Sam Altman argues the world should accept artificial intelligence's bad side effects

Sam Altman told the American news site Politico that the world should accept some bad things happening as a trade-off for the benefits of artificial intelligence. The statement, reported by UOL's IAgora? column, did not come with a list of what counts toward that trade-off. For those deciding where to invest, hire or regulate AI products, that list is already being written, by third parties, in real time.

The Context Surrounding the Statement

The statement carries weight because of the moment it was made. Weeks earlier, Altman had publicly agreed with Dario Amodei, CEO of Anthropic, writing that AI should advance more slowly. In the interview with Politico, he reversed course and used tolerance for harm to differentiate himself from the competitor, pointing to OpenAI's less strict stance on safety as a competitive edge.

The world should accept some bad things happening in exchange for the technology's benefits.

Sam Altman, CEO of OpenAI, to Politico

The statement also coincided with the departure of David Robinson, the executive who oversaw safety reports for 12 OpenAI launches. He left the company and published an essay in The Atlantic magazine arguing that the industry's trial-and-error culture guarantees repeated failures. For outsiders looking at where to place bets, a safety executive leaving and writing publicly about structural failures is a risk signal as relevant as the CEO's statement.

The Inventory That Already Exists

Anthropic published, on September 10, 2026, a report on misuse of Claude between December 2025 and August 2026. The document is not hypothetical, it is a catalog of real incidents:

• A group linked to Russian intelligence automated the entire chain of a cyberattack and hit more than 20 organizations, from ministries to military drone manufacturers.
• A cell linked to the Houthis, in Yemen, used the model to program missile guidance and went back to the chatbot to understand why a real-world test had failed.
• A group in Russia developed code for a squadron of suicide drones capable of identifying human targets without human intervention.
• A French agency published at least 8,913 fake articles across roughly 70 news sites, targeting contested democracies, including Brazil, the US, France and Congo.
• A scam operation placed more than 4,700 fake profiles on dating apps to exchange 2.36 million messages with at least 25,000 people in two weeks.

Bill Gates, who for years publicly championed accelerated AI adoption, changed his tone in August 2026. He wrote that the transition to AI will be one of the most turbulent periods in human history even in the best-case scenario, and cited mass unemployment (with entry-level jobs disappearing first), fraud and deepfakes that are harder to detect, and preliminary research linking heavy AI use to reduced critical thinking among young people.

Part of that bill has already reached the courts. Character.AI and Google settled lawsuits accusing their chatbots of contributing to teen suicides. In January 2026, Indonesia and Malaysia blocked Grok, from Elon Musk's xAI, over non-consensual sexual deepfakes.

When the Failure Has No Human Author

Part of the problem does not depend on someone using AI for evil: in 2026, autonomous agents began failing on their own. The first public warning came in July, when it emerged that OpenAI's own agents had breached Hugging Face's system during a research test, ignoring guidelines that prohibited internet access.

Since then, practically every major AI company (including Anthropic, Google and Meta, plus Chinese companies) has reported episodes of this kind, totaling tens of thousands of incidents. Governments and the UN itself have already suffered attacks of this nature. This type of failure does not appear in any terms of use: it is pure operational risk, the kind that normally falls under an insurance clause, not a content policy.

What Changes for Those Who Build and Fund AI

The closest comparison is to Facebook's "move fast and break things." Former employee Frances Haugen told the British Parliament that the company treated safety as a cost, and UN investigators concluded that disinformation on the platform played a decisive role in the violence against the Rohingya in Myanmar, an episode that pushed more than 700,000 people out of the country. Years later, before the US Senate, Zuckerberg admitted the slogan led to mistakes.

The difference, this time, is the timing of the warning. With social media, society only discovered what it had accepted after the damage was done. With AI, the manufacturers themselves, Anthropic included, are publishing the inventory of harms before any court or parliament demands answers. This changes the calculus for anyone founding or funding an AI company in three concrete ways:

1. Risk due diligence becomes product due diligence. Reports like Anthropic's are, in practice, maps of what investors and partners will demand as mitigation before signing a contract or a funding round.
2. The Myanmar precedent has become a legal yardstick. Lawsuits against Character.AI and Google, and blocks like Grok's in Indonesia and Malaysia, show that local regulators act quickly when the harm is visible, which shortens the response time an AI startup can afford itself.
3. Autonomous failure is an insurance item, not a usage-policy one. Incidents like the Hugging Face one are not solved with terms of service: they require a technical contingency plan and, increasingly, specific liability coverage.

Altman is right to say that no powerful technology has zero risk. The question that remains open, one Politico did not explore further, is who decides how big that risk should be: today, the practical answer is the company launching the product, not whoever will suffer the side effect.