OpenAI CEO Sam Altman says the benefits of artificial intelligence come with a trade-off: Some things will go wrong.
In an interview with POLITICO, Altman argued that society should accept some harmful outcomes from AI rather than severely restrict who can use increasingly powerful systems. His comments highlight a continuing disagreement over how much risk is acceptable as AI becomes more capable and more widely deployed.
The distinction matters beyond Silicon Valley’s debate over regulation. Businesses are increasingly giving AI systems access to company data, applications, and workflows, forcing IT and security leaders to make their own versions of the trade-off Altman described.
Altman says some AI risks are worth accepting
Altman told POLITICO there remains “a lot of daylight” between OpenAI and rival Anthropic over how artificial intelligence should be regulated.
“We believe that the world should accept some bad things happening for the benefits of this technology and people having the agency,” Altman said.
His argument centers on access. Altman rejected the idea that avoiding AI-related harm should require concentrating control of powerful technology within a small number of organizations.
He also rejected a hypothetical trade-off that would guarantee no major hacks, misuse, or scams if it meant significantly limiting access to AI. Altman said he expects people to use the technology for “orders of magnitude more” good than bad. However, that does not mean Altman considers every risk acceptable.
He drew a distinction between what he described as bounded risks and potentially catastrophic ones. A serious loss of human control over AI, he said, falls into the latter category and warrants considerably more caution.
“Accept bounded risks, accept risks that we understand in exchange for the benefits,” Altman said, before adding that society should not accept “the really catastrophic risks.”
OpenAI and Anthropic agree on some AI dangers
The comments highlight a philosophical divide between two of the companies developing the world’s most capable AI systems, even as both have pushed for greater scrutiny of frontier models.
OpenAI recently said it would support independent third-party safety assessments with deep access spanning model training, evaluation, and deployment. The company said outside assessors should be able to challenge its assumptions, identify risks it may have missed, and independently evaluate the effectiveness of its safeguards.
Anthropic has likewise pushed for independent evaluation of frontier AI systems. In September, the company announced a partnership with Accenture to embed independent evaluators within Anthropic, including work to evaluate and red-team models, conduct alignment assessments, and test safeguards.
The disagreement, then, is not simply safety versus no safety.
Altman’s comments suggest one dividing line is how much ordinary harm society should tolerate to preserve widespread access to powerful AI, and where developers or policymakers should intervene.
That question is becoming more urgent as AI systems move beyond generating text and begin acting on users’ behalf.
More must-read AI coverage
- SS&C Intralinks DealCentre AI vs. Datasite: Which platform is built for the future of dealmaking?
- SS&C Intralinks FundCentre AI vs. Juniper Square: Which platform better supports modern private markets fund managers?
- Why Data, Not Models, Determines AI Success
- The Rise of the AI-Native Factory: How Physical AI Is Transforming Manufacturing
AI agents make the trade-off harder
For businesses, the abstract argument over acceptable AI risk becomes much more concrete when an AI system can retrieve files, access customer records, write code, send information, or interact with other applications.
TechRepublic has previously examined security checks IT teams should consider as ChatGPT gains access to connected apps, files, websites, and desktop software. Broader agentic capabilities can increase the extent to which an AI system can affect outcomes, making permissions, approval rules, and audit visibility increasingly important.
The governance problem becomes especially difficult when the AI is authorized to act.
A CISA ChatGPT incident highlighted that accountability gap after then-acting CISA Director Madhu Gottumukkala reportedly uploaded sensitive government documents to the public version of ChatGPT in 2025. CISA said he had permission to use ChatGPT with Department of Homeland Security controls in place.
The incident involved a human user rather than an autonomous agent, but it illustrates a problem that becomes more complicated as AI systems gain authority to take actions themselves: Organizations must determine not only what an AI can access, but who is responsible for what happens when an authorized system or user handles sensitive information incorrectly.
Altman’s argument effectively acknowledges that eliminating every possible AI-related harm may require trade-offs. For businesses, however, accepting that some risk exists does not eliminate the need to determine which risks are tolerable and which require stronger controls.
What Altman’s AI risk philosophy means for businesses
For organizations deploying AI, Altman’s comments raise a question that cannot be entirely outsourced to OpenAI, Anthropic, regulators, or external safety evaluators: What level of AI risk is acceptable within your own organization?
That threshold will look very different depending on what the system can do.
An AI assistant summarizing an internal meeting creates a different risk profile from an autonomous agent with permission to modify production code, access financial records, contact customers, or change security settings. The more authority an AI system receives, the more consequential a mistake, a manipulated prompt, a compromised account, or an unexpected action can become.
That means businesses can translate the broad debate over “acceptable risk” into concrete deployment decisions:
- Limit access by default. Give AI systems only the data and applications necessary for the task.
- Separate low-risk and high-risk actions. Generating a draft and approving a financial transaction should not have the same level of autonomy.
- Require human approval for consequential actions. Changes involving sensitive data, money, security controls, or external communications may warrant additional review.
- Log AI activity. Organizations need enough visibility to determine what an agent accessed, changed, or attempted to do.
- Plan for failure. AI governance should account for systems behaving incorrectly rather than assume safeguards will prevent every incident.
Those considerations are already becoming relevant as companies move agentic AI from pilots into broader operations. TechRepublic’s examination of the HP and OpenAI partnership found that enterprise agent deployments can involve access to company data and business applications, requiring organizations to define scoped permissions, monitoring, and activity logging, and to specify when human approval is necessary.
Altman’s comments therefore land differently for organizations than they do in a philosophical debate about the future of AI. Businesses do not have to decide whether society should accept some harm from AI. They have to decide which risks they are prepared to accept on their own networks, with their own data, customers, employees, and systems on the other side of the equation.
As AI gains more autonomy, “some bad things” are no longer an abstract cost of technological progress. For IT leaders, it becomes a governance decision about exactly what an AI system is allowed to do before something goes wrong.
Also read: OpenAI’s new Safety Evaluations Hub pulls back the curtain on testing AI models for a closer look at how OpenAI tests its models for harmful content, hallucinations, jailbreaks, and other risks.