The Alignment Agenda

A research roadmap for making advanced AI systems evaluable, interpretable, and controllable.

Summary

Alignment is not one problem. It is a portfolio of four bets. The question "is this system aligned?" has no single test, because alignment is not a single property. This volume reframes the field as four compounding bets (evaluation science, interpretability, scalable oversight, and control) each attacking a different root difficulty, none sufficient alone. The task of the decade is not to pick the winning bet. It is to industrialize all four, and to know at every moment which layer of the defense is thinnest.

Alignment is often described as a single unsolved problem. This volume treats it as a portfolio of four bets (evaluation science, interpretability, scalable oversight, and control) that only pay reliably when held together, as defense in depth. For each bet the report assesses what exists, what it cannot yet do, and what maturity would look like, then proposes a funding and institutional program to make assurance compound industrially. It closes with twelve open problems that would move the field if solved this decade.

Key findings

1. Alignment is a portfolio, and portfolios need weights Specification, generalization, and oversight are distinct root difficulties; no one method addresses all three. The four bets fail in different ways under different futures, which is exactly what makes holding them together rational.
2. Evaluation is the most mature bet, and still lower-bounds Every evaluation is a floor, not a ceiling: a model's failure to show a capability mostly reflects the questioner's elicitation. Turning floors into defensible ceilings is the field's first open problem.
3. Interpretability has instruments, not yet coverage Features, circuits, and production monitors are real and useful. But nobody can state what fraction of a model's safety-relevant computation they explain, and audits without denominators are testimony, not evidence.‐
4. Scalable oversight is the hinge of the decade When systems exceed their checkers, trustworthy judgment must come from protocol structure (critique, debate, decomposition, weak-to-strong generalization) rather than judge acuity. Whether that works at frontier scale is genuinely undetermined.
5. Control pays out exactly when the other bets fail Designing deployments to be safe even if the model is misaligned (interruptibility, authority budgets, tripwires, control evaluations) is the cheapest insurance in the portfolio and the least fashionable work in the field.
6. The binding constraint is industrial, not conceptual Ideas outnumber people, access, and career paths. Making the four bets compound at the scale of what they must assure, Volume 01's test, is chiefly a problem of funding, institutions, and the independent bench.

How to cite this report

The Alignment Agenda: A research roadmap for making advanced AI systems evaluable, interpretable, and controllable. FAIR Labs Frontier Series, Vol. 05. Fair Artificial Intelligence Research Labs, 2026. Available at https://fairlabs.ai/research/frontier-vol05-the-alignment-agenda

Topics: alignment, interpretability, evaluations, scalable oversight, AI control, research agenda, open problems

FAIR Labs (Fair Artificial Intelligence Research Labs) is a nonpartisan 501(c)(3) research institute in Washington, DC. Its reports may be reproduced with attribution for non-commercial purposes. For interviews, briefings or data requests, use the contact form.

Related FAIR Labs reports