AI safety discourse has become unverifiable, and that’s a bigger problem than whatever the models are actually doing.
I review AI tools for a living. That means I spend most of my week separating marketing claims from measurable behavior, and I’ve gotten reasonably good at it. But the safety conversation in 2026 has crossed into territory where I genuinely cannot tell you which alarming story is real, which is exaggerated, and which is a screenshot someone generated in four seconds. TechCrunch put it plainly this week: two viral AI safety conversations demonstrated just how hard it has become to discern AI fact from fiction. One involved Andrew Yang. That’s the part I can confirm. Everything downstream of it turned into the usual reply-guy sediment.
What we actually know
Strip away the noise and the verified picture is short, which is exactly why it gets padded out with speculation:
- Safety discussions have intensified in 2026, driven by unexpected model behaviors and concerns about potential risks.
- OpenAI, Anthropic, and Google have been in talks on AI safety for weeks, per OpenAI global policy chief Chris Lehane, who spoke to reporters in mid-September.
- Regulatory discussions are underway.
- Mainstream outlets including ABC News and CBS News ran segments in September on concerning incidents of AI behavior, featuring outside experts.
That’s the whole verified pile. Four bullet points. Notice what’s missing: specifics. Which behaviors? In which models? Measured how? Reproduced by whom? The public record right now consists of major labs confirming they’re talking, news networks confirming that experts are concerned, and a viral discourse cycle that sits on top of both like foam on a beer.
The verification gap is the story
I’m not arguing the concerns are fake. Unexpected model behavior is real, documented in plenty of technical work, and worth taking seriously. My complaint is narrower and, I think, more useful: the infrastructure for checking safety claims has not kept pace with the volume of safety claims.
When I test a coding agent, I can run it. I can give it the same repo ten times and count failures. When someone posts that a frontier model did something unsettling, I usually can’t reproduce it. I don’t have the model version, the system prompt, the temperature, the conversation history, or often even confirmation that the screenshot came from the product it claims to. The claim goes viral. The correction, if one arrives, gets 3% of the reach.
Look at the engagement numbers on those news segments. The ABC News piece: 802 views, 15 likes, on a channel with 19.8 million subscribers. The CBS segment: 439 views, 21 likes, 7.09 million subscribers. Careful, sourced reporting from outlets with enormous audiences is getting effectively zero attention, while unsourced posts about model behavior circulate for days. That asymmetry shapes what the public believes about AI risk far more than any lab’s safety framework does.
Why the lab talks matter more than the discourse
The most consequential fact in this batch is the least dramatic one. OpenAI, Anthropic, and Google have been in conversation about safety for weeks. Three competitors who normally communicate through launch blogs are apparently comparing notes in private.
I read that two ways, and both are worth holding at once. The generous read: this is what responsible coordination looks like, and it’s better than each lab defining safety alone. The skeptical read: coordinating on safety standards ahead of regulatory action is also how incumbents shape the rules they’ll be graded against. Lehane is a policy chief, not a researcher. When policy chiefs talk to reporters about industry collaboration while regulatory discussions are underway, that’s positioning as much as it is progress.
Neither read requires cynicism about the people involved. It just requires noticing who’s in the room.
What I’d want before I believe the next viral claim
My standard, which I’d suggest you borrow: treat safety incidents like you’d treat a benchmark result from a vendor. Ask for the version number. Ask whether anyone independent reproduced it. Ask what the base rate is, because a behavior that appears once in 50,000 turns and a behavior that appears reliably are different stories told with identical screenshots.
And ask who benefits from you being scared. Sometimes the answer is nobody, and the concern is just genuine. Sometimes the answer is a security vendor, a competitor, or an engagement-farming account.
The honest position on AI safety in 2026 is discomfort. The labs are worried enough to talk to each other. Regulators are moving. Experts on national broadcasts are raising flags. And the public conversation carrying all of that is so contaminated that a professional skeptic like me is reduced to saying, repeatedly, “I don’t know, and neither do you.”
That’s not a satisfying verdict. It’s just the accurate one. Anyone offering you more certainty than that is selling something.
🕒 Published: