Submission 59
Attune Legal
Submitter group: Companies.
An AI model identified the submitter group and wrote the summary, key points and recommendation summaries. Counts do not measure importance or merit.
Summary
Read the original submission for the full text.
Argues that AI models have no motivation of their own, so risk sits with the people who set goals and act on outputs, and obligations should attach at the input (configuration) and output (verification, irreversible action). Says existing law reaches much of this, questions reliance on human oversight, and recommends configuration retention, effect-based disclosure and prohibiting self-motivated AI.
Key points
Quotes are the submitter’s own words. Text taken from documents, including scanned pages read by text-recognition software, can differ from the PDF. Check the original. Listed in the same order as on the submission’s page in the interactive guide.
Argues risk lies primarily with people directing systems or acting on outputs, rather than within the model.
It is also sobering, because it removes the comfortable idea that risk is contained inside the machine where a technical fix or a better probe will eventually reach it. The risk is not primarily in the box. It sits with the people who point the system and the people who act on what it returns.
Argues obligations should attach at the input, where direction enters, and the output, where consequences land.
Because direction enters at the input and consequences land at the output, these are the two points at which obligations should attach. They call for different things.
Argues that the presence of a human in the loop is not itself evidence of oversight and may produce accountability without control.
Taken on its own, the presence of a human is not evidence of oversight, and treating it as though it were will produce accountability without control.
Argues that for agentic systems the most consequential safeguard is a human check before irreversible actions.
For agentic systems the single most consequential safeguard is a constraint on autonomous, irreversible action: the more irreversible the effect, the more a human check must be interposed before it occurs, rather than offered as a review after the fact.
Says existing technology-neutral law reaches much of the issue and that the gap lies in guidance not locating duties at the human control points.
Existing, technology-neutral law reaches much of this, and the National AI Plan is right that privacy, consumer, anti- discrimination and sector-specific regulation apply to AI. The gap is not missing statutory power. It is that guidance, and any forthcoming standard, has not located the duties at the two points where a human controls the system
Argues that deliberate development of AI designed to generate and pursue its own goals should be prohibited and treated as categorically different from current agentic systems.
The deliberate development of AI architected to generate and pursue its own goals, independent of a human-supplied objective, should be prohibited, and treated as categorically different from the agentic systems in use today
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Frame AI standards, guidance and the AI Safety Institute’s frontier-risk work around human-set goals, with input duties for specification and configuration and output duties for verification and controlling irreversible action.
Frame regulation and guidance around the human control points, not machine intent. Any AI standard, and the AI Safety Institute's treatment of frontier risk, should treat an agentic system as the competent pursuit of a human-set goal, and locate obligations at the input (specification and configuration) and the output (verification and control of irreversible action).
Prohibit the deliberate development of systems architected to generate and pursue their own goals, independent of a human-supplied objective.
The deliberate development of systems architected to generate and pursue their own goals, independent of a human-supplied objective, should be prohibited.
Require configuration retention and auditability instead of prospective mapping; deployers whose agentic systems affect people must retain operative instructions and configurations so they can answer for the system’s actions.
Require configuration retention and auditability for agentic systems, as the practical record-keeping mechanism, in place of prospective system mapping. An entity deploying an agentic system in a way that affects people should retain the instruction artefacts and configuration under which it operated, so it can answer afterwards for what the system did.
Make deployers responsible for adequate instructions and limits when agentic tasks affect people; disclaimers must not transfer that responsibility to individual users.
The second is that the burden of specification should not be discharged by a disclaimer. Where a deployer sets an agentic system loose on a task that affects people, the adequacy of the instruction, and of the limits placed on the system's freedom to act, is the deployer's responsibility, not the individual user's.
Place native versioning and retention of configuration artefacts on platform providers, not users, where a platform rewrites its own instruction files in response to prompts.
Place native versioning of configuration artefacts on platform providers. Where a platform rewrites its own instruction files in response to prompts, the obligation to version and retain them should sit with the provider, not the user.
Require output verification proportionate to reversibility, with a human check interposed before irreversible action rather than as after-the-fact review.
Require output verification proportionate to reversibility, and interpose a human check before irreversible action. The more irreversible the effect of an action, the more a human control must sit before it, not as an after-the-fact review.
Adopt quality-assurance re-review for decision-shaping systems: a defined proportion of cases re-reviewed independently without seeing the automated output, with systematic divergence triggering review of the system's configuration.
Adopt quality-assurance re-review for decision-shaping systems. A defined proportion of cases should be re-reviewed by a person who approaches the underlying record independently, without first seeing the automated output, with systematic divergence triggering a review of the system's configuration, not merely the individual case.
State in guidance that a human in the loop is not, on its own, evidence of oversight, and adopt the submission's functional factors as the test for genuine oversight.
State in guidance that a human in the loop is not, on its own, evidence of oversight, and adopt the functional factors set out above as the test for whether oversight is genuine.
Test oversight by whether outputs precede independent judgement, agreement rates, and the cognitive load and time pressure that can make oversight nominal.
The first is whether the output reaches the person before or after they have formed an independent view, because a pre-formed output anchors a decision while a post- hoc one checks it. The second is the proportion of cases in which the person simply agrees with the system. The third is the cognitive load and time pressure they work under, which turn oversight nominal at volume.
Test oversight by disagreement costs, upstream pipeline effects and whether people know if their views came from records or automated summaries; presume AI shaped decisions when they cannot say.
The fourth is the cost of disagreeing: where overriding the system requires the person to document reasons, escalate or accept personal exposure, the system is functionally determinative even where it is formally advisory. The fifth is the system's position in the pipeline, since intake, triage and prioritisation upstream determine which matters reach the decision-maker and on what framing. The sixth is whether, afterwards, the person can even say whether their view was formed on the record or on an automated summary of it. Where they cannot, the system should be presumed to have shaped the decision.
Require disclosure that describes a system's effect rather than its architecture, consistent with the ADM obligation commencing 10 December 2026, and carry it into the federal-agency ADM framework.
Require disclosure that describes effect, not architecture, consistent with the automated decision-making obligation commencing on 10 December 2026, and carry the same approach into the federal-agency automated decision-making framework.
Require disclosure of the kinds of decisions AI shapes and the information it uses.
Disclosure should describe a system's effect, the kinds of decisions it shapes and the information it uses, rather than its architecture.
Avoid personal accountability for oversight staff lacking time and information to audit outputs; treat that gap as the deploying entity’s governance failure.
Protect the people placed in oversight roles. A framework should not create personal accountability for staff who lack the time and information to audit the output they are asked to ratify; that gap is a governance failure of the deploying entity, and should be treated as one.