Sampura Research emphasizes that human judgment remains the true safety valve for evaluating the behavior of hyper-models and bridging critical gaps in automated oversight systems.
Article Index:
- Founding of Sampura Research for intellectual safety
- An alternative to the methodology of AI monitoring itself
- Building the infrastructure for alignment and harm evaluation
- The context of the safety crisis and brain drain from laboratories
- Frequently asked questions
Founding of Sampura Research for intellectual safety
A group of senior former researchers at Google DeepMind has launched a new nonprofit organization headquartered in London called “Sampura Research,” aiming to dedicate scientific efforts to keeping humans at the core of supervision and oversight systems for ultra-advanced artificial intelligence models. The organization, officially announced on August 24, seeks to build and develop what it calls certified “judges”—hybrid systems combining human expertise and artificial intelligence to evaluate model safety and behavioral discipline according to designated objectives.
The organization was founded through a joint initiative by Rishub Jain, Joshua Jacob, and Alex Adams, all of whom previously worked at Google DeepMind. Jain spent seven years at the lab, focusing for two years on scalable oversight research and contributing to pioneering projects such as AlphaFold, while Jacob led human data engineering at DeepMind and previously worked at self-driving vehicle company Waymo.
An alternative to the methodology of AI monitoring itself
The organization presents itself as a direct intellectual and scientific alternative to the current prevailing approach in the tech industry, which relies on using artificial intelligence models to monitor and evaluate other artificial intelligence models. While this method offers extreme speed, Jain notes that it leaves critical gaps that can only be bridged by balanced human insight and judgment. According to Bloomberg reports, Jain believes humans hold an indispensable role in verifying model behavior even as their reasoning capabilities grow.
Building the infrastructure for alignment and harm evaluation
Sampura’s research agenda focuses on developing standardized datasets that accurately measure the proficiency of humans and AI systems in verifying model outputs, and using those metrics to improve oversight methods. The organization defines a “judge” as “a human, artificial intelligence, or hybrid system that evaluates the correctness and alignment of model behavior during conversations and agent workflows.”>
Planned workflows include building forward-looking datasets for future harms and risks, innovating methodologies for distributing sub-tasks between humans and machines, and developing an open-source human evaluation platform. The organization has begun hiring at least six seasoned AI researchers in London with annual salaries ranging from £100,000 to £290,000.
The context of the safety crisis and brain drain from laboratories
The organization’s launch coincides with growing concern over the capacity to monitor advanced models. A Reuters report noted that artificial intelligence companies “cannot yet contain what they have built,” while the independent nonprofit organization METR—which tests models from OpenAI, Anthropic, Google, and Meta—faces severe difficulties in recruiting talent despite offering salaries exceeding $500,000.
More than 1,100 researchers and practitioners signed an open letter in late July calling for support for international mechanisms to slow down the development of extreme models when necessary. Jain’s departure was part of a major talent exodus from DeepMind in June 2026, during which the lab lost five of its senior researchers within just six days.
Frequently asked questions
Question: What is Sampura Research and what is its goal?
Answer: It is a nonprofit research organization in London founded by former DeepMind researchers to ensure the continuation of human oversight and arbitration for intelligent models.
Question: Why does the organization reject total reliance on AI to monitor itself?
Answer: Because purely automated monitoring involves gaps and alignment risks that can only be accurately detected and addressed through direct human insight and evaluation.
Question: What are the most prominent research projects Sampura is working on?
Answer: They include building future harm datasets, methodologies for dividing oversight tasks between humans and machines, and an open-source human evaluation platform.
