Forecasting Research Institute

Stated Theory of Change

FRI's stated theory is that better forecasting methodology, rigorously applied, will produce more reliable probability estimates of catastrophic risks. These estimates create shared baselines for policy discussions. Decision-makers who start from calibrated probabilities will make better choices about risk mitigation. Over time, tracking how beliefs change produces an "inflation expectations for x-risk" -- a public good that improves collective decision-making.

The specific mechanism is not forecasting-as-prediction but forecasting-as-structured-disagreement: FRI's most important contribution is mapping where experts and superforecasters agree and disagree, identifying the cruxes that drive those disagreements, and publishing this information so others can decide what to do with it.

Revealed Theory of Change

FRI's actions substantially align with their stated mission but reveal some telling divergences:

Alignment. FRI genuinely produces rigorous forecasting research. The XPT, adversarial collaboration, and LEAP are all methodologically serious, transparent about their limitations, and openly published. The 750,000x question-wording effect was discovered and published by FRI, not by critics -- that takes intellectual honesty. They revised their LEAP report when publicly challenged on methodology. The research team (especially Karger) is unusually candid about what forecasting can and cannot do.

Divergence 1: The translational gap is widening, not closing. FRI was founded to make forecasting "useful to decision-makers." Three years and $18M+ later, there's no documented case of a decision being changed by FRI forecasts. Epistemic Audits -- the planned translational product -- appear to have been abandoned or deprioritized. The research output is excellent, but the causal chain from "better forecasts" to "better decisions" remains undemonstrated.

Divergence 2: The methodology keeps revealing its own limitations. FRI's most important findings are about why forecasting doesn't work for x-risk: the 750,000x question-wording effect, the systematic underestimation of AI progress, the near-zero convergence in adversarial collaborations, and the lack of correlation between short-term accuracy and long-term risk views. Each study strengthens the case that x-risk probabilities are unreliable -- which is the opposite of what you'd want from a methodology shop trying to make those probabilities useful.

Divergence 3: FRI is an academic research program, not a policy translation engine. The team is overwhelmingly academic. The CEO comes from GiveWell (evidence-based philanthropy), not policy. The research director's day job is at the Federal Reserve Bank of Chicago. The publications look like academic papers, not policy briefs. The Strategy Fellow role (newly created) explicitly acknowledges the gap.

Key Assumptions

Assumption 1: Quantified probability estimates for existential risk are useful, even if imprecise.

• Evidence for: The XPT created reference numbers that didn't exist before. Researchers (Matt Clancy) use them in formal models. The numbers provide a starting point for structured discussion. Even imprecise numbers reveal the structure of disagreement.
• Evidence against: Normaltech/AI Snake Oil argues the numbers are "feelings dressed up as numbers" that "launder speculation through pseudo-quantification." The 750,000x sensitivity to question wording suggests the numbers aren't measuring a stable quantity. Lina Khan cited 15% as "optimistic" on AI extinction, demonstrating how non-specialists misuse the numbers.
• Testable: Partially. Near-term components can be tested (and have been -- with unflattering results). Long-term components cannot be.
• If wrong: FRI produces an expensive public good that looks like information but isn't. The numbers anchor discussions without informing them.

Assumption 2: Superforecasters add value over domain experts for unprecedented risks.

• Evidence for: Superforecasters have demonstrated accuracy on geopolitical questions. They're more calibrated on base rates and less susceptible to domain-specific biases. Their agreement with experts on non-anthropogenic extinction (asteroid risk etc.) shows calibration where base rates exist.
• Evidence against: On AI specifically, domain experts have been more accurate on near-term benchmarks (24.6% vs. 9.7% average probability on observed outcomes). Superforecasters' reference-class reasoning breaks down when there IS no reference class. Carlsmith reports their arguments weren't persuasive on object-level AI risk questions.
• Testable: Yes, as more XPT/LEAP questions resolve. Early evidence favors domain experts for AI-specific questions.
• If wrong: The core innovation of the superforecasting methodology -- that generalist forecasting skill transfers across domains -- fails in the most important domain.

Assumption 3: Mapping disagreement structure is intrinsically valuable, even if it doesn't produce convergence.

• Evidence for: The adversarial collaboration identified concrete cruxes (METR evaluations, major powers war) and revealed that disagreement is about worldviews (rate of change, trust in theoretical arguments) rather than factual disputes. This is genuinely new information. Knowing that both groups agree powerful AI is coming (88-90%) but disagree on timing (2045 vs. 2450) reframes the debate productively.
• Evidence against: If the mapped disagreements reflect "personality factors" and "non-epistemic" influences (as FRI's own report suggests), then the mapping is sociologically interesting but epistemically uninformative. You've documented a cultural divide, not an intellectual one.
• Testable: Partially, through long-run resolution of cruxes.
• If wrong: FRI becomes a very expensive sociology of knowledge project with limited policy relevance.

Assumption 4: FRI can maintain intellectual independence despite extreme funder concentration.

• Evidence for: FRI publishes findings unflattering to the methodology their funder invested in. The 750,000x effect, the lack of convergence, and the systematic AI underestimation are all findings CG/OP might prefer didn't exist. Karger is candid in interviews.
• Evidence against: 99%+ funder concentration creates structural dependency regardless of individual integrity. Karger advises OP while directing FRI research. CG referred "concerned" participants for the adversarial collaboration. No conflict-of-interest policy exists. The board has no independent members.
• Testable: Would FRI publish a finding that directly threatened CG funding? We don't know because it hasn't happened.
• If wrong: FRI's research is subtly shaped by what's fundable rather than what's most important.

Strengths

1. Intellectual honesty is genuine and unusual. FRI consistently publishes findings that undermine their own methodology's credibility (750,000x effect, lack of convergence, systematic AI underestimation). This is the opposite of the institutional incentive. Most orgs hide their failures; FRI publishes them as research findings. Karger's candor in the 80K Hours interview is remarkable for an org leader.

2. Created the reference dataset for x-risk discussions. Before the XPT, there was no systematic, comparable dataset of expert and superforecaster estimates of catastrophic risks. Now there is. Whatever one thinks of the numbers' reliability, they provide a common reference point. Researchers cite them, policymakers reference them, and the structured disagreement data is genuinely new.

3. Tetlock's credibility provides unmatched institutional legitimacy. FRI benefits from decades of forecasting research credibility. This isn't purchased prestige -- it's earned through pioneering work (Expert Political Judgment, Superforecasting, IARPA tournaments). No other organization could have convened the XPT or LEAP with comparable participant quality.

4. LEAP is the most comprehensive expert panel on AI forecasting ever assembled. 339 experts including top-cited AI scientists, frontier lab employees, leading economists, and policy researchers. With 600K+ words of rationales across first three waves. The longitudinal design (3 years of monthly surveys) will produce uniquely valuable data on how expert views evolve.

5. Research quality is high. Publications in peer-reviewed venues (ICLR 2025 for ForecastBench), XPT book, multiple working papers. The adversarial collaboration methodology is genuinely innovative. The conditional trees method for question generation is novel.

Weaknesses and Risks

1. The translational gap is FRI's existential challenge. Three years and $18M+ with no documented policy impact. The founding promise was to make forecasting useful to decision-makers. Epistemic Audits were abandoned. The Strategy Fellow role is a tacit admission of failure on this front. If LEAP and XPT data don't reach policymakers in actionable form, FRI is an expensive academic exercise.

2. FRI's own findings undermine confidence in its core product. The 750,000x question-wording effect, systematic AI underestimation, and adversarial collaboration non-convergence all suggest that x-risk probability estimates are unreliable. FRI is honest about this, but it creates a paradox: the organization best positioned to produce x-risk forecasts has produced the strongest evidence that such forecasts can't be trusted.

3. Extreme funder concentration (99%+ from one source) is a governance failure. Open Phil/CG could terminate FRI by declining to renew grants. There's no institutional resilience. The Karger-OP advisory relationship further blurs the boundary. The board has no independent members. No conflict-of-interest policy exists. For an organization receiving $18M+, this governance structure is inadequate.

4. Superforecasters' track record on AI is bad and getting worse. XPT superforecasters substantially underestimated AI progress on every measured benchmark. Epoch AI found they gave only 9.7% probability to observed outcomes. If superforecasters are systematically biased on the most important variable (AI progress), the entire methodology is compromised for the most important application.

5. The academic orientation limits impact. FRI's team, publications, and culture are academic. This produces rigorous research but limits engagement with the policymakers and industry leaders who could actually use it. The intelligence community's failure to adopt forecasting (which Karger himself cites) is a cautionary precedent.

Cross-References

Complementary to Metaculus: Metaculus is the platform (crowd predictions on resolvable questions). FRI is the methodology shop (how to elicit, aggregate, and interpret forecasts, especially for unprecedented events). They share Open Phil as primary funder and forecasting as core domain. FRI's Project Improbable (low-probability elicitation) could directly improve Metaculus's accuracy.

Complementary to Epoch AI: Epoch AI forecasts AI progress via compute trends (data-driven). FRI forecasts via expert elicitation (judgment-driven). These are fundamentally different approaches that should be compared: when compute trend extrapolation and expert judgment disagree, which is more reliable? Neither org has pursued this comparison. Epoch's finding that XPT forecasters underestimated AI progress validates Epoch's data-driven approach over FRI's judgment-driven one, at least for near-term capability questions.

Relationship to Good Judgment Inc.: Tetlock co-founded both. GJI provides superforecasters that FRI uses. FRI research validates the methodology GJI commercializes. GJI has pushed back on ForecastBench. No institutional firewall exists.

Funded by the same source as many AI safety orgs: CG/OP funds FRI, Metaculus, Epoch, METR, and many others. This creates correlated risk: a shift in CG priorities could affect the entire ecosystem simultaneously.

What Would Change This Assessment

• Documented policy impact: If a government or major institution published evidence that FRI forecasts changed a consequential decision, that would significantly strengthen the theory of change.
• Resolution of LEAP/XPT cruxes: As questions resolve by 2030, we'll learn whether superforecasters or domain experts were more accurate on AI progress. If superforecasters dramatically outperform, that vindicates the methodology. If domain experts outperform (early evidence suggests this), FRI's reliance on superforecasters needs rethinking.
• Revenue diversification: If FRI secured government contracts, lab partnerships, or diversified philanthropic support, the governance concern would diminish.
• Epistemic Audits becoming real: If FRI actually delivered organizational decision tools (not just research papers), the translational gap would close.
• A forecasting failure with consequences: If a policy decision explicitly based on FRI-type forecasts led to a bad outcome, it would reveal whether the methodology does harm as well as whether it does good.

Self-Critique

What sources should I have checked but didn't?

• The full 800-page XPT technical report. I read the summaries and the Karger interview but not the full document. Key methodological details may be buried there.
• The later sections of both Tetlock 80K Hours interviews (Oct 2025). I read summaries but didn't read them in their entirety. These likely contain updated thinking on FRI's direction.
• Titotal's detailed post "Explaining a subtle but important error in the LEAP survey" (EA Forum, Dec 2025) -- referenced but not fetched due to forum blocking.

Where is this analysis potentially biased?

• I may be too sympathetic to FRI's intellectual honesty, treating "we published our own failures" as sufficient evidence of rigor when it might be a strategy to preempt criticism.
• I may overweight the Normaltech critique because it's analytically elegant. There may be strong defenses of x-risk forecasting that I haven't engaged with.
• I may underweight the diffuse value of creating reference forecasts. Like GDP statistics, the value may be in how thousands of people use the numbers, not in any single traceable decision.

What would a thoughtful person who disagrees say? "You're demanding documented policy impact that no methodology organization can show. Climate science took decades to influence policy. Tetlock's first generation of forecasting research (Expert Political Judgment, 2005) influenced the intelligence community over 15+ years. FRI is 3 years old. The XPT numbers are cited everywhere -- in academic papers, EA strategy documents, policy briefs, and the IAISR 2026. You can't trace these citations to specific decisions because that's not how information influences policy. It diffuses. FRI's job is to produce the best possible information and trust the system. The fact that they publish findings that undermine their own methodology shows they're doing real science, not advocacy -- and real science is exactly what this field needs."

What's my single weakest claim? The claim that FRI has no documented policy impact. "No documented" is doing a lot of work. It's entirely possible that IAISR 2026 authors, Open Phil grantmakers, and government officials in various countries regularly consult FRI data and it meaningfully shapes their decisions. The absence of a paper trail doesn't mean the absence of influence. This may be a documentation gap, not an impact gap.

What information would most change my view? A candid account from a government policymaker or institutional decision-maker saying: "We used FRI's XPT/LEAP data to calibrate our risk assessment, and it changed what we decided to do about X." One verified instance of this would substantially change my assessment of the theory of change from "broken at the decision link" to "working but undersold."