Charter

The committee exists to keep the safety of advanced AI and the material cost of building it in the same conversation. Both are commonly treated as separate specialisms, discussed by different people in different rooms, and the result is that progress on one is regularly used to excuse the other.

Its purpose is to produce work that a person outside the committee can check: measurement methods, benchmark results, written standards, and a public record of what was said to policymakers.

Catalyst. A system is acting as a catalyst when it reduces the effort required to reach an outcome that a person has chosen, while leaving the choice, the judgement and the responsibility with that person. A system that selects the outcome is not acting as a catalyst, regardless of how it is described.

Human decision maker. A person is the decision maker when they can reach a different conclusion than the system without penalty, have the information required to do so, and have the time to use it. A person who can only approve or reject a recommendation is a reviewer, not a decision maker, and the committee treats the two as different.

Material cost. Electricity, fresh water, land, and hardware, measured at the site where they are consumed rather than netted off against purchases made elsewhere.

The committee commits to publishing its methods in full, including the parts that make its own results look worse.

It commits to publishing negative results from its efficiency research. A benchmark that fails to show the effect we expected is a finding, and withholding it would make every other number the committee reports worth less.

It commits to publishing the substance of what it says to governments and regulators within a reasonable period, so that its policy work can be examined by the people it affects.

It commits to stating what it does not know. Where reliable figures do not exist, the committee will say so rather than produce an estimate that will later be quoted as a measurement.

The committee will not endorse systems designed to remove a human decision maker from decisions that materially affect a person's liberty, health, livelihood or safety.

It will not accept membership or funding conditioned on a finding, on the suppression of a finding, or on advance approval of what it publishes.

It will not list an institution, adviser or endorsement that has not agreed in writing.

It will not describe a target as an achievement. Where this document or the committee's other work states an intention, it will be marked as an intention.

The committee's first output is a reporting method for the material cost of training and inference. It is intended to be adopted by laboratories other than ours, which means it has to be practical to apply and hard to satisfy selectively.

The method will specify: energy per training run and per million inferences, measured at the facility; fresh water withdrawn and consumed at the site, reported separately, because they are not the same figure and are frequently conflated; the grid carbon intensity at the time of use rather than an annual average, since when a load runs changes what it costs; and the boundary of what is counted, stated explicitly, so that two reports can be compared.

Where a laboratory cannot report a figure, the method requires it to say which figure is missing and why, rather than omitting the line.

Members serve as individuals. An institution may support a member's participation, but the member's positions are their own and the committee does not claim institutional endorsement from an individual's membership.

The committee is convened by the Eunoia Foundation and its work is not subject to approval by SHV Groups. A committee assembled by a laboratory to examine the laboratory's own field carries an obvious problem, and the only workable answer is that its outputs are checkable by people who do not work for us.

Members may publish dissent. Where the committee issues a finding that a member disagrees with, the disagreement is published alongside it and attributed.

Members declare commercial interests in the systems, facilities or policies under discussion, and declarations are published with the membership list.

A declared interest does not disqualify a member. Much of the relevant expertise sits inside the organisations whose work is being examined, and excluding it would leave the committee talking to itself. An undeclared interest does disqualify.

Outputs are published openly and without registration. Methods, data and the code used to produce results are released together, because a method published without the means to reproduce it is a claim rather than a contribution.

Where material cannot be released for reasons of safety or third party confidentiality, the committee states what is withheld and on what grounds.

This charter is reviewed annually and whenever a member formally requests it. Amendments are published with the reasoning, and previous versions remain available.

The committee will state each year what it said it would do, what it did, and what it did not get to. A body that reports only its completed work is reporting selectively.

Version 1.0, in effect from publication. The committee is newly formed. It has no completed outputs, no published measurement standard, and no member list, because it has not yet finished assembling one.

Everything in this document describes what the committee is for and what it has bound itself to. None of it describes what it has achieved.