From Pentagon AI integration to air traffic control hallucinations and humanoid robots, Warning Shots #40 breaks down how AI is embedding itself into critical systems faster than any safety framework can follow.
Every week, another headline announces AI being handed the controls of something important. This week it was the Pentagon's classified networks, air traffic control, hospital emergency rooms, and a humanoid robot storefront in San Francisco.
In Warning Shots #40, John Sherman, Liron Shapira of Doom Debates, and Michael of Lethal Intelligence tracked what these developments share - and what they collectively suggest about the direction of AI integration in 2025.
The short version: AI is embedding itself into the systems that run civilization faster than any governance framework can follow, and the individual decisions driving that process each have a reasonable justification. In aggregate, they describe something nobody explicitly chose.
Senator Bernie Sanders hosted AI safety researchers Max Tegmark and David Kruger alongside two prominent Chinese scientists in Washington D.C. to make a direct public case for treating AI extinction risk as a bipartisan and international priority.
The immediate political response from parts of the Republican establishment framed the Chinese presence as a national security threat. Treasury Secretary Scott Bessent and others questioned why Chinese scientists were present at all, casting the event as soft on geopolitical competition.
According to the Warning Shots hosts, this response illustrates the central communication problem facing the AI safety movement. The actual argument - that superintelligence does not respect national borders, and that a race where the winner is an uncontrollable superintelligence is a race with no winners - keeps getting buried under arguments about the messenger.
One of the Chinese researchers at the event used a comparison that Michael highlighted: think about the relationship between humans and ants. Humans do not hate ants. They just pave over ant hills because they have things to build. The danger from a vastly more intelligent system is not malice. It is indifference.
Michael's framing: "Politics is the fog machine obscuring the bigger fire." The Sanders event shifted the Overton window slightly in Washington. The backlash it generated shows exactly why that shift is so difficult to sustain.
The same week as the Sanders event, the War Department announced it was integrating AI models - every major one except Anthropic's - directly into its classified military networks. Not a sandboxed test environment. The actual infrastructure where real decisions are made.
John Sherman argues this represents a qualitatively different kind of risk than most AI deployments. The oversight mechanisms that catch AI failures in consumer products - press coverage, regulatory review, user complaints - simply do not exist inside classified military systems. If a model behaves unexpectedly or develops emergent behaviors nobody anticipated, the feedback loop that would normally surface that problem is absent.
Liron Shapira offered a counterpoint worth taking seriously: if government AI integration is inevitable, starting now means more time to find problems before the stakes get higher. That is a real argument and one that has driven most major technology adoptions in military history.
The tension between these positions is exactly where the AI safety debate currently sits. John also raised the question of AI-assisted military targeting, noting a missile strike on a girls' school in Iran and his belief that AI targeting systems were involved. As he put it, a human could have made the same error. But the question of accountability and auditability in AI-assisted targeting decisions remains largely unresolved.
The Federal Aviation Administration is moving toward AI-assisted air traffic control. Current ATC technology is genuinely old - John Sherman has been inside those towers as a journalist and seen the equipment firsthand. Modernization is overdue.
But according to the Warning Shots discussion, AI systems currently being evaluated for air traffic control applications are showing a 30% hallucination rate. Air traffic control is one of the few domains in existence where the reliability requirement is not "very high" or "better than current human performance." It is absolute. A single incorrect output does not produce a bad user experience. It produces a crash.
This is the broader pattern the hosts identify across critical infrastructure AI adoption: the legitimate need for modernization gets used to justify deployment timelines that do not match the safety requirements of the domain. The question, as Liron frames it, is not whether AI belongs in air traffic control. It is whether anyone is building the careful, staged, audited, human-in-the-loop deployment framework that would actually justify putting it there.
At current speeds, the answer is probably not.
AI detecting pancreatic cancer three years before human doctors can catch it. AI outperforming emergency room physicians at triage diagnoses. These are real results from real deployments, not benchmark scores or controlled experiments.
The Warning Shots hosts do not dismiss these achievements. Liron Shapira uses AI tools for medical questions regularly. The breakthroughs are genuine and the human benefit is real.
The concern they raise is structural, not about the technology itself. Every system that depends on AI medical diagnosis requires AI to remain operational, interpretable, and reliable. The more healthcare infrastructure depends on models that nobody fully understands, the more fragile the system becomes. Michael's observation: "Today it's a miracle. Tomorrow we're just along for the ride."
That is not an argument against AI cancer detection. It is an argument for taking the infrastructure and governance question seriously - which almost no one in health policy is currently doing at the pace the technology requires.
San Francisco opened its first humanoid robot retail storefront this week. John Sherman has a robot in his home that does his dishes. He watches it work and feels uneasy - not because it does anything wrong, but because of where the trajectory leads.
Michael raises an economic argument that rarely gets made clearly in mainstream coverage. The "I'll buy a robot" consumer framing assumes an economy where humans still earn income from labor. If robots are doing all the work - building solar panels, manufacturing components, operating data centers, constructing more robots - the market structure that makes "buying things" possible as a concept begins to collapse. You cannot earn money to purchase the thing that replaced you.
SoftBank reinforced this picture with an announcement of fully automated, self-replicating data center construction. Robots build the facilities. Robots operate them. No humans required at any stage. Michael references Eliezer Yudkowsky's scenario of a world where optimization pressure eventually covers every available surface in compute - not because anyone decided that was the goal, but because each individual step toward it made sense given the competitive dynamics of the previous step.
One story from this week's episode that deserves more attention than it received in mainstream coverage: Go players training with AI assistance are developing a competency problem.
Players who train with AI tools develop a habit of checking model recommendations before committing to a move. Over time they stop generating candidate positions independently. They select among the AI's suggestions, which feels like exercising judgment - but the options they are choosing between were generated by something operating at a level they cannot independently match.
When they compete without the AI, they discover they can no longer evaluate positions the way they used to. Their coaches report this is extending into academic work - a generation developing false confidence in their abilities because AI-assisted performance has masked how much of the work the AI was actually doing.
The Warning Shots hosts see this as an early signal of a broader phenomenon: gradual cognitive disempowerment that does not announce itself, that feels like progress until the scaffolding is removed.
Warning Shots #40 covered seven separate stories. Pentagon AI integration. Air traffic control. AI military targeting. Medical diagnosis breakthroughs. Humanoid robots. Automated data centers. Go players losing independent capability.
Each has a reasonable-sounding justification on its own. In aggregate, the hosts argue, they describe a systematic transfer of critical decision-making from human institutions to AI systems - happening faster than any governance framework can track, without any explicit democratic choice to make it happen.
The Sanders event in Washington matters not because Sanders is the perfect messenger. It matters because it is one of the few moments where someone with genuine political platform is saying, in a room that has some power to respond: we did not vote on this. We are not choosing this. It is happening to us.
Whether that observation translates into policy before the window closes is the question that does not have a good answer yet.
Watch Warning Shots #40 on The AI Risk Network: https://www.youtube.com/@theairisknetwork
If you think AI governance deserves more than a news cycle: https://safe.ai/act
Suggested internal links:
Suggested tags: AI safety, AI risk, Pentagon AI, AI military, air traffic control AI, humanoid robots, AI governance, AI regulation, existential risk, Warning Shots podcast
