AI researchers warn superintelligence could pose an extinction risk in new video interviews

Researchers and former employees from OpenAI, Google DeepMind and Anthropic are publicly warning that advanced AI could become an existential threat, with some saying the chance of human extinction is worryingly high if superintelligence is ever built without reliable control systems. The interviews, released by nonprofit Palisade Research, underscore how sharply the AI safety debate has escalated inside the industry itself.

The new video series matters because the people sounding the alarm are not outsiders or critics at the margins. They are insiders from companies building the systems most likely to shape the next phase of AI development, and their remarks show that concern about catastrophic risk has moved from abstract theory into the mainstream of the field.

In the clips, researchers argue that the technology could outstrip human oversight, create incentives that are hard to reverse, and become powerful enough to cause irreversible harm if deployed carelessly. The interviews do not produce a single solution, but they do offer a rare public look at the range of fear, uncertainty and disagreement inside the AI community.

What did the researchers say?

Several interviewees used unusually stark language to describe the possible consequences of creating systems smarter than people in most domains.

Geoffrey Irving, who previously worked at OpenAI and Google DeepMind, said he believes the odds of human extinction are roughly even if the technology is mishandled. Neel Nanda, a research scientist at Google DeepMind, said the possibility of AI causing human extinction is still uncomfortably large, putting the risk in the double digits. Former OpenAI researcher Daniel Kokotajilo described a future superintelligent system as unimaginably capable and argued that humanity currently has no dependable way to control one.

Those views are not presented as a consensus, but they reflect a growing thread of concern among researchers who think the field is advancing faster than safety work can keep up.

Geoffrey Irving, a former OpenAI and Google DeepMind employee, said in the interview series that the chance of human extinction is, in his view, about a coin flip.

Neel Nanda of Google DeepMind said there is still at least a 10% chance that AI leads to human extinction, which he described as an alarmingly high figure.

Daniel Kokotajilo, formerly of OpenAI, argued that a superintelligent system would be “god-like” in power and that current methods do not provide reliable control.

Why these warnings are getting louder now

The interviews arrive amid a broader wave of anxiety about how quickly frontier AI systems are improving, how widely they are being deployed, and whether companies are moving faster than regulators or researchers can respond.

AI safety has become a major topic within the industry, but it remains a concept with blurred edges. For some researchers, safety means preventing catastrophic misuse. For others, it means reducing bias, improving robustness or avoiding harmful behaviors in everyday products. The uncertainty over definitions makes it harder to agree on what the real risks are and what should be done first.

That ambiguity is part of what makes the new videos notable. Rather than presenting a neat roadmap, the researchers describe an unresolved problem: they believe the stakes are enormous, yet they do not share a single practical plan for guaranteeing control over systems that may eventually surpass human capability in key areas.

How close are researchers to a solution?

They are not close to a universally accepted solution. The interviews suggest that the field has more urgency than certainty, with some experts saying safety tools are improving but remain far behind the speed of model development.

That gap matters because the most dangerous scenario discussed in the videos is not a clearly malicious AI, but one that is powerful, misaligned and difficult to steer. In that world, even minor errors in design or deployment could have large consequences.

Who is behind the new interview series?

The videos were gathered and released by Palisade Research, a nonprofit that says it studies AI capabilities and motivations. The organization created the frominside.ai project as a way to collect firsthand perspectives from people who build frontier systems and study them up close.

Rather than a single documentary or one-off interview, the project is structured as a set of shorter conversations. Viewers can watch the full discussions or jump between edited highlights grouped around common themes such as hype, corporate incentives and the personal reasons researchers continue working in a field they say could be dangerous.

That format gives the series a double value: it acts as both a warning and a window into how industry insiders rationalize their own participation in the race to build more capable systems.

How do AI researchers justify working on dangerous technology?

Many say they believe the only way to reduce the risk is to stay inside the companies building the models.

In one of the more pointed comments, Google researcher Mary Phuong said listeners should be skeptical of her remarks because she is employed by a major lab. Her framing reflects an important tension in AI safety debates: the same companies that stand to benefit commercially from advanced models are also the ones funding much of the safety research.

Nanda offered a similar argument for remaining in the field. He suggested that leaving would not stop the work from continuing, and said he would only stay if he believed his work meaningfully reduced existential risk. In essence, the researchers portray themselves as trying to improve an inevitable technology rather than pretending they can halt it.

• Some researchers see their jobs as a form of harm reduction.
• Others believe internal pressure is more effective than public criticism from the outside.
• Several acknowledge that company incentives make objectivity difficult.

What superintelligence means in this debate

Superintelligence refers to AI systems that would exceed the best human minds across most or all important tasks. In practice, that could mean systems with the ability to solve scientific problems, write software, influence people, and carry out long-horizon planning better than any individual or institution could.

The concern is not just that such systems would be smart. It is that they might become strategically powerful in ways humans cannot reliably predict, monitor or constrain. If a system can outthink its operators and pursue goals misaligned with human interests, the consequences could be severe.

That is why some researchers in the video series frame the issue in existential terms. Once a system becomes hard to control, they argue, there may be no second chance to get deployment right.

Why the AI safety debate is so hard to pin down

AI safety sounds straightforward, but in practice it covers several different worries that do not always point to the same solution. One camp focuses on near-term harms such as bias, hallucinations, fraud or misuse. Another is focused on far more serious long-term scenarios, including a loss of human control over powerful systems.

That split creates confusion outside the field and frustration inside it. A researcher might use the phrase “AI safety” to mean tightening a model’s behavior before release, while another uses it to mean preventing a future civilization-level accident. Those are related goals, but they are not interchangeable.

The new interviews highlight that divide. The people involved clearly share alarm, but they do not share a single theory of the problem or a single policy agenda. Some want more research. Others want stricter deployment limits. Still others seem to believe the work must continue under close internal scrutiny because stopping entirely is unrealistic.

What makes the issue different from earlier tech scares?

The scale and autonomy of frontier AI systems are what make the current debate so intense. Unlike many past technologies, advanced AI may be able to improve workflows, influence human decision-making and generate novel strategies at machine speed.

That creates a scenario in which the system itself could become part of the risk rather than just a tool used by people. If models can plan, persuade and adapt faster than humans can supervise them, then traditional safety oversight may no longer be enough.

How the industry is responding

So far, the response has been uneven. AI companies continue to race ahead with more capable models, larger training runs and broader product rollouts, while also funding red-teaming, alignment research and policy engagement.

That combination reflects the central contradiction of the field: the same organizations pushing the technology forward are also the ones being asked to restrain it. Critics say this produces predictable conflicts of interest. Supporters argue the labs are among the few institutions with the expertise and resources to do the work responsibly.

The interviews do not resolve that contradiction, but they sharpen it. If insiders believe the worst-case outcome could be catastrophic, then the question becomes whether current governance, corporate incentives and technical safeguards are remotely adequate.

What happens next?

Public pressure around AI safety is likely to keep building as models become more capable and more embedded in everyday products. The more that developers talk openly about extinction-level risk, the harder it becomes to dismiss safety concerns as fringe speculation.

At the same time, the field still lacks shared benchmarks for proving a system is safe in the long term. That means future debates are likely to focus not only on model performance and profitability, but also on control, containment and whether any lab can credibly guarantee that a future system will remain aligned with human intentions.

The Palisade Research interviews may not settle those arguments, but they capture a rare moment in which some of the people closest to the technology are saying the quiet part out loud: if superintelligence is built before control problems are solved, the consequences could be far more serious than most of the public has been led to expect.

Key timeline of the debate

For now, the central message from the videos is not that disaster is inevitable, but that a growing number of people building advanced AI believe the downside risks are too large to ignore. That alone makes the debate harder for the rest of the industry — and the public — to avoid.