Part 2 of a series on AI safety, speed and the products entering everyday life

At CES 2026, one of the products that stayed with me was a toilet. Pondo's display said its AI could turn observations such as color, shape, volume and frequency into digestive and hydration scores. At the same show, my wife and I saw health mirrors promising to turn ordinary reflections into useful signals. One mirror prompted her to see a cardiologist; weeks later, a doctor confirmed a concern. That is our experience, not a clinical validation of the mirror or a reason to replace medical care with a device.

It is also a good place to begin a conversation about superintelligence. No one has established that artificial superintelligence exists. The term describes a hypothetical system that would exceed human ability across a broad range of intellectual tasks. CES 2027 will not be evidence that such a system has arrived. What we can already see is a race to put today's more capable AI into products that observe our bodies, run our homes, help us work and move us through the world.

In Part 1 of this series, I asked whether AI needs a brake pedal. The question at CES is more immediate: as companies press ahead, who checks whether the products they sell deserve our trust?

Why the show matters now

CES began in 1967 as a consumer electronics show. Sixty years later, its scope includes computing, vehicles, robotics and digital health. The significance of CES is not that every demonstration becomes a hit. It is that the show gathers the people who can make an idea widely available: manufacturers, retailers, investors, journalists and officials.

The scale is real. According to the independently audited CES 2026 attendance summary, the show had 148,392 verified attendees. Some 55,841 came from outside the United States, representing 141 countries, regions and territories. The same report counted 189 government guests through CES's Leaders in Technology program. These are 2026 figures, not a forecast for 2027.

That global reach creates an unusual opportunity. A company can unveil an AI health device in front of a retailer who might stock it, a clinician who might question its evidence, a journalist who might inspect its claims and an official who might ask whether existing rules apply. Those conversations can affect what gets sold and how it is described. They do not, by themselves, create a global safety agreement.

There is another limit that matters: CES is a trade-only event, closed to the general public. Consumers do not get a vote on the show floor. Their interests have to be represented through independent coverage, patient and consumer advocates, researchers, clinicians, buyers who demand evidence and companies willing to answer hard questions. A bigger crowd of executives is not the same thing as public accountability.

What changes when AI leaves the screen

An AI assistant that makes a bad suggestion can waste an afternoon. A device that misreads a health signal may delay care or create needless alarm. A home robot can damage property. A vehicle system can make a consequential mistake in traffic. As AI gains access to sensors, software tools and physical controls, the setting and the permissions matter as much as the model's intelligence.

The Pondo display I photographed at CES 2026 illustrates the promise and the gap. Its proposed digestive and hydration scores could make a private, easily ignored health signal easier to track. But a booth sign cannot establish whether a score is accurate for different users, what a false alert means, or who receives the underlying data. The same questions apply to the mirrors my wife and I saw. Her doctor's later finding makes the episode personally meaningful; it does not tell us how often such devices are right or wrong. Pondo's own site describes the product, but its claims need independent assessment.

This is the sort of reporting CES should provoke. For every AI health product, ask: What does it measure directly, and what does it infer? Was it tested against an appropriate clinical reference? On whom? How often does it miss a problem or raise a false alarm? Is the claim about general wellness or medical diagnosis? Can users delete sensitive data, and does a score still work if they refuse cloud processing?

Other categories demand their own questions. For an AI PC, which tasks run on the device and which send data to a server? For a home robot, what happens when it fails to recognize a child, pet or obstacle? For a connected vehicle, what actions can the system take without human confirmation, and how is that limit enforced? A useful demonstration answers a specific question under repeatable conditions. A cinematic demo merely raises one.

CES 2027, scheduled for January 6–9 in Las Vegas, already has relevant stages. The announced CES Foundry is devoted to AI and quantum innovation, while the new Health Hub will bring together health technology, clinicians, investors and media. As of October 13, the announced keynote roster includes leaders from Lenovo, Waymo and Stellantis, putting computing and mobility in view. Their presence does not confirm any particular product launch or safety claim. Those details deserve scrutiny when they are announced.

Where policy meets the product pitch

CES already hosts policy debates alongside product launches. At CES 2026, the announced government speakers included presidential science adviser Michael Kratsios, Federal Trade Commission Chair Andrew Ferguson and Federal Communications Commission Chair Brendan Carr. Kratsios was scheduled to discuss America's AI future at CES Foundry. The presence of officials gives manufacturers a chance to hear policy concerns and gives officials a view of the technology being proposed for sale.

The Consumer Technology Association, which owns and produces CES, can help turn broad concerns into practical standards and questions. In August 2026, for example, CTA published a standard for monitoring the accuracy of predictive health AI solutions after they reach the market. That focus on performance after launch is valuable: a health model may change, its users may differ from a study population, and problems may only become visible in routine use.

CTA also represents the technology industry. Its AI policy position favors a national framework based on risk while warning against regulation it sees as excessive. Readers should understand both roles. CTA can convene and set standards; it is not an independent regulator, and a company appearing at CES has not earned a safety seal. The stronger test is whether a product maker publishes meaningful evidence, allows independent scrutiny and remains accountable after the applause ends.

The problem a Las Vegas gathering cannot solve

What if the United States slows development and another country does not? What about models already available to download? These are serious objections to any simple “pause the race” slogan.

The 2026 International AI Safety Report explains the distinction between open-weight and open-source AI. Open-weight means the trained model's numerical parameters can be downloaded and run elsewhere; the training data and code often remain private. Such releases can widen access for researchers and smaller developers. They also make safeguards easier to remove and use harder to monitor. Once the weights have spread, they cannot be recalled like a defective appliance.

Even Anthropic CEO Dario Amodei, a prominent advocate of stronger safeguards, has said that he does not support a blanket ban on open-weight models. He argues for testing sufficiently capable models, whether open or closed, before release, while acknowledging that effective testing would require international participation. That is a policy proposal, not an existing worldwide rule.

No U.S. law or CES panel can compel every government or developer to adopt the same restrictions. Still, imperfect global cooperation is not the same as having no options. Governments can share incident information and evaluate dangerous capabilities. Buyers can demand security, logging and update commitments in contracts. Developers can limit an agent's permissions and require approval before it reaches money, private records or physical controls. Standards can give organizations a common language for risk management: NIST's AI Risk Management Framework is voluntary, while ISO/IEC 42001 sets requirements for an organizational AI management system. Neither certifies that a particular CES demo is safe.

The wider diplomatic work belongs in wider institutions. The United Nations' Global Dialogue on AI Governance brought representatives from 163 countries to its first meeting in July 2026. Dialogue does not guarantee agreement or enforcement, but it reaches governments that a U.S. industry trade show cannot bind. CES is best used as a public-facing test of claims and as one place to compare practical approaches; international negotiations, national regulators, independent labs and standards bodies have different jobs.

Are the warnings about humanity's end evidence that the end is near?

Some Anthropic researchers and former employees have made stark public forecasts. In September, Axios reported that alignment researcher Evan Hubinger personally put the chance of AI killing all humans within the next decade above 10%. Former researcher Jacob Coxon and others voiced grave concerns. These are their judgments about possible futures, not an Anthropic finding that catastrophe is imminent and not a measured probability derived from a repeatable experiment.

The independent International AI Safety Report gives a more careful picture. It says experts disagree widely about the likelihood of a future loss of human control. Current systems show early signs of some relevant abilities but do not have the sustained capabilities needed for the severe scenarios the report examines. The report also says that present decisions about access, permissions, testing and oversight could affect future risk. We can take the warnings seriously without presenting a deadline for the end of the world as established fact.

For consumers, the nearer question is already difficult enough: How much independence should we give a system that can be wrong, opaque and connected to our most intimate data? That question applies whether or not superintelligence ever arrives.

The test for CES 2027

I will be watching for five things. First, a clear statement of what each AI feature actually does. Second, evidence that its claimed benefit holds up outside a staged demonstration. Third, limits on what an agent or device can access and do without permission. Fourth, a plain account of data storage, security and software support. Fifth, a way for users and independent investigators to report problems and see how a company responds.

The booth that earns attention may still be the one with the oddest object. Last year, for me, it was a toilet. But the measure of progress will be less amusing: whether its maker can explain the score, show the evidence and remain responsible when the product enters someone's home. That is why CES matters in this age of superintelligence talk. It puts the future on display while giving us a chance to ask what, exactly, we are being asked to trust.

Quick answers

Has superintelligence been achieved? No. It remains a hypothetical, disputed future capability. CES showcases present and proposed products; it does not establish that superintelligence exists.

Can consumers attend CES 2027? CES is a trade-only event and is not open to the general public. Consumers can follow independent reporting and ask manufacturers for evidence behind product claims.

Could CES create worldwide AI rules? No. CES can convene companies and officials and expose products to scrutiny. Binding rules depend on governments and other institutions, with international cooperation where possible.

Can released open-weight models be recalled? No. Once model weights have been downloaded and copied, the original developer cannot reliably withdraw every copy. That makes testing before release and safeguards around use especially important.

Does a CES health demo prove medical accuracy? No. A demo shows a product's intended behavior in a selected setting. Medical claims require suitable clinical evidence and, where applicable, regulatory review.