OpenAI CEO Sam Altman says society may need to accept some harmful consequences of artificial intelligence in order to preserve its benefits and people's ability to use the technology. His comments have reopened a harder question for the AI industry: which harms can reasonably be accepted, and who gets to make that decision?
Speaking on POLITICO's Decoded podcast, Altman said OpenAI believes the world should accept some negative outcomes in exchange for the benefits of AI and individual agency, according to India Today's report on the interview and expert reactions. The remark did not specify exactly which harms Altman considered acceptable.
That distinction matters because Altman has separately taken a much harder line on catastrophic AI risk. In remarks to the United Nations Security Council on September 23, he said estimates of catastrophic risk at 10 percent, 1 percent, 12 percent or even 0.1 percent were all "not remotely acceptable" and said developers should not train models unless they can make an extremely strong case that the systems can remain under human control.
Altman's remarks are not an acceptance of extinction risk
The two statements address different levels of risk. Altman's podcast comment suggests that some negative consequences may accompany widespread AI adoption. His official UN Security Council remarks, however, make clear that OpenAI does not publicly describe catastrophic loss of human control as an acceptable cost of progress.
At the UN, Altman said powerful AI systems should remain understandable enough to monitor and controllable enough for humans to intervene. He also said competitive pressure between companies was not a justification for taking excessive technological risk and that OpenAI had slowed development in the past when necessary.
That leaves a difficult middle ground. AI systems can cause serious damage without reaching a science-fiction-style extinction scenario, and experts contacted by India Today pointed to cybercrime, fraud, deepfakes, job disruption and the concentration of economic power as examples of risks that are already easier to imagine and measure.
Why experts are talking about a 5 percent extinction figure
Amit Das, CEO of Think360.ai, pointed to a major survey of AI researchers in explaining why even uncertain catastrophic risks deserve attention. He argued that people would not willingly board an aircraft if they were told it carried a 5 percent chance of disaster.
The figure comes from the 2023 Expert Survey on Progress in AI, published in 2024. The survey involved 2,778 researchers who had published at six major AI research venues.
In one question, the median respondent assigned a 5 percent probability to future AI advances causing human extinction or similarly permanent and severe disempowerment of humanity. Another wording produced a median estimate of 10 percent, while 38 percent to 51 percent of respondents across related questions assigned at least a 10 percent probability to outcomes as severe as human extinction.
Those numbers are not a measured probability that AI will end humanity. They are subjective forecasts collected from researchers under uncertainty, and the results varied depending on how questions were worded. The survey methodology and results published by AI Impacts make that distinction important.
The airplane comparison is therefore an argument about risk tolerance, not evidence that AI has a scientifically established 5 percent chance of causing extinction. It asks whether a low-probability outcome should still demand strong precautions when the possible damage is extraordinarily large.
Experts point to harms that do not require a doomsday scenario
Anirban Mukherji, CEO of cybersecurity company miniOrange, told India Today that AI can make existing threats cheaper, faster and easier to scale. He cited convincing phishing attempts, deepfakes, voice cloning, automated fraud and attempts to steal credentials or manipulate users.
Those concerns are particularly relevant because the people harmed by AI-enabled abuse may never have chosen to use an AI system themselves. Someone whose voice is cloned for a scam, whose identity is used in manipulated media or whose account is targeted by automated fraud is not making the same voluntary trade-off as a person choosing to use an AI chatbot.
Nishant Chandra, co-founder of Newton School of Technology, highlighted the risk of deepfake and voice-cloning scams across Indian languages, where misuse could spread faster than public awareness and detection tools. He also identified disruption to entry-level work, overdependence on AI and concentration of opportunity among a small number of companies or countries as concerns.
These are materially different from claims that AI will cause human extinction. They are nearer-term risks that can be assessed through fraud reports, employment data, cybersecurity incidents and the ways AI systems are deployed.
Who should decide which AI harms are acceptable?
The experts also challenged the idea that technology companies alone should determine the acceptable cost of AI deployment.
Dr N Vivek, chairman of the Centre of AI in Business at Great Lakes Institute, told India Today that no single company, government or individual should unilaterally decide which social costs are acceptable in exchange for technological progress. He also noted that the people receiving AI's benefits may not be the same people who bear its costs.
That creates a practical policy problem. A business may gain productivity from automation while a worker faces displacement. A consumer may gain access to a powerful AI assistant while another person becomes the target of an AI-generated impersonation scam.
Das similarly argued that harms imposed on people who did not choose to accept the risk require decisions beyond product teams, pointing to regulators and courts as institutions with a role in determining responsibility.
The benefits remain part of the calculation
The experts interviewed by India Today did not argue that AI has no value. Potential benefits cited in the discussion included higher productivity, automation of repetitive work, more time for tasks involving judgment and creativity, and wider access to useful services.
Prashanth Tondapu, CEO of software company Innostax, pointed to potential gains in areas such as education and healthcare access in India. Bajrang Saharan, founder of Pressmate OS, said AI and automation could reduce repetitive manual work and allow employees to spend more time on problem-solving, customer interaction and creative tasks.
The dispute is therefore not simply between people who believe AI is beneficial and people who believe it is dangerous. The more precise disagreement concerns how much risk should accompany those benefits and whether the people exposed to the consequences have meaningful protection.
What happens next
Altman's two sets of comments establish an important boundary but do not fully resolve it. He has argued that society may need to tolerate some negative effects of AI, while also saying publicly that even comparatively small estimates of catastrophic risk are unacceptable and that advanced systems must remain under human control.
What remains unresolved is where acceptable disruption ends and unacceptable risk begins. Cybercrime, deepfakes, employment disruption and other foreseeable harms can be addressed through technical safeguards, company policies and regulation, but the debate becomes much harder when researchers are discussing uncertain events with potentially irreversible consequences.
For now, there is no scientifically established percentage representing the chance that AI will cause human extinction. What is confirmed is that large surveys have found substantial concern among AI researchers, while Altman himself has publicly said catastrophic risk should not be accepted merely because AI also promises major benefits.