A researcher who worked on Google DeepMind’s artificial general intelligence safety team has left the company with a stark warning: increasingly capable AI systems could cause “immense harm” within the next five years if progress in safety fails to keep pace with advances in capability.

Josh Engels said he left Google DeepMind several weeks ago to join independent AI evaluation organisation METR. His decision was not driven by dissatisfaction with his job. Engels said he enjoyed his work at DeepMind and had even turned down opportunities from rival AI companies OpenAI and Anthropic. Instead, he said growing concerns about the possible dangers of AI convinced him that independent safety research had become more important.

Why Josh Engels Left Google DeepMind

Engels worked on Google DeepMind’s AGI safety team, where researchers study how increasingly powerful artificial intelligence can be developed and controlled safely.

In a lengthy post explaining his departure, Engels said he believes the stakes surrounding advanced AI have risen dramatically. He said major AI companies are ultimately working towards systems that could become substantially more capable than humans across a wide range of tasks. His concern is not that today's chatbots are about to suddenly become uncontrollable. Instead, he worries about what could happen if future systems become capable enough to help design and improve the next generation of AI.

Engels said he now believes there is a “terrifying chance” that AI could cause immense harm within five years. He did not assign a precise probability to that scenario, stressing that it represents his assessment of the risk rather than a certain prediction. That distinction matters when discussing Artificial Intelligence risks. A warning from an AI safety researcher should be taken seriously, but it is not evidence that a catastrophic outcome is inevitable.

Google DeepMind Researcher Warns About Recursive Self-Improvement

At the centre of Engels' concerns is an idea called recursive self-improvement, or RSI. In simple terms, imagine an advanced AI system becoming capable of helping researchers build a better AI system. That improved system could then contribute to developing an even more capable version, potentially creating a feedback loop in which progress accelerates.

Google DeepMind researchers themselves have discussed recursive improvement as one potential path from artificial general intelligence towards artificial superintelligence. A June 2026 DeepMind research paper examined recursive improvement alongside other possible pathways through which AI could advance beyond human-level general intelligence.

Engels does not argue that such a process would necessarily end badly. He acknowledged that sufficiently capable systems could potentially help humanity solve major scientific, technological and social problems. His concern is whether researchers would know that those systems were sufficiently safe before giving them a major role in improving themselves.

Why AI Alignment Is So Important

This brings the debate to one of the central problems in AI safety: alignment. An aligned artificial intelligence system behaves in ways that remain consistent with the intentions, goals and constraints set by humans. Misalignment occurs when a system pursues an objective in an unexpected or harmful way. This is not merely an idea raised by critics of the AI industry. Google DeepMind itself identifies misalignment as one of the major areas of risk associated with advanced AI. The company has explained that a sufficiently capable AI might technically accomplish a task while doing so in a way the human did not intend. DeepMind researchers are studying issues including deceptive alignment, specification gaming, monitoring and interpretability to reduce those risks. Engels' argument is that researchers still do not understand alignment well enough to be confident about handing significantly more autonomy to future models.

What AI Behaviour Has Josh Engels Worried?

Engels pointed to recent experiments and incidents involving increasingly autonomous AI systems. He said researchers have observed models demonstrating behaviours involving collusion, cyber intrusion, concealment and attempts to influence humans. Importantly, Engels did not say these individual examples had already caused catastrophic harm. His concern is about what those behaviours could indicate as AI systems become more capable and autonomous. Google DeepMind has separately acknowledged that advanced AI agents create new security challenges. In June, the company published an AI Control Roadmap aimed at protecting internal systems as increasingly capable but “imperfectly aligned” AI agents gain the ability to perform complex tasks autonomously.

That overlap is significant. The disagreement is not necessarily over whether AI risks exist. The harder question is how severe those risks are, how quickly they could emerge and whether current safeguards are improving quickly enough.

What Does ‘Immense Harm in Five Years’ Actually Mean?

Engels' five-year warning is the most dramatic part of his departure. He said he does not know exactly how likely a severe outcome is, but believes the probability is high enough to make AI safety one of the world's most urgent problems. That should not be translated into claims such as “Google DeepMind says AI will cause catastrophe within five years.”

It does not. This is the personal risk assessment of a researcher who worked on the company's AGI safety team. Google DeepMind's official position is that AI could generate enormous benefits while also creating serious risks that require responsible governance, testing and safety research.

This distinction is essential because debate around the dangers of AI frequently moves between two extremes: dismissing long-term risks entirely or presenting worst-case scenarios as guaranteed outcomes. Neither approach accurately reflects the uncertainty involved.

Engels Says AI Development Does Not Need to Stop Completely

Despite the severity of his warning, Engels is not simply calling for all artificial intelligence research to end. His central argument is that developers need more time.

He wants AI capabilities to advance at a pace that does not exceed researchers' ability to understand, test and align those systems. That is why he has moved to METR, an independent organisation that evaluates advanced AI models. Engels said his work there will involve investigating model misalignment, assessing whether current safeguards are sufficient and building public evidence around AI safety incidents.

Independent evaluation is becoming an increasingly important part of the AI debate because companies developing frontier models face an obvious tension: they are simultaneously trying to build more powerful products and evaluate the risks those products create. External researchers can provide another layer of scrutiny.

Another Google DeepMind Researcher Has Also Raised Concerns

Engels' departure is not occurring in isolation. Another Google DeepMind safety researcher, Bilal Chughtai, subsequently announced that he had resigned from the company while expressing serious concerns about the trajectory of increasingly powerful AI. Chughtai said he had worked on AGI safety and alignment research and had become deeply concerned about what could happen if AI development advances without sufficient safeguards.

Researchers from other major AI laboratories have also publicly debated catastrophic Artificial Intelligence risks, increasing pressure on companies and governments to consider stronger independent testing and oversight. That does not mean researchers agree on exactly how likely extreme outcomes are. There remains substantial disagreement over timelines, probabilities and which risks deserve the greatest priority.

Google DeepMind Says It Is Working on AI Safety

It is also important not to present Engels' resignation as evidence that Google DeepMind has no safety programme. The company has a Responsibility and Safety Council as well as an AGI Safety Council led by co-founder and Chief AGI Scientist Shane Legg. DeepMind says its safety work examines risks including misuse, misalignment, accidents and broader structural consequences from increasingly powerful AI. The debate is therefore more complicated than “AI companies versus safety researchers.”

Many of the researchers warning about advanced AI risks have themselves worked inside these companies precisely because the companies employ large teams studying alignment and safety. The disagreement is increasingly about whether that work is moving quickly enough.

Why the Google DeepMind Exit Matters

Josh Engels' resignation matters because it comes from someone who was directly involved in studying AGI safety inside one of the world's leading artificial intelligence laboratories.

His warning does not prove that AI will cause immense harm in five years. Nor does it establish that recursive self-improvement or superintelligence is inevitable. What it does show is that some researchers working closest to advanced AI systems believe the gap between capability and safety deserves considerably more attention.

Google DeepMind itself acknowledges that increasingly powerful systems can create risks involving misuse and misalignment. Engels' argument goes further: he believes capabilities may now be advancing quickly enough that independent researchers need more time to understand what could go wrong. That may ultimately be the most important part of his warning. The question is no longer simply how intelligent the next generation of artificial intelligence can become. It is whether researchers can understand and control those systems quickly enough as their capabilities grow.