AIMM’s public evaluation framework: the criteria by which we judge whether an external theory, tool, technology, or organization satisfies our strategy. Published so a recommendation is never a quiet commercial endorsement: anyone can see what we require, what we considered, and why an approach did or didn’t meet the bar.
Criteria
-
Preserves or expands human agency. Looking for approaches that demonstrate a viable way to work with AI on human-centered terms, not only warnings about what to avoid.
-
Practically adoptable by real people. An approach aimed only at policymakers, regulators, or frontier labs doesn’t meet this bar on its own. Priority goes to what individuals, families, businesses, and local governments can actually put into practice.
-
Affirmative and constructive. A rigorous critique of a harm is valuable but insufficient by itself; we look for what someone can do differently as a result.
-
Evaluable and transparent on its own terms. The approach should be auditable against its own claims, not a black box we’d be asking people to trust on faith.
-
Compatible with non-ownership. We can recommend it without owning, controlling, or being the sole distributor of it. A proposer’s financial or commercial stake is disclosed before evaluation, not after.
-
Advances a named mission area. Advances at least one of the three mission areas (philosophical/ ethical foundation, practical AI strategy, or alternative computing research, see mission.yaml) explicitly, not just gestured at.
Process
-
Anyone can propose an approach for evaluation.
-
Conflict-of-interest disclosure is required first, from both the submitter and whoever at AIMM evaluates it. An evaluator with a disclosed conflict recuses.
-
The evaluation is recorded against these criteria specifically, not criteria invented ad hoc for one submission.
-
The record of what was considered, and why it was or wasn’t recommended, is public wherever the recommendation itself is published.
-
Evaluations are dated, not permanent. An approach that doesn’t meet the bar today can be resubmitted later, because the approach improved, a new version shipped, or our own criteria evolved. Every evaluation is its own dated record; a later evaluation of the same approach doesn’t erase or replace an earlier one, it supersedes it for current-recommendation purposes while the full history stays visible.
We don’t claim to have built the answer ourselves. Where a program produces AIMM’s own strategy documents, templates, or evaluation criteria (see programs.yaml), we own that. Where a submission implements part of that strategy (an open-source project, an academic theory, a commercial service, independent research, a future computing architecture), it stays outside AIMM, evaluated and recommended, not acquired.