OpenAI dissolves Superalignment after Jan Leike's departure
OpenAI has dissolved the team created to research how to control a future superintelligence. The decision comes as its two leaders, Ilya Sutskever and Jan Leike, leave the company, with Leike questioning OpenAI's safety priorities.
On May 17, 2024, Jan Leike explained his departure from OpenAI and said the Superalignment team had struggled to secure compute; the unit ceased to exist as a separate group. His public statement gives the former co-lead’s account, while OpenAI’s original plan provides the promised objectives and resources for comparison. A resignation evidences disagreement, not a complete company audit.
The team's disappearance does not mean the company is abandoning safety research altogether, but it does eliminate a division with its own mandate, dedicated leadership and a singular promise: to devote a significant share of its computing capacity to solving the problem of aligning a hypothetical superintelligence with human interests.
Two departures in three days
Leike shared leadership of Superalignment with Ilya Sutskever, OpenAI's co-founder and, until now, chief scientist. Sutskever announced his departure last Tuesday, May 14, without explaining his reasons in detail at the time. His role will be taken over by Jakub Pachocki, previously the company's head of research.
Leike's departure came with a much more explicit message. The researcher wrote that he had disagreed with OpenAI's leadership for quite some time over the company's fundamental priorities, until those disagreements reached a breaking point.
Leike wrote that, in recent years, the company's safety culture and processes had taken a back seat to flashy products. He also argued that OpenAI should devote a much larger share of its resources to the safety challenges posed by advanced systems.
OpenAI CEO Sam Altman responded on the social network X that he appreciated Leike's contributions and that the company still had a great deal of safety work ahead of it. He did not explain how the functions handled by Superalignment would be reorganized.
What Superalignment was
OpenAI introduced the team in July 2023, when Sutskever and Leike announced a four-year research plan. The premise was that AI models could one day surpass human capabilities in most intellectual tasks, and that current methods for controlling them would prove insufficient.
In simple terms, aligning an AI means getting it to pursue its intended goals and respect human limits even in unfamiliar situations. That is different from fixing a specific bug in ChatGPT or preventing a harmful response: it involves designing oversight mechanisms for systems that, in theory, could outperform their supervisors at many tasks.
The group promised to use about 20% of OpenAI's computing capacity over four years for this work. Its research areas included training models to help evaluate more powerful models and finding ways to detect deceptive or dangerous behavior before it caused harm. Primary source.
The unit's very existence carried symbolic weight. OpenAI was founded in 2015 with safety at the center of its mission, but since ChatGPT launched in late 2022, it has also become a company competing to launch products, attract customers and maintain its position against Google, Anthropic and other technology companies.
Safety integrated, but without its own team
Members of Superalignment will be redistributed among other OpenAI research teams. The company still has groups dedicated to safety, model evaluations and risk preparedness, so the dissolution does not mean those activities are going away.
However, it changes how they are organized. An independent team can focus on long-term problems that do not yet generate revenue or immediately improve a product. By integrating it into other divisions, OpenAI will have to show that the research retains resources, autonomy and the ability to challenge launch decisions.
The timing deepens the questions because the move comes six months after the board crisis that led to Altman's temporary ouster and subsequent reinstatement. Sutskever initially took part in that decision, although he later expressed regret and backed the CEO's return.
For users and businesses, the news does not change access to ChatGPT or OpenAI's models today. Its significance lies elsewhere: the safety of the most capable models depends not only on public statements, but also on who has internal authority, budget and the power to stop a launch. Sutskever and Leike's departures leave open the question of how OpenAI plans to preserve that function while accelerating the commercial development of its systems.
A team is measured through operating capacity
A unit’s name states its mission; budget, compute, people, authority and artefacts show whether it can perform it. Evaluate Superalignment by comparing the published plan with effective access to resources, results and release decisions. Leike’s account identifies a discrepancy the company can answer with evidence.
A compute reservation needs a denominator and calendar. Theoretical capacity, allocated hours and used hours are different. A percentage without project records may look large and remain unavailable when needed. An audit asks for a series rather than one commitment.
Safety should not depend on one person
Departures matter, but the programme also lives in owners, escalation processes and criteria that stop deployment. If it disappears with its leaders, it was not institutionalised. If it continues elsewhere, ownership, objectives and results should be visible without the old label.
Research lines are judged through reproducible experiments: oversight using weaker models, detection of deceptive behaviour and generalisation of control. Each method needs a threat, metric and failure condition. “Align superintelligence” is a mission, not a measure.
Priority conflict is not resolved by counting releases. Follow resources, deadlines, published evaluations, incidents and veto power. Preserve each party’s account too: internal testimony adds information but does not reveal every decision by itself.
The transferable skill is to audit a safety commitment through objective, real resources, authority, outputs and continuity. The framework distinguishes reorganisation from loss of function without turning a resignation into rumour or a corporate promise into proof.
Reorganisation needs a continuity table
For every plan line, record the new owner, resources, next publication and review date. If a line moves into another team, the table shows continuity; if nobody owns it, it documents loss. Renaming without transferring responsibility would be continuity only in rhetoric.
Negative results matter. A method that fails to scale or an evaluation the model learns to game can close a false route if published. Measuring safety only through successes rewards hiding the most valuable evidence.
External oversight can review processes without model secrets: team composition, compute use, incidents, release criteria and treatment of objections. Operational transparency does not require publishing dangerous capabilities.
Readers should preserve dates. The original plan and later criticism belong to different moments; a future response cannot erase the discrepancy retroactively. Chronology is part of evidence.
The company can answer through allocations, publications and verifiable decisions. If only release language changes, the discrepancy remains open. That symmetry—attributed testimony and required corporate evidence—supports investigation without choosing a side by sympathy.
A useful conclusion does not predict whether OpenAI will be safe. It leaves a method for checking continuity and evidence of loss: no owner, unmeasured resources, unpublished evaluations or removed veto power.
This article was produced with artificial intelligence under human editorial oversight.
