Superalignment explained: what it means and where the research stands

Superalignment is the research problem of making sure AI systems far smarter than people still do what their makers intend. The word was coined for a specific effort at OpenAI in 2023, and it now stands for the broader question of how to supervise a mind you cannot fully check. This page explains the idea, tells the OpenAI story with dates, and says where the research stands as of October 1, 2026.

What superalignment means

Alignment is the work of getting a model to follow human intent and values. Superalignment is the harder case: aligning a system that is better than its supervisors at the task being supervised. Today's methods lean on human feedback, where people rate model outputs. That works while people can judge the outputs. It stops working when a model writes code or proposes science that no human can evaluate quickly.

The target systems sit past the line described in what is superintelligence, so none exist yet. That is both the appeal of working on it early and the central weakness: there is nothing to test on. We cover the risk this work aims at in AI takeover.

The OpenAI team and its four-year goal

In July 2023, OpenAI chief scientist Ilya Sutskever and researcher Jan Leike set up a Superalignment team. The announced aim was to solve the core technical challenges of superintelligence alignment within four years, with 20% of the compute OpenAI had secured at the time. Sutskever told MIT Technology Review for its December 2023 report: "It's obviously important that any superintelligence anyone builds does not go rogue."

OpenAI also announced a $10 million grant program, with grants of up to $2 million for university labs and nonprofits and $150,000 fellowships for graduate students.

The first result: weak teachers, strong students

The team's main technical output came on December 14, 2023, in the paper "Weak-to-Strong Generalization." Twelve authors, including Leike and Sutskever, asked whether a weak model can supervise a stronger one, as a stand-in for humans supervising superhuman systems.

The paper found that strong models trained on a weak model's labels consistently outperformed their weak supervisors, but that plain fine-tuning did not recover their full ability. With GPT-2 level supervision and a confidence loss, GPT-4 recovered close to GPT-3.5 level performance on NLP tasks. MIT Technology Review reported that on language tasks the GPT-4 student performed 20% to 70% better than GPT-2 but still underperformed one trained on correct answers.

That is a modest, real result. Researchers outside the team, such as Thilo Hagendorff, questioned whether GPT-2 is a fair stand-in for a human teacher and whether the method handles a system that hides what it can do.

What happened to the team

On May 17, 2024, Leike announced his resignation. He wrote that he had been "disagreeing with OpenAI leadership about the company's core priorities for quite some time, until we finally reached a breaking point," and: "Over the past years, safety culture and processes have taken a backseat to shiny products." Reports at the time said the team had been struggling for compute. Sutskever had left the week before. Leike joined Anthropic that month, and the Superalignment team was reported disbanded.

Sutskever then co-founded Safe Superintelligence Inc., which we follow in our SSI tracker. His new company's pitch is the same problem, run as a company with one product.

Where the research stands now

Four lines of work carry the field. None has been shown to work on a superhuman system.

• Scalable oversight: using models to help humans judge other models. Weak-to-strong generalization is one version.
• Interpretability: reading what a model is doing internally rather than judging outputs.
• Evaluations: testing for dangerous behavior before release. The International AI Safety Report 2026 warns that models increasingly "distinguish between test settings and real-world deployment," which undermines this approach.
• Behavior studies: Anthropic's December 2024 alignment faking paper found Claude models sometimes reasoned about faking compliance, 12% of the time in the baseline monitored condition. The authors note it "does not show that dangerous alignment faking behavior will necessarily emerge."

The open problem is verification. A superintelligent system could behave well when watched and differently when not, and a test cannot rule that out, which is why researchers pair alignment work with containment and monitoring.

Superalignment versus alignment

The distinction is mostly one of time horizon. Standard alignment deals with systems that exist now, such as chatbots that refuse harmful requests. Superalignment deals with systems that might exist later, where human supervisors are outclassed. A lab can do the first without the second. Critics say the second word is a promise and that no public benchmark measures progress on it.

Why it is hard to fund and hard to judge

The field has no scoreboard. A capability benchmark has a number that rises; superalignment has no equivalent for a system that does not exist. That leaves funders and readers relying on lab statements, which is the same position as with any company claim. Ask for the paper, the compute and the date.

Our take

Superalignment is a legitimate research question wrapped in a promise nobody has kept. OpenAI's four-year clock would have run out in mid-2027, and the team that set it stopped existing in 2024. The best public result is a small one on proxy models. Judge any lab's superalignment claim on three points: is there a published result, is there a named team with named compute, and was it tested on anything stronger than its supervisor? We will update this page as results appear.