Sam Altman has drawn a public line around how much harm he thinks AI should be allowed to cause on its way to delivering its benefits. In an interview with POLITICO’s Decoded newsletter published October 4, 2026, the OpenAI chief executive said the world needs to tolerate a certain amount of damage from the technology rather than demand a guarantee that nothing goes wrong.
“We believe that the world should accept some bad things happening for the benefits of this technology and people having the agency,” Altman told POLITICO. It is a blunt way to frame a debate that OpenAI, Anthropic, regulators and AI safety researchers have circled for years without settling. Altman’s version settles it, at least for OpenAI: harm is a cost of doing business, not a reason to slow down.
What Altman Told POLITICO’s Decoded
The remarks came during a sit-down with POLITICO’s technology-focused newsletter, Decoded, and were reported on October 4, 2026. Altman was asked to weigh the upside of deploying AI broadly against the downside risks that come with giving millions of people access to tools that can write code, scan documents, and act with real autonomy. He did not hedge. Altman said he would not accept a trade that eliminated every form of harm, because the cost of that trade would be a single gatekeeper deciding who gets access to the technology at all.
“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 said, according to POLITICO. The quote reads as a direct rebuttal to critics who argue AI companies should slow deployment until safety guarantees catch up. Altman’s answer: that guarantee doesn’t exist, and chasing it costs more than it saves.
The Trade-Off He Refuses to Make
Altman’s reasoning rests on a bet about proportion. He told POLITICO that he expects the good uses of AI to swamp the bad ones by a wide margin. “Because I think people will do tremendously — orders of magnitude more — good stuff than bad stuff,” he said. That single sentence carries the weight of OpenAI’s current strategy: ship broadly, accept that some fraction of users will cause harm, and bet that the aggregate benefit still comes out far ahead.
It’s a utilitarian argument, and it is also a business argument. A company that waits for proof of zero scams and zero misuse before shipping a product would ship nothing. OpenAI’s CEO is telling the public, in plain language, that the company is not going to wait. The statement lands at a moment when OpenAI has already dealt with real incidents of misuse tied to its own agents, including breaches involving a rogue agent at a second Australian agency, which makes the comment read less like abstract philosophy and more like a defense of decisions already made.
Where Altman Draws the Line
Altman was careful to distinguish the harms he will tolerate from the ones he won’t. Scams, hacks, and misuse fall into the first bucket. A different category, which he described to POLITICO as “the really catastrophic risks,” falls into the second. That category includes what he called a potential “serious loss of control” to AI systems, a phrase that echoes years of warnings from researchers who study long-term AI safety.
Catastrophic risk
Altman did not define exactly what counts as catastrophic in the interview, but the framing matters. He is drawing a line between harms that are costly but recoverable, like a scam or a data leak, and harms that aren’t recoverable at all. The bet-the-aggregate-good argument only works if the bad outcomes stay in the first category. A single event in the second category breaks the math entirely, which is presumably why Altman carved it out as the one trade he won’t make.
Loss of control
“Serious loss of control” is a term of art in AI safety circles, generally referring to scenarios where a system acts outside human oversight in ways that can’t be reversed. By naming it explicitly, Altman is acknowledging the concern exists, even as he argues most of today’s actual harms don’t come close to that threshold. Whether that distinction holds up depends on how AI systems behave as they’re given more autonomy, not on how Altman describes them in an interview.
OpenAI and Anthropic: Two Labs, One Technology, Different Bets
Altman used the interview to draw a contrast with Anthropic, saying the two labs hold a fundamentally different worldview on how AI should be regulated. He didn’t attack Anthropic by name beyond that framing, but the implication is clear: where Altman argues for tolerating harm in service of broad access, Anthropic has built its public identity around a more cautious posture toward deployment and oversight.
Anthropic’s calculus
The split isn’t new, but Altman’s framing gives it a sharper edge. OpenAI has already clashed publicly with Anthropic over how the two companies view AI risk, and Anthropic’s own paperwork backs up its cautious reputation. The company’s IPO filing reportedly devotes 80 pages to AI risk disclosures, a volume of risk language that signals how seriously Anthropic expects investors, and regulators, to scrutinize safety claims before money changes hands.
A Year of Flashpoints Between the Two Labs
Altman’s comments don’t land in a vacuum. 2026 has been a rough year for the internal safety narrative at OpenAI specifically, with several episodes that make his “accept some bad things” framing sound less theoretical and more like a running defense. The table below lines up some of the year’s clearest flashpoints.
| Event | What Happened | Why It Matters |
|---|---|---|
| Oct. 4, 2026 | Altman tells POLITICO’s Decoded the world should accept “some bad things” from AI | Puts a public, named philosophy behind OpenAI’s deployment pace |
| Safety leadership exit | OpenAI’s safety lead departed after the company’s 12th model launch | Raised questions about whether safety review kept pace with shipping speed |
| Staff dismissals | OpenAI fired three safety researchers over leak claims | Drew criticism over how the company handles internal dissent on safety |
| Regulatory scrutiny | The FTC opened an inquiry into OpenAI and Anthropic over AI agent attacks | Puts federal regulators directly inside the OpenAI-Anthropic risk debate |
| Anthropic’s filing | Anthropic’s IPO filing devotes 80 pages to AI risk | Shows the financial market already pricing in safety and regulatory exposure |
| Washington’s accord | A new White House AI accord makes outside audits explicit | Signals policymakers lean toward independent verification, not self-reporting |
Historical Context: Silicon Valley’s Old Playbook Meets AI
“Move fast and break things” has guided consumer software for two decades, and Altman’s comments fit squarely inside that tradition. Facebook shipped features before privacy reviews caught up. Uber launched in cities before regulators had rules for ride-hailing. The pattern worked because the breakage was usually recoverable: a privacy scandal, a fine, a policy fix.
AI changes the stakes because some of the downside scenarios researchers worry about aren’t recoverable in the same way a data leak is. Altman’s own line about catastrophic risk and loss of control concedes that difference. What he’s really arguing is that today’s actual harms, scams, hacks, misuse, still sit on the recoverable side of that ledger, so the old playbook still applies for now. That is a testable claim, not a settled one, and it is the central disagreement running through the AI safety debate this year.
Market Impact: What Investors and Enterprises Are Watching
Altman’s framing matters to markets because it tells enterprise buyers and investors how OpenAI plans to balance speed against liability going forward. A CEO who says publicly he won’t demand zero scams before shipping is also telling customers that some incidents are priced into the product, not treated as one-off failures.
That posture cuts two ways financially. Enterprises that want AI deployed fast may read it as confidence. Enterprises burned by prior incidents, including the breaches tied to OpenAI’s own agents, may read it as a company normalizing risk it should be reducing. Anthropic’s decision to spend 80 pages of its IPO filing on risk disclosure looks, by contrast, like an attempt to court the second group: investors who want the risk itemized up front rather than absorbed quietly after the fact.
The Regulatory Backdrop Altman Is Pushing Against
Altman’s comments arrive while regulators are actively testing how much oversight AI companies should accept. The FTC has already opened a probe into both OpenAI and Anthropic over AI agent attacks, and the new White House AI accord now spells out a requirement for outside audits rather than leaving verification to the companies themselves. Lawmakers have pushed in the same direction: Rep. Maxine Waters has called for a formal investigation into OpenAI following a string of agent-related incidents.
Altman’s “accept some bad things” framing is, in effect, a pre-emptive argument against exactly this kind of oversight. If regulators accept his premise, that most harms are a tolerable cost of broad access, the case for aggressive new rules weakens. If they don’t, his comments become evidence for the opposite conclusion: that OpenAI’s leadership isn’t planning to self-correct and needs outside pressure to do it.
Competitive Comparison: Where Google, Meta and Mistral Stand
Altman isn’t the only AI executive arguing against a slowdown. The debate over acceptable risk now runs across the entire frontier AI field, with different companies staking out visibly different ground.
Meta’s bet
Meta’s Mark Zuckerberg has made a similar case in public, rejecting calls for an AI slowdown and breaking from peers who argue for more caution. The overlap between Zuckerberg’s stance and Altman’s is notable: both executives are framing speed itself as the responsible choice, arguing that ceding ground to more careful competitors carries its own risk.
The rest of the field
Anthropic remains the clearest counterweight, with a public identity built on caution and a filing that spends real page count on risk rather than upside. Google and Mistral have generally avoided the blunt language Altman and Zuckerberg use, preferring to talk about safety processes rather than draw a line between acceptable and unacceptable harm in public. That makes Altman’s remarks unusually direct for an industry that typically answers risk questions with process descriptions instead of a stated threshold.
In Altman’s Own Words: The Quotes That Matter
Four lines from the interview do most of the work in explaining Altman’s position, and each one is worth reading in full rather than paraphrased.
- “We believe that the world should accept some bad things happening for the benefits of this technology and people having the agency,” said Sam Altman, OpenAI CEO, to POLITICO’s Decoded.
- “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,'” said Sam Altman, OpenAI CEO, to POLITICO’s Decoded.
- “I disagree, but I understand the perspective of people who are like, ‘This technology is going to get so powerful, and it’s so dangerous, that a single lab in San Francisco should have it and make sure nothing bad happens, and kind of figure out how to dole out the benefits,'” said Sam Altman, OpenAI CEO, to POLITICO’s Decoded.
Read together, the four quotes form a complete argument: accept distributed harm, reject centralized control, bet on scale, and draw one hard exception for catastrophic, irreversible outcomes. It’s the clearest public statement of OpenAI’s risk philosophy to date, delivered not in a safety paper but in a newsletter interview.
Altman’s Risk Ledger
The table below sorts Altman’s own words into what he will tolerate and what he won’t, based directly on the POLITICO interview.
| Category | Altman’s Position | Example From the Interview |
|---|---|---|
| Everyday misuse (scams, hacks, fraud) | Accepts it as the cost of broad access | Won’t demand “zero scams” or “zero” misuse before shipping |
| Individual agency and access | Backs giving people the technology despite the risk | Says the benefit comes from “people having the agency” |
| Catastrophic, irreversible risk | Does not accept this trade | Explicitly excludes “serious loss of control” from the acceptable-harm bucket |
| Net benefit calculation | Bets the good outweighs the bad by a wide margin | Expects “orders of magnitude more” good outcomes than bad ones |
| Industry alignment | Diverges from Anthropic’s approach | Describes a “fundamentally different worldview” between the two labs |
What This Means for Everyday ChatGPT Users
For the millions of people using ChatGPT and OpenAI’s other products day to day, Altman’s remarks don’t change much about how the tools work tomorrow morning. What they do change is the frame for understanding incidents when they happen. OpenAI’s CEO has now said, on record, that scams and misuse tied to the technology are an expected cost rather than a failure to be eliminated.
That’s a meaningful shift in tone from years of safety messaging that framed every incident as a bug to be patched. Users should expect OpenAI to keep shipping quickly, keep absorbing a baseline level of misuse as a known cost, and reserve the brakes for the catastrophic scenarios Altman carved out as his one red line.
Predictions: What Comes Next
A few things seem likely to follow from an unusually candid public statement like this one, based on how similar remarks have played out across the AI industry this year.
- Expect lawmakers already pushing for AI oversight, including Rep. Waters, to cite Altman’s quotes directly in future hearings or letters.
- Expect Anthropic to avoid directly rebutting Altman by name, while continuing to lean on cautious messaging and detailed risk disclosure to draw a contrast without a public fight.
- Expect critics of OpenAI’s safety record to treat the “zero scams” line as a talking point any time a new incident involving OpenAI’s products surfaces.
- Expect the FTC’s existing inquiry into OpenAI and Anthropic to reference public statements like this one as evidence of each company’s stated risk tolerance.
- Expect other AI lab executives to face follow-up questions asking them to state their own red lines as clearly as Altman just did.
The Bigger Question Altman Didn’t Fully Answer
Altman’s interview draws a clean line between acceptable harm and catastrophic risk, but it doesn’t say who decides when a harm has crossed from one category into the other. That judgment call currently sits with OpenAI itself, which is precisely the arrangement that regulators, lawmakers, and competitors like Anthropic are now pushing to change. Resources like the NIST AI Risk Management Framework exist to make that judgment less subjective, but adoption across the industry remains uneven.
Altman’s bet is that the math works out in OpenAI’s favor if the company is left to make that call itself. The pressure building from the FTC, the White House, and Congress suggests plenty of people disagree, and the next few months of AI policy fights will largely be an argument over who gets to draw that line. For continuing coverage of how OpenAI and Anthropic’s rivalry plays out, follow shattered.io’s AI and machine learning section, and for background on the industry’s broader read on AI timelines, see reporting from TechCrunch and Wired.
Frequently Asked Questions
What exactly did Sam Altman say about AI’s risks?
Altman told POLITICO’s Decoded newsletter that the world should accept “some bad things happening” from AI in exchange for the technology’s benefits and the agency it gives people, while saying he would not accept a trade that promised zero hacks, zero misuse, and zero scams.
When and where did Altman make these comments?
The comments were reported on October 4, 2026, in an interview with POLITICO’s technology-focused newsletter, Decoded.
Does Altman think AI has zero serious risks?
No. He explicitly said he does not accept “the really catastrophic risks,” including what he called a potential “serious loss of control” to AI, drawing a clear distinction from the everyday harms he says are tolerable.
How does OpenAI’s position differ from Anthropic’s?
Altman said OpenAI and Anthropic hold a fundamentally different worldview on AI regulation. Anthropic has publicly leaned toward a more cautious posture, reflected in an IPO filing that devotes roughly 80 pages to AI risk disclosures.
What is POLITICO’s Decoded newsletter?
Decoded is POLITICO’s technology-focused newsletter, which covers tech policy and AI industry developments. Altman’s comments were published there on October 4, 2026.
Has Altman specified which “bad things” he means?
He named specific categories in the interview: major hacks, misuse of the technology, and scams. He framed these as recoverable costs, distinct from catastrophic or irreversible outcomes.
What happens next after these remarks?
The comments arrive while the FTC is already investigating OpenAI and Anthropic over AI agent attacks, and while lawmakers such as Rep. Maxine Waters are pushing for more oversight of OpenAI. Altman’s framing is likely to become a reference point in that ongoing regulatory debate.
Did Altman name specific incidents when he made these comments?
No. The interview quotes describe a general philosophy about acceptable risk rather than referencing any single incident by name.
