About this Report
Historical forecasts of AI progress typically predicted the timeline to human-level AI (“AGI”) to be measured in decades or even centuries. In the largest survey of AI researchers’ timeline predictions (Grace et al., 2025), completed in October 2023, the point at which there would be a greater than even chance of AI outperforming humans in every possible task was estimated to occur in 2047.
This is in sharp contrast to statements made by several frontier AI company leaders throughout 2024 and early 2025. In April 2024, xAI CEO Elon Musk predicted that AI would be “smarter than the smartest human” before the end of 2026 (Reuters, 2024). Anthropic CEO Dario Amodei said in November 2024 that the rate of capability increase is such that a rough estimate suggests “we’ll get there by 2026 or 2027” (Fridman, 2024). Amodei emphasized a high degree of uncertainty in this prediction, but also said that we are “rapidly running out of truly convincing blockers, truly compelling reasons why this will not happen in the next few years.” Shortly thereafter, in January 2025, OpenAI CEO Sam Altman said that he thinks AGI will probably be developed before the end of Trump’s second term (Pillay, 2025).
Although many researchers disagree that AI progress will advance this rapidly (Nellis, 2024), the fact that several leaders of the companies that are at the forefront of AI development predict such time horizons ought to be sufficient to warrant substantial preparation, considering the magnitude of the impact if they are right. However, no comprehensive analysis of what AI risk mitigation interventions might be viable under these timeline assumptions appears to exist.
This report addresses that gap. It began as a private planning aid, developed into guidance for a workshop with senior AI experts, and an early version eventually circulated within the community of AI safety practitioners. The interest in this report was substantial enough that we ultimately decided to write it up in a form suitable for public release. It is for this reason that this study, which focuses on viable interventions under conditions of great urgency, is being released months after its initiation. Nonetheless, we are pleased that it has finally been published today for a broader audience.
How To Use This Report
This report serves as both a decision aid and an inspirational resource. Its intended primary audience is individuals and organizations seeking to take actions for reducing AI risk themselves. Secondary audiences include policy staff and advisors who prepare options for policymakers, as well as philanthropic funders who shape their grantmaking strategy.
We recommend that all three audiences follow a similar reading path:
1. Begin with the executive summary, which explains the scope and purpose of the report, summarizes the evidence base, and outlines the main conclusions. We highlighted a set of illustrative projects – one within each domain which can serve as useful starting points. Time-constrained readers may stop here, though it should be noted that a portfolio of this kind cannot be easily summarized, and readers surveying only this section will derive limited benefit from this work.
2. Consult the full list of the 208 initiatives. This list should be treated as a menu rather than a narrative – we recommend skimming by category and delving deeper into items more relevant to your remit and interests. For many initiatives, external resources exist and we encourage readers to seek out further information.
3. Read the section on broad strategic considerations for systemic factors affecting the viability of the initiatives in the portfolio.
Readers interested in methodology, participant demographics, or the questions used, may consult the methodology and disclosures section, and the appendices containing the participant demographics and the questionnaire, respectively.
Methodology
Interviews
We pooled 279 potential interview candidates. ~70 individuals were selected to represent a cross-section of the AI ecosystem, including funders, AI labs, non-profits, governments, academia, independent researchers and entrepreneurs – and supplemented by expert referrals and targeted outreach to dissenting views. This ensured that each participant added complementary expertise to the conversation. Participant demographics are presented in the appendix.
The interviewers took notes as well as recordings of the conversation.
48 individuals participated in a 45-60 minute call. Interviewer 1 conducted 70% of the interviews, Interviewer 2 20%, and Interviewer 3 10%. Participants were not required to prepare for the interview. Participants also volunteered five proposals after their respective interviews, and these have been processed as well.
The interviews followed a semi-structured format built around six thematic clusters: the current AI risk landscape, challenges within it, concrete projects, execution power, funding, and infrastructure for the field.
The prompts were deliberately openended and allowed interviewers to go deeper flexibly, which resulted in almost every interview skipping over some of the themes and questions. We prioritized getting a better picture of the interviewee’s opinions on near-term AI risks over following strict standardized protocol. Early interviews were used to determine the best questions and interview flow for later ones. The interview questionnaire we used as a guide is provided in the appendix.
The spontaneous nature of live conversation likely biases the suggestions towards novelty and away from ideas they may endorse given more time to think. The ideas listed should be read as “opinions and projects treated as worthy of serious consideration by several interviewed experts.”
To maximize candor, the interviews were conducted under the Chatham House Rule, and as a result the ideas are not attributed. We thank all collaborators for their contributions.
A list of all 208 concrete project ideas was distilled and grouped into 8 distinct categories (Technical AI Alignment Research, Evaluation & Auditing Systems, Intelligence Gathering & Monitoring, AI Governance & Policy Development, International Coordination, Preparedness & Response, Public Communication & Awareness, and Miscellaneous). Novel ideas which came up only once were placed in an appendix.
Interviews were also processed into broader strategic considerations from concrete project proposals mentioned by at least 10% (5/48) interviewees, grouped into four top-level themes (“readiness”, “coordination across the ecosystem”, “standards and common interfaces”, and “capacity constraints”).
Respondent Demographics
Interviewees include key staff at OpenAI, Anthropic, Google DeepMind, Mila, AMD, the EU AI Office, and multiple AI Safety Institutes, METR, RAND, Scale AI, GovAI, Transluce, and ARIA.
The geographic distribution is dominated by the United States (54%), with the UK and Germany accounting for a further 23%.
The gender gap is significant with 85% male representation (41 individuals) compared to 15% female representation (7 individuals).
Seniority is concentrated at the senior level, with 46% of participants having 11-15 years of general professional experience.
Most participants (~85%) had at least 3 years of AI experience.
“Independent” and “public sector” tie as the most common professional backgrounds (17% each), followed closely by “Non-AI lab corporate” and “Non-academic research institutions” (15% each).
Fig.1: Career experience.
Bar total heights are general experience; stacked bars are experience in AI.
Fig.2: Count of respondents by the main sector of their career.
Fig.3: Count of respondents by nationality
Fig.4: Treemap of areas, depicting the frequency with which they were mentioned in all interviews
Authors & Acknowledgements
Maximilian Schons, MD, focuses on safety and assurance work at the intersection of biotech and AI. He has held senior positions in national research consortia, worked as Chief Medical Officer for life-science startups, and recently published the State of Brain Emulation Report 2025.
Samuel Härgestam is a former technology entrepreneur whose AI security work has mainly focused on mobilizing capital for the field through investments, advisory work, and targeted risk communication. He serves on the boards of the Astralis Foundation and the Effective Institutions Project, and contributes to the work of LawZero.
Gavin Leech, PhD, is a co-founder of the research consultancy Arb and a fellow at Cosmos and Foresight. He was previously head of research at the Dwarkesh Podcast. He runs the annual Shallow Review of Technical AI Safety.
Raymund Bermejo specializes in operations and project management for AI security organizations. He co-founded HIRe, a recruiting firm, and directed Anti Entropy, a consulting firm – serving organizations in the field.
Further contributions to this report come from:
- Dr Sören Mindermann, Mila-Quebec AI Institute
- Philip Harrison, Arb Research
- Stag Lynn, Arb Research
- Rory Švarc, Arb Research