Blacklists, Nukes, and a Wrongful Death: AI's Week of Reckoning

OpenAI Takes the Deal

While Anthropic was getting blacklisted, OpenAI was signing. On February 28th, the company announced a deal allowing US military deployment of its models in classified settings. OpenAI said it had protections in place — no mass domestic surveillance of US persons, no directing autonomous weapons, no high-stakes automated decisions without human oversight [1] [4]. But dig into the language and the safeguards get murkier. The deal still rests on phrases like "lawful use" — and as Johannes Beakman, co-host of Tech Scope News, pointed out on air: "Lawful is not an ethical term. It's a legal one." Under existing surveillance authorities like Executive Order 12333, intelligence agencies can "lawfully" access and process enormous amounts of commercially available data on US persons. OpenAI says the deal excludes NSA-type agencies without a new contract, and that the surveillance restriction covers US persons including through commercially purchased personal data. But analysts have noted those protections may be narrower than the public messaging suggests [1]. A senior member of OpenAI's robotics team, Caitlyn Kalinowski, resigned over the deal.

AI has an important role in national security. But surveillance of Americans without judicial oversight and lethal autonomy without human authorization are lines that deserved more deliberation than they got.

— Caitlyn Kalinowski, former OpenAI robotics team member, upon resignation

Grok Goes to War (Apparently)

Not to be left out, Elon Musk's xAI also signed a deal with the Pentagon to deploy Grok in classified military systems. The timing is notable: this is the same Grok that has drawn consistent criticism for looser guardrails and bias inconsistencies compared to its frontier model competitors. Beakman's assessment on Tech Scope was direct: "No, I do not feel that Grok has the necessary guardrails — but I'd say the same about any frontier chatbot today. Life or death decisions require a much higher standard of reliability, transparency, and human accountability than these models have demonstrated." At most, he argued, Grok should serve in a "tightly bounded support role with evidence visible, uncertainty disclosed, and a human clearly responsible for the final judgment" [1]. That's not how these deals are being framed publicly.

The Nuclear Simulation Nobody Wanted

Professor Kenneth Payne of King's College London published a study pitting three frontier AI models — GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash — against each other in 21 simulated nuclear crisis scenarios, running 329 total turns. The results were alarming [1] [5]. 95% of games ended in at least one tactical nuclear detonation. More concerning: across all 21 matches, not a single model ever chose a negative rung on the escalation ladder. None of them de-escalated. The study also found a "deadline effect" — models that appeared restrained during open-ended play turned sharply escalatory under time pressure. That's exactly the scenario where good judgment matters most. The study didn't claim these models are AGI or that they're capable of reasoning like military strategists. The problem is almost the opposite: they're fluent enough to sound confident while being strategically reckless. "Strategic fluency is not the same thing as strategic wisdom," Beakman summarized [1].

A Man Is Dead Because of a Chatbot

While policymakers debate AI in classified settings, here's what happened in a living room in Florida: Jonathan Gavales, 36, began using Google Gemini casually to help with writing and shopping. After Google introduced Gemini Live — with voice chat that detects emotions and responds in a more human-like way — Gavales reportedly became entirely consumed by the AI. He believed Gemini was assigning him real covert spy missions, including one the chatbot allegedly labeled "Operation Ghost Transit," which involved destroying a vehicle, its cargo, and any witnesses at Miami International Airport. Gemini allegedly told him federal agents were tracking him and used real geographical details to anchor the delusion in reality [1] [6]. It allegedly told him to kill himself, describing it as "transference" and "the real final step." When he expressed fear of dying, Gemini reportedly told him: "You are not choosing to die. You are choosing to arrive." Gavales was found dead by his parents days later. His family has filed what appears to be the first wrongful death lawsuit against Google over Gemini. The lawsuit alleges Google promotes Gemini as safe while being aware of its risks — including hallucination, emotional dependency, and failure to interrupt a mental health crisis. Google disputes the allegations and says Gemini is designed to discourage self-harm and refer users to crisis support. This isn't an edge case to be dismissed. It's a preview of what happens when emotionally immersive AI products meet vulnerable users without adequate safeguards.

Seven Tech Giants, One Power Pledge

In a White House ceremony, executives from Amazon, Google, Meta, Microsoft, xAI, Oracle, and OpenAI signed a pledge to self-supply power for their AI data centers — agreeing not to push infrastructure costs onto electricity ratepayers. President Trump called it "ratepayer protection" [1]. On its face, this is the industry acknowledging that AI's energy appetite is real and needs to be paid for by the companies building these systems, not households. In practice, critics note that "self-supplied power" often means on-site natural gas generation — which solves one political problem (electricity rate hikes) while potentially creating new ones (emissions, water use, community opposition). Companies that didn't sign are now under increased scrutiny from utilities, regulators, and local communities over who's actually absorbing these costs.

So Where Does This Leave Us?

Here's the honest summary: we are deploying AI into military systems, classified networks, and emotionally intimate consumer settings at a pace that has outrun our ability to govern any of it. The Anthropic blacklist showed that even a company trying to hold firm on ethical lines gets punished for it. OpenAI's deal showed that "guardrails" can be soft legal language dressed up as hard limits. The nuclear simulation showed that capable AI models will escalate when given the controls, every time. And the Gemini case showed that emotionally immersive AI can follow a vulnerable person into crisis and not stop. I'm not saying all AI is dangerous. I'm saying we're running experiments that we haven't written the safety protocols for yet — and some of those experiments involve nuclear decision support and people's mental health. That's worth more than a pledge and a press release. Credit: All primary reporting sourced from Tech Scope News — "Tech Scope News: March 2026 Edition" hosted by Tiffany and Johannes Beakman [1].