Sixty-six likes. That’s how much engagement ABC News got on a video titled “AI safety concerns grow after models show unexpected behavior,” posted September 17, 2026, on a channel with 19.8 million subscribers. The video pulled 13,655 views. CBS News published a similar segment the same day, “Experts weigh in on concerning incidents of AI behavior,” and landed 9,709 views and 83 likes from a 7.09-million-subscriber audience.
I keep coming back to those numbers because they tell you something the headlines don’t. Two of the largest news organizations in the United States covered what may be the most consequential technology story of the year, and the response was roughly the size of a high school assembly. Meanwhile, the same week, TechCrunch reported that two AI safety conversations went viral — viral in the real sense, the kind that fills your feed for days.
What actually went viral
Per TechCrunch, the two viral conversations “demonstrate just how hard it is to discern AI fact from fiction.” One involved Andrew Yang, the former presidential candidate. I’m not going to characterize the specifics beyond that, because the reporting I have access to cuts off mid-sentence, and inventing the rest would make me part of the problem I’m describing.
That’s the whole shape of this story, honestly. The most-shared AI safety content of the week was not the network news segment with named experts. It was the stuff that was hard to verify.
Why this is a specific kind of hard for regular people
If you follow AI casually — you use a chatbot at work, you’ve heard the word “agent” a lot lately — you’re in a rough spot. You’re being asked to form an opinion about systems you can’t inspect, based on incidents you can’t independently check, described by people whose incentives you can’t fully see.
Traditional media literacy advice doesn’t help much here. “Check the source” assumes the source can show you something. But a claim like “this model did something unexpected” is genuinely hard to verify from the outside. The model isn’t sitting on a shelf where you can go look at it. Its behavior may not repeat. The company that runs it controls the logs.
So the stories that spread are the ones that feel true. And feeling true and being true are different properties that happen to travel together often enough to fool all of us.
The part that’s easier to pin down
Here’s a fact that’s checkable: OpenAI, Anthropic, and Google have been in talks about AI safety for weeks. Chris Lehane, OpenAI’s global policy chief, spoke to reporters about it, and TechCrunch’s Rebecca Bellan reported the story on September 15, 2026.
Three companies that compete hard against each other are sitting in rooms together about safety. That’s the signal I’d pay attention to over any individual viral incident. Competitors coordinate when the shared downside of not coordinating exceeds the private benefit of going it alone. You don’t need to know the contents of those meetings to read something from the fact that they’re happening.
Separately, experts continue to warn about the pace of AI advancement and the risks that come with it. That’s the broad consensus position, not a fringe one.
What I’d actually do with this information
A few things that have kept me sane covering this beat:
- Separate “a model did something weird” from “a model did something dangerous.” These get collapsed constantly. Unexpected behavior in a system with hundreds of millions of users is statistically guaranteed. Whether a given instance matters depends on details that rarely survive the trip into a viral post.
- Notice who is claiming what, and what they get from the claim. This cuts in every direction. Doom sells. So does dismissal. So does “our model is so capable it’s almost scary.”
- Weight institutional behavior over anecdotes. Coordinated talks between rivals, hiring patterns, policy filings — these are slower and duller than a screenshot, and they’re much harder to fake.
- Let “I don’t know yet” be a real answer. On a story where the verifiable details are still thin, holding a loose opinion is the accurate response, not a failure of nerve.
The uncomfortable summary
The gap between those view counts and those viral threads is the actual story. The careful version of AI safety reporting — named experts, network newsrooms, specific incidents — reaches a modest audience. The ambiguous version reaches everyone.
That’s not a knock on anyone who shared the viral stuff. It’s an honest description of what it’s like to try to understand a fast-moving technology from the outside, using tools built for a slower world. The best thing I can offer you isn’t certainty. It’s permission to stay curious and unconvinced at the same time.
🕒 Published: