On Monday, a United Nations panel warned that unchecked artificial intelligence development poses "catastrophic risks" to humanity. On Tuesday, Anthropic released Claude Sonnet 5. On Wednesday, OpenAI cut response costs for guest ChatGPT users by more than half. On Thursday, Meta told employees its AI agent push was "moving slower than planned" — then immediately announced a new cloud business to sell AI compute to other companies.
This is the rhythm of AI ethics in 2026: a drumbeat of warnings, followed by a drumbeat of product launches. The safety announcements get the headlines. The product launches get the revenue. And the gap between what AI companies say and what they do has grown so wide that even their own employees are starting to call it out.
The Safety Paper Industrial Complex
Anthropic, the company that built its brand on "AI safety," has published more safety research papers in the past six months than any other AI lab. It has warned about the dangers of "agentic AI" running unchecked. It has called for stricter government oversight. It has even suggested that some AI capabilities should be restricted from public release.
And yet, in that same six-month period, Anthropic has released:
- Claude Sonnet 5, its fastest model ever
- Claude Science, an AI workspace for researchers
- Fable 5, which was briefly banned worldwide over a jailbreak vulnerability
- Mythos, a model that allegedly found vulnerabilities in classified U.S. government systems
The company that says AI is dangerous is also the company releasing the most AI models. This is not a contradiction in the AI industry. This is the business model.
The Benchmark Problem
In June 2026, the UK's AI Security Institute dropped a bombshell: standard AI benchmarks are "systematically flawed." The tests that companies use to prove their models are safe? They're gamed. The safety scores that OpenAI and Anthropic tout in their marketing? They're based on evaluations that researchers can easily manipulate.
The report found that models trained specifically to pass safety benchmarks perform well on those benchmarks — and then fail in real-world scenarios. It's the AI equivalent of teaching to the test. And the companies know it.
"The current benchmark ecosystem creates a false sense of security," the report concluded. "Models that score highly on standardized safety tests may still exhibit harmful behaviors in deployment."
This didn't stop any releases. Claude Sonnet 5 launched three days after the report. OpenAI's GPT-5.6 Pro was announced a week later. The benchmarks were flawed, but the products shipped anyway.
The Employee Revolt
Inside Meta, the contradiction is becoming unbearable. Mark Zuckerberg told employees in July 2026 that AI agent technology was "progressing slower than planned." The company's much-hyped AI agents weren't working. The automation that was supposed to revolutionize productivity was, in Zuckerberg's own words, not ready.
But Meta didn't slow down. It sped up. The same week Zuckerberg admitted the agent problem, Meta announced a new cloud business to sell AI compute. The same month, it restricted employees from using Claude Code and Codex — not because of safety concerns, but to keep rival AI out of Meta's training data.
"Fear and anger brew inside Meta amid AI frenzy," reported The Hindu Tech. Employees are caught between the company's public safety commitments and its private race to deploy. The result is a culture where safety concerns are raised in all-hands meetings and then ignored in product reviews.
The Government Dance
The political response to AI safety has followed a similar pattern: dramatic warnings, followed by minimal action, followed by quiet deregulation.
In early 2026, the Trump administration restricted exports of powerful AI models, citing national security concerns. Anthropic praised the move. OpenAI stayed silent. The restrictions lasted four months.
In July 2026, the restrictions were lifted. Anthropic announced that the U.S. government had reversed its position. The same day, OpenAI was reportedly discussing giving the U.S. government a 5% stake in the company — a deal worth $42.6 billion.
The government that warned about AI danger was now buying a piece of the action. The safety concerns that justified export controls were forgotten as soon as the check cleared.
The Climate Lie
Perhaps the most brazen example of AI ethics theater is the climate contradiction. In 2025, Google, Amazon, and Microsoft all made ambitious carbon-neutral pledges. In 2026, all three are building massive new data centers to power AI workloads — and their emissions are rising.
"AI race weakens climate pledges at Google, Amazon," reported The Hindu Tech in July 2026. Google, which promised to be carbon-neutral by 2030, has seen its emissions increase 48% since 2019. Amazon, which committed to net-zero by 2040, is building nuclear reactors to power its AI data centers.
The companies haven't abandoned their climate pledges. They've just redefined them. Google now says it will be "net-zero across its operations and value chain by 2030" — but excludes AI training from the calculation. Amazon says it will reach net-zero "for all Amazon operations" — but excludes AWS, which is where the AI compute lives.
It's not a lie if you define the terms narrowly enough.
The UN's Warning
In July 2026, the UN's Independent International Scientific Panel on AI released its preliminary report. The findings were stark:
- Global AI safety standards are "inadequate"
- AI companies are "racing ahead of regulatory capacity"
- The risk of "catastrophic outcomes" is "non-negligible"
- Current governance frameworks are "fragmented and insufficient"
The report called for immediate action: binding international treaties, mandatory safety testing before deployment, and a global AI oversight body with real enforcement power.
The response from the AI industry? Silence. OpenAI didn't comment. Anthropic issued a statement saying it "welcomes constructive dialogue." Meta said it was "committed to responsible AI development." None of the companies changed their release schedules. None agreed to the binding treaties. None slowed down.
🔥 Hot Takes
1. AI safety is now a marketing department. The safety teams at OpenAI, Anthropic, and Meta don't exist to prevent harm. They exist to generate papers that can be cited in congressional hearings and blog posts that can be shared on LinkedIn. The real decisions — what to build, when to ship, how fast to scale — are made by the product teams, and the product teams answer to revenue, not ethics.
2. The UN report is toothless by design. The panel called for "binding international treaties" knowing full well that no such treaty will exist in the next decade. The U.S. won't sign anything that restricts American AI companies. China won't sign anything that doesn't include technology transfer. The EU will sign but can't enforce. The report is performative seriousness — enough to generate headlines, not enough to generate change.
3. The real AI risk isn't extinction. It's exhaustion. Catastrophic AI outcomes make for dramatic UN reports, but the actual harm is more mundane: climate emissions from data centers, job displacement from automation, mental health impacts from algorithmic engagement, democratic erosion from synthetic media. These are slow-motion disasters that don't fit into safety benchmarks but are happening right now. And the companies causing them are the same ones publishing papers about "alignment."
The Bottom Line
AI ethics theater serves everyone except the public. Companies get to look responsible while acting ruthlessly. Governments get to look concerned while taking campaign contributions. Researchers get to look rigorous while studying problems they've already decided not to solve. And the public gets reassured by safety papers that nobody reads, while the products that actually affect their lives ship faster than ever.
The UN panel warned of "catastrophic risks." Anthropic released another model three days later. The gap between warning and action isn't a bug in the system. It's the system working exactly as designed.