What is Sampura Research, the London-based nonprofit launched on August 24, 2026 by former Google DeepMind researchers Rishub Jain, Joshua Jacob, and Alex Adams, why does it position itself as a human-centered alternative to the industry’s “AI supervising AI” approach, and how wi

Sampura Research is a London AI-safety nonprofit founded by former Google DeepMind researchers Rishub Jain and Josh Jacob—not, based on its own announcement, by Alex Adams. It launched publicly on August 25, 2026, with an $11 million commitment from Coefficient Giving; Adams is credited for feedback on the announcement. 4

• Its central claim: current “scalable oversight” often replaces human judgment with LLM judges because they are cheaper and faster. Sampura argues that humans will retain useful, distinct strengths: catching AI’s uneven (“jagged”) failures, providing an independent check against model–judge collusion, and deliberating over genuinely new or evolving human values. 4

• What “hybrid judges” mean: a judge is a human, AI, or combined system that assesses whether an AI response or agent trajectory is correct and aligned. Sampura plans methods that estimate an AI judge’s confidence and escalate uncertain cases to people; use routers to allocate tasks or subtasks to humans versus AI; and give people task-specific AI assistance. The goal is not human review of everything, but targeted human involvement where it adds the most safety value. 4

• Benchmarks, datasets, and ratings: its initial six-month programme is a public, continuously updated leaderboard with more than 20 sub-datasets, including deception detection, cultural bias, and unsafe agent actions. It intends to test judges not only on static examples but also when the supervised model can optimize against them during reinforcement learning—an important test for reward hacking. Sampura says it will publicly release papers, code, datasets, and leaderboards. 4

• Forward-looking harms: the announced scope includes harm beyond classic “rogue AI” scenarios—such as bias, unsafe interactions with children, and companion-AI dependence—and later expands from judging outputs to producing and evaluating the underlying instructions, rubrics, and test cases. That can improve anticipation of failures, but it is a research programme, not demonstrated containment capability. 4

• Deployment test: Sampura proposes measuring whether better judges actually reduce reward hacking, determining when to retrain them as models adapt, and optimizing the cost–speed–quality trade-off of human input. It may first stress-test deployment with third-party evaluators, then aim for use in frontier-lab training, evaluations, and monitoring. 4

• Why the staffing problem matters: METR’s published evaluation-execution role lists $285,548–$503,116 in annual base pay and says it may go higher for exceptional candidates, illustrating how scarce experienced AI-evaluation talent is. 9 Sampura’s plan to recruit London technical staff is therefore strategically relevant, but I found insufficient primary-source evidence for the specific claim that it will hire “at least six” people at £100,000–£290,000; its primary announcement merely says it is hiring founding technical staff in London. 4

• Relation to calls for a slowdown mechanism: more than 1,100 frontier-AI practitioners signed a July 2026 letter seeking international capacity to slow development if it outpaces safe human supervision. 1 Sampura supplies a possible technical prerequisite—more credible, auditable evaluations—but cannot itself create legal authority, international coordination, model access, or enforcement. Its work therefore complements rather than substitutes for governance.

• DeepMind departures and the “containment” framing: Jain’s exit was reported amid a broader succession of DeepMind departures, though the “five core researchers in six days” formulation appears to come from lower-quality secondary reporting and should not be treated as independently established. 5 Likewise, the evidence available here does not establish the sweeping proposition that AI companies “cannot yet contain what they have built.” The defensible conclusion is narrower: robust oversight of increasingly capable models remains unproven, evaluation talent is scarce, and Sampura is attempting to make human input more scalable and resilient rather than assuming AI can safely supervise AI alone.