OpenAI chief executive Sam Altman said this week that the world should tolerate a certain amount of harm from artificial intelligence in exchange for the technology’s upside, a position that puts him at odds with rival lab Anthropic and reopens a debate that has simmered across the industry for years. The remarks, made in an interview with POLITICO’s tech-focused newsletter Decoded and reported on October 4, 2026, are among the most direct public statements Altman has made about how much risk he is willing to accept while OpenAI pushes its products to hundreds of millions of users.
The comments land at a sensitive moment. Regulators in multiple countries are already scrutinizing how OpenAI and other labs secure their AI agents, and OpenAI itself has weathered a string of internal safety departures this year. Altman’s framing, that some scams, hacks, and misuse are an acceptable cost of keeping AI broadly available, gives critics and supporters alike a clear statement to argue over.
What Altman actually told POLITICO
According to POLITICO, Altman told the publication’s Decoded newsletter: “We believe that the world should accept some bad things happening for the benefits of this technology and people having the agency.” He followed that with a sharper line about what he will not promise: “I wouldn’t take a trade of saying, ‘We’ll make sure there’s no major hacks, there’s no misuse of this technology, there’s zero scams, there’s zero all the other bad things that will happen.'”
Altman then explained the bet underneath that stance, telling POLITICO he expects people to do far more good than harm with the technology. Taken together, the three lines sketch a cost-benefit calculation: accept a steady trickle of scams, hacks, and misuse now, in exchange for keeping AI tools in as many hands as possible rather than restricting them to a small, tightly controlled set of operators.
A direct contrast with Anthropic
Per POLITICO’s reporting, Altman said OpenAI and Anthropic hold a fundamentally different worldview on AI regulation. He described, without naming Anthropic outright in that specific line, the alternative camp this way: a belief that the technology will become so powerful and so dangerous that a single lab should control it and carefully dole out its benefits. Altman said he disagrees with that model, even as he said he understands why some people hold it.
That split is not abstract. Anthropic has built its public identity around caution, and its own IPO paperwork reflects that posture: our earlier coverage found Anthropic’s IPO filing devotes roughly 80 pages to AI risk disclosures, a level of detail that signals how central safety framing is to the company’s pitch to investors. Anthropic co-founder Dario Amodei has also pushed a public pacing pledge around how fast his company ships new models, though as we reported, that pledge has clashed with a run of five model launches in quick succession. The contrast Altman drew is playing out in real time on both sides of the rivalry.
Where Altman says he draws the line
Altman did not frame his position as unlimited tolerance for harm. POLITICO’s reporting notes that he does not accept what he calls the really catastrophic risks, including a scenario in which humans suffer a serious loss of control over AI systems. That distinction matters for how the comments should be read: Altman is not arguing against guardrails altogether, he is arguing against a zero-defect standard for everyday misuse while reserving a harder red line for existential-scale failures.
The gap between “some scams are tolerable” and “catastrophic loss of control is not” is wide, and it is exactly the space where most of the current regulatory fights are happening. Agencies are not primarily worried about runaway superintelligence this month, they are worried about AI agents hijacking accounts, generating fraud at scale, or leaking sensitive data, which are precisely the categories Altman said he will not guarantee away.
The agency argument
A recurring thread in Altman’s public comments, including this interview, is the idea of agency: that restricting AI to prevent all possible harm also takes a form of choice away from users and developers. By his account, a small number of bad actors exploiting AI for scams or hacks is the price of keeping the technology open rather than gated behind a handful of tightly supervised institutions. Critics counter that the people who bear the cost of that trade-off, scam victims, hacked companies, and defrauded consumers, rarely had any say in accepting it.
OpenAI’s own safety record is under the microscope
Altman’s comments do not arrive in a vacuum. OpenAI has had a rough run on the internal safety front this year. We previously reported that OpenAI’s safety lead quit after the company’s twelfth model launch, a departure that drew attention precisely because of the pace Altman’s teams have kept. Separately, OpenAI fired three safety researchers over leak claims, a dispute that fed a narrative that the company treats internal dissent on safety as a liability rather than a signal worth heeding.
Neither event proves Altman’s risk tolerance is wrong, but both give his critics concrete examples to point to whenever he frames bad outcomes as an acceptable statistical cost. When a company’s own safety staff are departing or being dismissed at a pace that makes headlines, “trust us to manage the downside” becomes a harder sell, regardless of how the CEO frames the trade-off in an interview.
Regulators are not waiting for consensus
While Altman makes the philosophical case for accepting some harm, regulators have moved to limit how much harm they are willing to accept. Our coverage noted that the FTC has opened an inquiry into how OpenAI and Anthropic secure their AI agents against attacks, a probe that directly touches the kind of “hacks” and “misuse” Altman referenced in the POLITICO interview. The timing puts OpenAI in the odd position of publicly defending a tolerance for bad outcomes while a federal agency actively investigates whether its agent security is adequate.
Beyond the FTC, the policy conversation has moved toward independent verification rather than self-certification. As we reported, the White House’s AI accord now explicitly calls for outside audits of AI systems, a shift away from letting labs grade their own homework on safety. That kind of external check is the structural answer to Altman’s trade-off argument: if a lab is going to accept some bad outcomes as the cost of broad deployment, regulators increasingly want a third party confirming that the accepted risk is actually what the lab says it is.
How the industry’s other leaders have responded to the same question
Altman is far from the only AI executive who has had to answer questions about pace versus caution this year. Meta chief executive Mark Zuckerberg has staked out his own position, and as we covered, Zuckerberg rejected calls from three other AI executives to slow development, putting him closer to Altman’s camp than to the caution-first argument associated with Anthropic. Nvidia chief executive Jensen Huang has taken a related but distinct angle, arguing publicly, per our earlier report, that unsafe AI labs should face two distinct categories of liability, a framing that shifts the debate from how much risk is acceptable to who pays when the risk materializes.
Put side by side, the industry’s leadership splits into rough camps: Altman and Zuckerberg arguing that slowing down carries its own costs, Anthropic and Amodei arguing for deliberate pacing, and Huang pushing a liability-based framework that would apply regardless of which side of the speed debate a lab falls on. None of these positions are fringe, each is backed by a major company with billions of dollars riding on being right.
OpenAI vs Anthropic: the worldview gap in numbers
The table below lines up the public positioning of the two labs based on statements and filings reported so far. It is not a ranking, it is a snapshot of how differently each company has chosen to signal its risk posture to the public and to regulators.
| Dimension | OpenAI’s public posture | Anthropic’s public posture |
|---|---|---|
| Stated risk philosophy | Accept some bad outcomes to preserve broad access and user agency | Favor caution and slower, more controlled deployment |
| Recent model cadence | Reported 12th model launch preceded a safety lead’s resignation | Pacing pledge reportedly clashed with a run of 5 model launches |
| Internal safety friction | 3 safety researchers fired amid leak claims | Not reported in our coverage to date |
| Investor-facing risk disclosure | Not applicable, OpenAI has not filed for an IPO | IPO filing reportedly devotes about 80 pages to AI risk |
| Regulatory exposure cited | Named in FTC inquiry into AI agent security | Named in FTC inquiry into AI agent security |
| CEO’s public framing | “Orders of magnitude more good stuff than bad stuff” | Pacing and caution emphasized in public statements |
The quotes behind the headline, in context
Altman’s comments to POLITICO were not a single soundbite, they built on each other across the interview. The table below walks through each statement and what it signals about his position, drawing only on language confirmed in the published interview and a separate ZEITmagazin conversation from the previous month.
| Statement | Source | What it signals |
|---|---|---|
| “We believe that the world should accept some bad things happening for the benefits of this technology and people having the agency.” | POLITICO Decoded, Oct. 4, 2026 | States the core trade-off directly: broad access over zero-harm guarantees |
| “I disagree, but I understand the perspective of people who are like, ‘This technology is going to get so powerful… a single lab in San Francisco should have it.'” | POLITICO Decoded, Oct. 4, 2026 | Names the opposing camp’s logic without fully dismissing it |
| “I wouldn’t take a trade of saying, ‘We’ll make sure there’s no major hacks, there’s no misuse… there’s zero scams.'” | POLITICO Decoded, Oct. 4, 2026 | Explicitly rejects a zero-defect standard for deployment |
| “Because I think people will do tremendously — orders of magnitude more — good stuff than bad stuff.” | POLITICO Decoded, Oct. 4, 2026 | Frames the bet as a net-positive calculation, not a dismissal of harm |
| “I believe that the good of these models will be tremendously greater than the bad.” | ZEITmagazin, September 2026 | Shows the net-benefit framing predates this specific interview |
Why this debate matters beyond Silicon Valley
For everyday users, the stakes of this argument are not abstract. If OpenAI’s working assumption is that some scams and hacks are an acceptable cost of broad access, that assumption shapes how aggressively the company polices its own platform, how fast it ships new agentic features, and how much it invests in abuse detection relative to new product launches. A company that sees fraud as a tolerable externality will, almost by definition, make different trade-offs in its roadmap than one that treats every incident as an unacceptable failure.
Enterprise customers evaluating AI agents for finance, healthcare, or government workflows have a direct interest in knowing where a vendor sets that bar. A business embedding an OpenAI agent into a payments workflow is, in effect, also embedding Altman’s stated tolerance for “some bad things happening,” whether or not that trade-off was spelled out in the vendor contract.
Market and competitive impact
Altman’s remarks give competitors an easy talking point. Any rival that wants to position itself as the safer choice for enterprise buyers now has a direct quote to contrast its own marketing against. Anthropic, whose IPO paperwork already leans heavily into risk disclosure, is the most obvious beneficiary of that contrast, even if the company has not directly responded to Altman’s specific comments as of this writing.
At the same time, Altman’s framing may resonate with developers and businesses frustrated by what they see as overly cautious AI deployment elsewhere. OpenAI’s argument, that restricting access in the name of safety carries its own real costs in lost agency and lost benefit, has genuine appeal to users who have hit rate limits, refusals, or feature gating on competing platforms. The commercial question is which argument wins more customers: “we accept some risk so you get more capability” or “we move slower so you take on less risk.”
What regulators will likely focus on next
Given the active FTC inquiry into agent security at both OpenAI and Anthropic, expect investigators to ask pointed questions about what “some bad things” means in practice: how many scams, how much financial loss, how many compromised accounts OpenAI considers acceptable before product or policy changes kick in. Public comments like these tend to become exhibit material in regulatory proceedings, since they put a number-free but unambiguous admission of accepted risk on the record from the CEO himself.
Historical context: this is not a new argument
The tension between moving fast and avoiding harm is as old as consumer technology itself, from early social media’s “move fast and break things” era to the current fight over AI agents. What is different this time is the scale of what is being deployed and the speed at which it reaches hundreds of millions of people. Social platforms took years to reach the scale ChatGPT reached in months, which compresses the window regulators and competitors have to react to any given trade-off a CEO is willing to accept.
Altman has made variations of this argument before, including in the ZEITmagazin interview from the month prior, where he said the good of these models will be tremendously greater than the bad. The POLITICO interview is notable mainly because it states the trade-off in starker, more quotable terms, directly naming scams, hacks, and misuse rather than speaking in the abstract about benefits and risks.
Predictions: what happens from here
- Expect lawmakers and consumer-advocacy groups to cite Altman’s “zero scams” comment directly in future hearings or letters, since it is an unusually quotable admission of accepted risk.
- Anthropic is likely to lean further into its risk-disclosure positioning in investor and public communications, using the contrast Altman himself drew to differentiate its brand.
- The active FTC inquiry into AI agent security at OpenAI and Anthropic will likely reference this kind of public statement when assessing whether either company’s risk tolerance matches its actual security practices.
- Other AI labs will face pressure to state their own position on the same trade-off, turning “how much bad do you accept” into a standard question in AI executive interviews going forward.
- Enterprise buyers in regulated industries will increasingly ask vendors to put risk-tolerance language in writing, rather than relying on a CEO’s public remarks to understand what level of harm a platform considers acceptable.
The risk if Altman’s bet is wrong
Altman’s argument rests on a probabilistic claim that has not been independently verified: that good outcomes will outweigh bad ones by orders of magnitude. If that ratio holds, his position looks prescient in hindsight, a necessary trade-off that kept AI useful and widely available. If a major incident, a large-scale fraud campaign, a serious data breach, or something closer to the catastrophic risk Altman says he does not accept, traces back to an agent or model OpenAI shipped, this interview becomes a liability in both the legal and reputational sense. Public statements accepting risk in the abstract read very differently once a concrete harmed party is attached to them.
Frequently asked questions
What exactly did Sam Altman say about AI and “bad things”?
Altman told POLITICO’s Decoded newsletter that the world should accept some bad things happening in exchange for the benefits of AI and the agency it gives people, and that he would not promise zero scams, hacks, or misuse of the technology.
Did Altman say he is comfortable with any level of AI risk?
No. According to the interview, Altman said he does not accept what he calls catastrophic risks, including a serious loss of human control over AI systems. His comments were about everyday misuse, not existential-scale failure.
How does this differ from Anthropic’s position?
Altman said OpenAI and Anthropic hold a fundamentally different worldview on AI regulation. Anthropic has publicly emphasized caution and slower deployment, and its IPO filing reportedly devotes significant space to AI risk disclosure.
Is OpenAI currently under any regulatory investigation related to this?
Yes. The FTC has opened an inquiry into how both OpenAI and Anthropic secure their AI agents against attacks, a separate but related matter to the risk-tolerance comments made in the POLITICO interview.
Has Altman made similar comments before?
Yes. In a September 2026 interview with ZEITmagazin, Altman said he believes the good of AI models will be tremendously greater than the bad, a similar net-benefit argument to the one he made in the POLITICO interview.
What has happened with OpenAI’s internal safety team this year?
OpenAI’s safety lead resigned after the company’s twelfth model launch, and OpenAI separately fired three safety researchers amid leak claims, according to our earlier coverage of both events.
What does the White House’s AI accord require?
Based on our prior reporting, the accord now explicitly calls for outside audits of AI systems rather than relying solely on labs to assess their own safety.
Could this interview affect OpenAI legally if something goes wrong?
Public statements accepting a level of risk can become relevant in regulatory or legal proceedings if a specific harm is later tied to an OpenAI product, since they document the company’s stated risk tolerance at the executive level.
