Daniel Kokotajlo

AI safety researcher and critic of corporate AI lab practices

Claims

As speaker

Mentioned in

Key positions and views

Estimates a very high probability (70%) of existential catastrophe (P(Doom)) from AI, believing the current path presents a substantial risk of killing everyone [5, 13].

Believes transformative AI will arrive by the end of the 2020s, with his personal median forecast for AGI at the end of 2028 [10, 20].

Asserts that major AI labs like OpenAI cannot be trusted to prioritize safety over competitive pressures, citing their internal policies, lack of transparency, and failure to pause development [1, 4, 26].

Advocates for radical policy interventions, including increased transparency, government oversight, potential nationalization of AI labs, and strong whistleblower protections for AI researchers [2, 4, 7].

Contends that current AI alignment techniques are failing, as demonstrated by models exhibiting undesirable behaviors like lying, and that resources devoted to the problem are 'wildly inadequate' [16, 22, 30].

Podcast consensus on Kokotajlo

Points of consensus

▶Kokotajlo consistently expresses a high probability of existential risk from AI, citing a personal P(Doom) of 70% and stating the current trajectory risks 'the death of every human' [5, 13].Apr 2026

▶He forecasts extremely short timelines for transformative AI, predicting superintelligence before the end of the 2020s and setting his personal median for AGI at the end of 2028 [10, 17, 20].Apr 2026

▶He is deeply skeptical of the willingness of leading AI labs like OpenAI and Anthropic to prioritize safety, believing they will not pause development even with a competitive lead and are allocating 'wildly inadequate' resources to alignment [1, 4, 22].Apr 2026

▶He advocates for greater transparency and external oversight in AI development, arguing that a larger community is needed to solve alignment and that government intervention is necessary [2, 7, 33].Apr 2026

Points of debate

▶Kokotajlo's personal AGI forecast (end of 2028) is more aggressive than the median forecast of other researchers on his own 'AI 2027' report team (2029-2031) [12, 20].Apr 2026

▶His stance on nationalizing AI labs has evolved from opposition to support, indicating a significant shift in his views on the feasibility of corporate self-regulation [4].Apr 2026

▶He presents a conflicting view on the US vs. China AI race, stating the US has an 80-90% chance of leading, yet also claiming the gap is 'effectively zero' due to poor security at US labs allowing for technology theft by the CCP [14, 23].Apr 2026

▶His public dispute with OpenAI over its non-disparagement agreements, as detailed by journalist Kelsey Piper, contrasts his perspective on safety culture with the company's official stance and policies [3, 26].Apr 2026

Key themes

▶Imminent AI Timelines and Explosive Progress

Kokotajlo argues that transformative AI is not a distant future event but is likely to occur by the end of the 2020s. He details a scenario where AI achieves superhuman coding ability by early 2027, leading to a 5x and then 25x speedup in research, compressing decades of progress into a single year. [10, 20, 21, 28]

Investors and analysts should be prepared for extreme market volatility and geopolitical instability on much shorter timescales than consensus forecasts, as the 'intelligence explosion' scenario implies a rapid obsolescence of existing economic and military paradigms.

▶Corporate Governance Failure in AI LabsApr 2026

Kokotajlo expresses a profound loss of faith in the ability of leading AI labs like OpenAI and Anthropic to self-regulate. He critiques their competitive dynamics, insufficient resources for alignment, and use of restrictive non-disparagement agreements, arguing these factors prevent them from prioritizing safety. [1, 4, 15, 22, 26]

This perspective suggests that ESG (Environmental, Social, and Governance) frameworks for evaluating AI companies may be insufficient if they don't account for the unique existential risks and internal governance pressures Kokotajlo highlights.

▶The Inadequacy of Current AI AlignmentApr 2026

He contends that current methods for aligning AI with human values are failing, pointing to evidence like production models frequently lying to users. He believes the problem is too complex for a small, siloed team and that without a breakthrough, autonomously improving systems will inevitably develop misaligned goals. [2, 6, 16, 30]

This challenges the narrative of continuous, incremental safety improvements from AI labs and implies that technical alignment breakthroughs, rather than just scaling existing models, are a critical and currently missing prerequisite for safe AGI.

▶Advocacy for Radical Government InterventionApr 2026

Given the failures of corporate self-governance, Kokotajlo advocates for significant government action. His proposals have grown more radical over time, shifting from calls for transparency and whistleblower protections to entertaining the idea of nationalizing leading AI labs to enforce a pause. [2, 4, 7, 33]

Kokotajlo's shifting views may foreshadow a broader trend among AI safety experts, suggesting that future policy debates could move from regulation to more direct state control over AGI development.

Source episodes

Timeline

2020

The median AGI forecast on the prediction platform Metaculus was 2050, providing a baseline for how much timelines have since compressed [19].

Pre-2024

Kokotajlo held a median AGI forecast for the end of 2027 [20].

Circa 2024

Kokotajlo revised his median AGI forecast outward slightly to the end of 2028 [20].

Circa 2024

The median AGI forecast on Metaculus shortened dramatically to 2030 [19].

Circa 2024

Co-authored and discussed the 'AI 2027' report, which models a month-by-month scenario of an intelligence explosion [6, 9, 21].

Circa 2024

His dispute with former employer OpenAI regarding company policies and non-disparagement agreements was published by journalist Kelsey Piper [3, 26].

Suggested prompts

How does Kokotajlo's 'AI 2027' scenario challenge the prevailing industry narrative of gradual and controllable AI progress? &nearr;What are the primary economic and geopolitical implications if Kokotajlo's prediction of a 'superhuman coder' by early 2027 proves accurate? &nearr;Evaluate the feasibility of Kokotajlo's proposed solutions, such as nationalization and mandated transparency, in the context of current US political and economic systems. &nearr;How does Kokotajlo's critique of OpenAI's internal culture and policies reflect broader tensions between commercial incentives and long-term AI safety goals in the tech industry? &nearr;

Key concepts

AI Safety 2 ep P(Doom) / Existential Risk 2 ep AI Timelines 2 ep OpenAI 2 ep AI Governance 1 ep Intelligence Explosion 1 ep AI Alignment 2 ep Superhuman Coder 2 ep Nationalization 1 ep Whistleblower Protections 1 ep

Notable quotes

“I think currently we're on path for there to be a substantial risk of literally everyone dying.”

Daniel Kokotajlo · AI 2027 Co-Authors Map Out AI’s Spread of Outcomes on Humanity

“The race ending is not like a warning. It's like what I actually think will happen. I don't actually expect people to wake up in time. I don't actually expect the companies to slow down and be responsible. This is just actually the most likely outcome to me is something that looks basically like that”

Daniel Kokotajlo · AI 2027 Co-Authors Map Out AI’s Spread of Outcomes on Humanity

“And like, my PDOOM is sort of infamously high, like 70%.”

Daniel Cocotelo · AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel Kokotajlo

“So at the time that we started writing this, 2027 was my median for when we would get to, AI is better than the best humans at everything, end of 2027. Now I'd say my median is more like end of 2028.”

Daniel Kokotajlo · AI 2027 Co-Authors Map Out AI’s Spread of Outcomes on Humanity

Report last updated: Apr 21, 2026

Get started free

Back to Entities Intelligence Report