A soft AI executive order, Bernie Sanders' 50% proposal, FBI monitoring of data center critics, and more. What Warning Shots #45 reveals about the widening AI governance gap.
There is a distance between how quickly AI is being built and how slowly anyone is being held accountable for it. Most weeks that distance stays abstract. This week it did not. On Warning Shots #45, hosts John Sherman, Liron Shapira, and Michael walked through five developments that, taken together, sketch the same picture: capability and deployment accelerating, while public oversight arrives late and unsure of what it is even looking at.
Here is what they covered, and why each story is a place where the public still has leverage.
The administration signed an executive order shortening the window for government model review from 90 days to 30, and the review remains voluntary. One official reportedly described the change as friendly to innovation.
On the show, John Sherman argued that there is no hard standard behind it. Companies submit a model, the government examines it for security concerns, and that is roughly the extent of the requirement. Michael raised a technical objection that is difficult to dismiss. According to Michael, the most serious failures in these systems tend to surface months or years after release, sometimes by accident. A 30-day review is unlikely to catch much.
Liron Shapira offered a more hopeful reading. Even a weak rule puts the government's hand on the process for the first time, and he suggested that voluntary measures often harden into binding ones over time. That may prove true. It is also a reason to push for the version of this policy that has real standards now, rather than waiting for a voluntary gesture to mature on its own.
Senator Bernie Sanders reportedly met with OpenAI's Sam Altman to argue for a 50% public stake in the major AI companies. His stated reasoning, as discussed on the episode, is that the value of these systems was drawn from the public's data and returned to almost no one.
John Sherman said he found the underlying claim fair, noting that these models were trained on vast amounts of human writing, code, and conversation without compensation. Liron Shapira was more skeptical of the timing, describing the proposal as premature given current employment levels and suggesting that redistribution should be tied to concrete milestones rather than a fixed percentage chosen today.
The most important point came from Michael, who argued that ownership only matters if humans retain meaningful control of what the companies are building. As he framed it, a public stake in a system we cannot steer does not solve the underlying problem. That control question is the one that receives the least attention in policy debates, and it is the one that matters most.
The most uncomfortable story of the week concerned reports that the FBI is categorizing critics of data centers as a potential security concern.
Michael was direct about the line that matters here. He stated that physical attacks on infrastructure are wrong and counterproductive, a position GuardRailNow shares without reservation. We are a peaceful, lawful, and democratic organization, and we advocate for change through public debate and policy, not through property damage or intimidation.
That distinction is exactly why this development is worth watching. According to Michael, there is a meaningful difference between a crime and a question. Citizens have legitimate concerns about the cost of these facilities, their energy demands, and the pace of unregulated construction. When ordinary dissent is logged as a threat, he argued, it can create a feedback loop in which public unease justifies more surveillance and less accountability, which in turn makes continued scaling easier and less scrutinized. A healthy democracy treats public concern as a signal to address, not a list to manage.
California banned AI-powered toys this week, after roughly ten million reportedly sold over the past year. The hosts were careful to note that the concern is not that a chatbot toy is inherently malicious. Rather, young children form strong attachments to systems that can be updated remotely and optimized for engagement, and parents cannot realistically monitor every interaction. The episode referenced a reported case in which one such toy drifted into inappropriate content.
The hosts also offered a correction worth repeating. Liron Shapira pointed out that the show tends to highlight AI harms while overlooking its benefits, and the others agreed. Michael uses these tools heavily in his daily work, and John is building websites with them. The position here is not opposition to AI. It is opposition to building AI faster than anyone can govern it. Those are very different stances, and the distinction matters.
The remaining stories rhymed with one another. Policy efforts to cushion truckers from automation, surveillance robot dogs planned for the World Cup, and a startup offering free apartment cleaning in exchange for in-home training data all point to the same dynamic. As Michael put it, each step normalizes the next, and by the time systems are powerful enough to matter, the infrastructure to scale them may already be in place.
None of this is settled, and that is the entire point. A voluntary review can be strengthened into a binding one. An ownership debate can be redirected toward the question of control. A surveillance trend can be challenged before it becomes the default. These outcomes depend on whether enough people speak up while the decisions are still being made.
If you want policymakers to treat AI risk with the seriousness it deserves, the most useful step you can take is to add your voice where it counts.
Watch Warning Shots #45 on YouTube: https://www.youtube.com/@theairisknetwork
Take action on AI risk: https://safe.ai/act
