AI does not have an adoption problem. It has a proof problem

Five signals are being mistaken for proof

Today's stories arrive from different institutions: an international expert exercise, a medical evidence review, an account of public backlash, a police pilot, and a reported takeover approach. Each produces a signal that can sound like validation. Experts express concern. A regulator authorizes a device. A department deploys a tool. A company attracts a powerful buyer. A technology becomes difficult to avoid.

None of those signals answers the question that legitimacy requires: did the system create a durable benefit for the people absorbing its consequences? Expert judgment can prioritize hazards. Clearance can permit market entry. A pilot can test feasibility. Capital can price strategic value. Adoption can reveal distribution. They are useful evidence, but they are not interchangeable with safety, effectiveness, consent, or public value.

Catastrophic risk deserves mitigation, not probability theater

A Delphi study summarized by MIT Sloan asked 272 specialists from 37 countries to assess 24 AI risk categories over a five-year horizon. Under a business-as-usual scenario, participants placed 18 categories above a 10 percent probability of catastrophic harm as defined by the study. With pragmatic mitigations, five categories remained above that threshold.

The categories overlap, and the researchers explicitly warn against adding those probabilities into one apocalyptic headline. The responsible conclusion is still severe: a mature industry would not treat repeated double-digit assessments of catastrophic harm as background noise. The value of the study lies in prioritizing mitigation and clarifying responsibility, not pretending that expert elicitation is a forecast with actuarial precision.

Medicine exposes the cost of substituting clearance for outcomes

A PLOS Digital Health evidence census examined 1,357 AI and machine-learning medical devices authorized by the FDA through December 5, 2025. Only 34 were linked to registered prospective trials, 12 had posted results, 12 had peer-reviewed publications, and three evaluated patient-centered outcomes such as mortality, morbidity, or readmission.

This does not establish that the other devices are ineffective. It establishes that the available evidence rarely answers whether they improve what happens to patients. Regulatory authorization, technical accuracy, and workflow efficiency can all matter, but a health system should not let them become substitutes for clinical benefit, subgroup performance, post-market surveillance, and transparent failure reporting.

Trust is a verdict on visible trade-offs

TechCrunch argues that ubiquity has not delivered acceptance because many people experience AI as an imposed feature alongside job anxiety, intellectual-property conflict, and data-center costs. Pew's 2026 survey provides a direct measure of that skepticism: 63 percent of Americans said AI was advancing too quickly, 71 percent expected it to make personal information less secure, and about six in ten lacked confidence that companies would develop and use it responsibly.

The industry can call that a narrative problem, but the public may be grading the deal accurately. People are being asked to accept disruption, infrastructure, surveillance risk, and uncertain labor consequences before the promised gains become measurable in their own lives. Trust will not return because adoption statistics rise. It will return when benefits are observable, harms are contestable, and claims can be audited by people outside the institution making them.

Public-safety pilots need rights before scale

Kennesaw State University and Technovative AI announced that Moultrie Police will pilot CaseFinder on department-owned hardware. The tool analyzes incident reports and 911 transcripts to identify possible behavioral-health crises for co-responder follow-up. That goal addresses a real capacity problem, and the pilot may produce useful evidence.

It also operates where sensitive health inference meets police data. A responsible pilot must measure more than cases surfaced. It should publish false-positive and false-negative rates, demographic performance, data-retention rules, who may inspect the output, whether a person can contest the classification, whether follow-up remained voluntary, and whether the intervention improved access to care without increasing coercion. The pilot is the question. Those outcomes must be the answer.

Capital validates demand, not public benefit

Bloomberg reports that SpaceX approached AI coding company Cognition about a possible acquisition, that Cognition did not engage with the approach, and that discussions about collaboration may continue. The report relies on unnamed sources because the talks were not public. There is no completed acquisition to evaluate.

Even if a partnership follows, strategic interest proves only that computing capacity and coding agents may be valuable to the parties. It cannot establish safety, labor benefit, competitive fairness, or social value. Capital is excellent at revealing what powerful institutions want. It is not a substitute for evidence about what everyone else receives.

Replace proxies with proof

AI can create enormous value, but the burden of proof should rise with the stakes and scale of deployment. The institutions asking for trust should stop presenting permission, prevalence, investment, and expert attention as if they were outcomes.

The better standard is blunt: measure the promised benefit, expose the failure rate, give affected people a path to challenge the system, and let independent scrutiny reach the evidence before a proxy hardens into policy.

• Publish outcome measures before celebrating adoption or market authorization.
• Separate expert risk judgments from forecasts and disclose uncertainty and overlap.
• Require independent validation across populations before high-stakes systems scale.
• Build notice, correction, appeal, and human review into public-sector deployments.
• Treat investment and acquisition interest as market evidence, not proof of public benefit.