Comments
ALTER’s work continues to focus primarily on AI policy, and especially on standards, evaluations, and related governance questions. We have also continued a smaller amount of research and policy work in adjacent areas, including biosecurity, AIxBio, philosophy of AI, and public health.
ALTER is currently funded tentatively through the end of 2028, pending finalizing new grants to cover our base budget. This primarily includes support from cG and SFF Speculation matching grants. This does not fill out room for funding; we have more ambitious plans for expanding our staff to do additional work in AIxBio policy and in international AI policy and standards.
Vanessa Kosoy’s work with CoRAL is supported separately through approximately the end of 2027, via a combination of AISI / ARIA and Open Philanthropy funding, including her salary via ALTER, and her research team.
AI policy and standards remain ALTER’s primary focus. As a recent example of our likely impact, LLMs estimated a 40% probability that Demis Hassabis’s recent institutional proposal (boosted by Sam Altman) was substantively informed, directly or indirectly, by our 2025 publication, The Necessity of AI Audit Standards Boards.
Our policy work in the past 6 months has increasingly focused on standards, where slow and bureaucratic processes can still be unusually important because they help define the terms that governments, companies, and courts later rely on when asking whether organizations have met expected practice.
David attended the semi-annual ISO/IEC JTC 1/SC 42 meeting in Singapore in mid-April. The draft standard on Human Oversight of AI Systems, ISO/IEC 42105, is now in FDIS, meaning that it is close to finalization. This is a slow process, and the standard is not a binding regulatory instrument, but we continue to think it is valuable to have clear language explaining that human oversight is not a trivial requirement, and cannot be satisfied merely by assigning a human to be notionally “in the loop.”
We have also begun work on new projects in JWG 2, the joint working group on AI evaluations, verification, and validation, partially led by James Gealy of Safer AI. This work is still at an early stage, but it is a natural continuation of our broader push to make AI evaluation more rigorous, more explicit about limits, and more useful for governance. This also builds on out AI Evaluation Standards project, below.
In addition, we have contributed to ISO/IEC TS 25568, on addressing risks in generative AI systems, with a focus on ensuring that loss of control, bioterrorism, and other global catastrophic risks are explicitly included in the set of risks to consider. We have also contributed to ISO/IEC TS 25570, on reliability, especially around the need to be clear about the limits of reliability testing.
There have also been preliminary international discussions about possible biorisk standards for cloud labs. These discussions are not yet a concrete project for ALTER, but we continue to think the topic is important, and expect that standards in this area could become valuable as cloud labs and AI-enabled biological work become more capable and widely accessible.
The AI evaluation standards project, which David was seconded to for a large portion of his time, has now been publicly launched as Evals-Consensus.AI. The project is a loose alliance of participants and organizations working to clarify what the AI evaluation community agrees on, where disagreement remains, and what practices are sufficiently well-supported to recommend for adoption.
The Delphi process has been completed, and the results paper is now being written. We expect this to be useful not as a final answer to what evaluations should look like, but as a way of making expert disagreement and agreement more legible to policymakers, standards bodies, developers, and researchers.
Related work has also continued on evaluation reporting. David contributed to a non-archival EvalEval workshop paper at ACL presenting the systematic review, and to an Evaluation Cards paper which incorporates the systematic review, as well as a platform for those cards that compiles and presents data on evaluations across models. He is also contributing to other papers being prepared for and with EvalEvals.
David presented our “oversight versus control” paper at AAAI 2026. This work continues to be useful for standards and policy discussions, because it gives a clearer vocabulary for distinguishing between systems that can actually be controlled or meaningfully overseen, and systems where human oversight is mostly nominal. David also attended IASEAI 2026, where he discussed several possible collaborations and participated in the AI Oversight workshop.
The Agents of Chaos paper, led by Natalie Shapiro, has also been released, with David as one of the authors. We are considering possible follow-up work, including more technical work on multi-agent monitoring with William Waites, using category theory to formalize agent actions and interactions. John Baez has a related post here. There is also additional conceptual follow-on work about context windows and vulnerabilities. We are also considering a more governance-oriented follow-up on oversight and monitoring of multi-agent systems, and joint work on other aspects of multi-agent safety.
The systematic review for evals-consensus was accepted at a non-archival ACL workshop, and the Evaluation Cards paper, which incorporated the systematic review, is currently under review.
David has also recently written a paper and accompanying explainer on AI Economics, exploring the implications of LLMs being primarily compute-limited, and what that means for profit margins of AI labs versus hyperscalers
Our work with RAND has now ended. We are grateful for the collaboration, and expect the work to continue informing our thinking on AIxBio and related policy questions.
The winner of the AI Safety Prize has been decided, and we are planning the announcement at the Israel AI conference. We are pleased to have been able to help create an incentive for AI safety research done in Israel, and hope that in coming years the prize will continue to help make this work more visible to both the Israeli AI community, and the broader policy and research ecosystem.
David presented at the Ministry of Health Council for Regulation of Research into Biological Disease Agents annual seminar, “Microbiology: Innovation, Technologies and Vaccines,” on the topic “When Should We Worry About AI Being Used for Nefarious Purposes, and What Can We Do About It?”
This led to useful conversations about possible updates to Israeli biorisk laws and regulations. Unfortunately, much of this is currently on hold because of the war, and the path forward is uncertain. We nevertheless continue to think that Israeli biosecurity regulation will need to adapt to AI-enabled biological risks, and that there may be useful opportunities to assist when the relevant government attention returns.
We have also finally published our paper on a ThreatNet-like system in Israel, which would serve as a useful pilot for such systems in other countries due to Israel’s combination of advanced healthcare, expertise in computational genomics, and small size.
As mentioned in our previous update, David also presented at a RAND / ALTER side event on AIxBio at the Biological Weapons Convention meeting on December 10th. This work remains part of our broader effort to make AI-enabled biological risk concrete and decision-relevant for policymakers.
David’s talk on philosophy and AI at the Technion was delayed, but took place at the end of June, immediately after the same work was presented at 5ICEAI in Zagreb the week before. It was warmly received, and a paper is being written, based in large part on the argument outlined here.
Our work on salt iodization in Israel has continued to move slowly, despite our nominal success. The regulations are supposedly being finalized, and are currently in the regulatory review process to be published for public comment soon. The extended delay is frustrating, but progress is currently limited by the broader political and governmental situation, and the glacial pace of regulatory changes.
Vanessa Kosoy and the CoRAL research team have continued making progress on their research agenda. This work is now substantively separate from ALTER, though ALTER continues to support it administratively to some extent. More detail is available from the (end of 2025) Vanessa / CoRAL research update; further updates will come from CoRAL.
Finally, we have re-done our web site to be easier to update, with the help of an LLM agent. It’s no longer a complex wordpress site, and it handles Hebrew / English versions more smoothly, as well as allowing easier updates of the site. (The visual aspects of the site are mostly unchanged, but feedback is welcome.)
