Competence requirements for high-impact AI systems

The regulatory gap

Most current and proposed oversight of artificial intelligence regulates artifacts: the model, the dataset, the disclosure, the impact assessment. Very little of it regulates people. This is unusual. In every other field where technical judgment protects the public — structural engineering, medicine, aviation maintenance, electrical installation — the regulated unit is a qualified person who is accountable for a decision, and the artifact requirements exist to support that person's judgment.

Artifact-only regimes have a predictable failure mode. They generate documents that satisfy the letter of the requirement and are authored by whoever has capacity, rather than by whoever is competent. The Society's position is that oversight of high-impact AI should name a competent accountable professional, and that the profession, not the legislature, should define and maintain the competence standard.

“A duty without authority transfers blame downward and improves nothing. Employers must give the accountable professional the standing to say no.”

What competence should mean

The Society defines competence for high-impact AI across four domains. Technical evaluation covers the methods used to characterize system behavior, including their limits. Systems reasoning covers the socio-technical context: who uses the output, under what pressure, with what recourse. Assurance argumentation covers the ability to construct and defend a claim about fitness for deployment. Professional ethics covers the duties owed to the public, the employer, the profession and the individuals affected.

Crucially, competence must be demonstrated on real work under real constraints. Examination alone is insufficient, because the failures that matter arise from judgment under commercial pressure, not from ignorance of technique.

• Technical evaluation, including the limits of the chosen methods.
• Socio-technical systems reasoning about use, pressure and recourse.
• Assurance argumentation that survives adversarial review.
• Professional ethics with duties that can override an employer instruction.

Recommendations

First, any regime governing high-impact AI should require a named, credentialed accountable professional for each deployed system, on the model of the responsible engineer in other disciplines. Second, competence standards should be maintained by an accredited professional body, subject to public consultation, so they can track practice faster than statute can. Third, registration should be revocable, because a credential with no consequence is a marketing asset rather than a public protection.

Fourth, and least discussed: employers should be required to give the accountable professional the authority to say no. A duty without authority transfers blame downward and improves nothing.

What ACS commits to

The Society will maintain the competence framework in the open, publish assessment criteria in full, operate a public register, and investigate complaints against registrants through an independent process. We will also publish annual data on assessment outcomes so that the standard's rigor can be scrutinized rather than assumed.