product spent roughly seven months on shelves after a public tester had already flagged the problem. That is the shape of the story: a wearable that carries a battery of 660 milliampere-hours against the side of a user's head, a known heat complaint, and a fix that arrived only when regulators stepped in.

A different kind of wearable problem in healthcare

The recall sits at the edge of the digital health sector, and it lands in the same week as a much bigger argument about what happens when software gets between patients and medical decisions. On 26 September, TechCrunch reported that hospitals' use of artificial intelligence tools when submitting insurance claims led to an additional $942 million in healthcare spending over a two-year period. The figure comes from an analysis by the Blue Cross Blue Shield Association.

The BCBSA analysis found what it called a sharp increase in patients being documented as having complex conditions. It argued there is a clear disconnect between medical coding and treatment, because there is no evidence of a corresponding change in care delivered. The New York Times, cited by TechCrunch, described the analysis as the latest sign that AI is contributing to rising healthcare costs. It noted that fights between hospitals and insurers over payments are nothing new, but that AI on both sides appears to be making them worse.

Two people quoted in that report frame the dispute differently. Dr. Shiv Rao, founder of the AI startup Abridge, acknowledged that AI use could lead to "a horrible dystopic future nobody wants to live in," with bots fighting bots and agents fighting agents, but said it might also reduce tensions and cut costs. Luke Chalker, a senior vice president at BCBSA, resisted the word battle: "It's not a war. It's a completely one-sided blood bath," he said, with insurers on the losing side.

That is a financial argument about coding and reimbursement. It is not, strictly speaking, a clinical safety story. But it shares a root with the Air3 recall: software is being asked to manage a problem that the hardware or the workflow created, and the people affected find out later.

What the evidence says about AI as a crutch

A study published on 26 September by The Decoder offers a more uncomfortable angle. Researchers ran five experiments with 3,132 participants to test whether access to a language model changes how people handle uncertainty. Just having AI advice available nearly wiped out people's willingness to say "I don't know."

The design was deliberately harsh. The researchers picked questions where the model they used, Step 3.5 Flash, was almost always wrong: fine visual details from movies, such as the color of a team uniform in Bend It Like Beckham. Those details rarely appear in online text, which makes them prime targets for hallucinations. GPT-5.5, Claude 4.6 Sonnet and Gemini 3.5 Flash got most of the other questions right but sometimes failed on the harder ones.

In the first two studies, participants could choose whether to ask an AI for advice. In the control group without AI access, they withheld judgment on 36 and 44 percent of questions. With AI access, those numbers fell to 6 and 3 percent. In Study 2, confidence with AI access hit 75.9 points on a scale of 100, against 29.6 points without AI, while the share of correct answers dropped from 27.6 to 10.0 percent. Across all studies, participants without incentives who had AI access got 9.2 percent of questions right, versus 27.5 percent without AI.

Studies 2 through 4 added money. Participants earned 10 cents for each correct answer, lost 10 cents for each wrong one, and got nothing for "I don't know." The researchers had pre-registered the hypothesis that incentives would increase willingness to abstain but that AI availability would weaken that effect. None of the three studies showed a statistically significant interaction. Incentives led participants to seek AI advice less often, 4.53 versus 5.27 requests out of six in Study 3, and to answer correctly more often when AI was available. But the effect was modest.

In Study 4, the AI answer appeared automatically, without the participant asking. The researchers say this mirrors search engines that display AI summaries and writing assistants that offer unsolicited suggestions. Judgment suspension dropped from 35 percent without AI to 1 percent with automatically displayed answers. With incentives, it fell from about 39 to 7 percent.

The authors frame the result around what they call "Epistemia": the tendency to accept AI answers because they sound convincing rather than checking them. A language model always has to produce an answer, they write. It never pauses, even when it does not know something. When users hand off their judgment to such a system, they may adopt its lack of restraint. The paper also cites a study from Swiss Business School with 666 participants that found a strong negative correlation between AI use and critical thinking.

Read together with the BCBSA claim data, the pattern is not flattering. Software that always answers, deployed in billing and now in clinical workflows, produces confident output. Whether that output is right is a separate question, and one that is often answered only after the invoice or the injury.

Background: health tech's centre of gravity is shifting

These debates are unfolding as the geography of health technology changes. In a Rest of World interview published on 24 September, physician and consultant Ruby Wang described how China's digital health infrastructure works: patients book appointments through WeChat or Alipay, consult doctors over telemedicine platforms, or chat with an AI avatar of a doctor about everything from insomnia to diabetes. Wang argued the West treats technology as an add-on to a health system, while in China the digital rails are enabled enough that society accepts them as normal. She also said China's digital health is not directly exportable, and that the West would never accept an AI doctor avatar.

On drug development, Wang said more clinical trials are being conducted in China than in the US, and that the country now accounts for half of global drug licensing deals, potentially as much as 60 percent this year. She noted caveats: China is strongest in oncology and less mature in immunology, and its biotech companies are weak at commercialising their own products, which is why out-licensing to partners such as Pfizer is common. She also criticised security-focused framing of Chinese medicine, pointing to the Biosecure Act, the COINS Act and the BINSA bill as potential barriers to drugs reaching patients elsewhere.

For anyone watching wearables specifically, the oldest item in this dossier is a reminder that the underlying radio technology is not new. IEEE Spectrum reported on 21 September that Rhizomatica, a Philadelphia nonprofit, has open-sourced a digital shortwave radio set called HERMES, which operates in the 3 to 30 megahertz band and uses the ionosphere instead of satellites or relay towers. Founder Peter Bloom said the radios put out about 20 watts and can reliably do 400 to 600 kilometre links, enough for telemedicine manifests and coordinates in places where terrestrial infrastructure is not simple to build.

That is a useful contrast to the Air3. One project uses cheap, well-understood radio to move small amounts of data to places with no other option. The other puts a 660 milliampere-hour battery against a user's temple and asks a firmware update to keep it cool.

Regulators have not reported injuries from the Air3, and INMO has a support channel open. Owners of units bought between December 2025 and August 2026 should check the label inside the right temple and update before wearing the glasses again. The wider lesson is less tidy: in health technology, the failure mode is rarely the algorithm alone. It is the gap between what a device or a model claims to do and what the people relying on it can actually verify.