PII Comment from Shepard, Matt

Re: AI Policy Statement; Matter No. P264200 I am a licensed general contractor and a fourth-generation tradesman. Over the last two years I taught myself to build and deploy AI systems, and I now run a consultancy helping blue-collar business owners put those tools to work. I write from the middle of two worlds the Commission is trying to bridge: the everyday operator who relies on these systems, and the builder who understands what they actually are. I offer this comment in support of the core principle in the proposed statement, with a few suggestions I believe would make it more durable. I support the Commission's central theory. A provider that secretly steers a system's outputs toward an objective the user did not ask for and would not expect, while marketing that system as accurate and best-effort, is deceiving its customers. That is a real harm, and the right thing for Section 5 to reach. The wrong here is not error, and not the system holding a point of view. The wrong is concealment: passing off a steered output as neutral, best-effort assistance. Keeping that distinction sharp is what makes the rule defensible. First, I encourage the Commission to keep a bright line between hidden objectives and ordinary mistakes. The statement already does this by carving out hallucinations, and that carve-out is essential. Modern AI systems are not deterministic, rule-based software that returns a fixed answer. They are reasoning systems that produce judgments, and like any reasoning process, human or machine, they will sometimes be wrong. A regime that treated ordinary error, or reasonable design tradeoffs between accuracy, brevity, and clarity, as if it were deception would be unworkable and chill exactly the innovation the Administration seeks to protect. The target should stay on deliberate, undisclosed steering, not on the fact that a reasoning system is fallible. Second, I want to name what I believe is the root of the consumer harm the statement identifies: the public does not understand the distinction I just drew. In my work I regularly watch capable business owners treat an AI output as a settled fact rather than a considered judgment. The statement's own figure, that users accept outputs without further checking the large majority of the time, is alarming precisely because the public's mental model is wrong. They think they are querying a database when they are consulting an advisor. That gap is why hidden steering is so damaging, and why disclosure, which the statement already identifies as the path to compliance, is the right instrument. I would go one step further and suggest a concrete, low-burden mechanism. Some providers already append fixed, deterministic notices to responses on sensitive topics, for example prompting users to verify election information or surfacing a crisis line on questions about self-harm. These notices do not touch the model's reasoning at all. They are fixed, auditable, and triggered by topic, not generated by the model. A general expectation that systems carry a plain notice on high-stakes topics, urging users to verify against independent sources, would help reset the public's mental model at the point of use, cheaply and without requiring anyone to inspect a model's internal objectives. To be clear, this is not a cure for a bad actor who secretly trains a model to mislead. That is what the anti-concealment rule is for. Disclosure and the rule work together: the rule handles deliberate manipulation, and the notice closes the everyday expectation gap for everyone else. The Commission might also consider whether a public-education effort on how these systems differ from conventional software would serve the same goal. Third, offered in the spirit of strengthening the rule rather than opposing it, I encourage the Commission to apply and illustrate this standard with strict neutrality. If the wrong is undisclosed manipulation, it is wrong in every direction. A provider that secretly skews its outputs to favor one political party is committing the same deception as one that skews toward the other. The illustrative examples in the proposed statement lean heavily in a single direction. A standard meant to protect truth in the marketplace will be more durable, and less vulnerable to a challenge that it discriminates by viewpoint, if its examples show it biting manipulation of any political stripe with equal force. Finally, on structure: I believe a single national standard is the right destination, and a patchwork of conflicting state rules serves no one well. I note only that a policy statement asserting the reach of Section 5 is a first step. Lasting national uniformity will likely require a clearer federal standard from Congress or through rulemaking, rather than resting on enforcement theory alone. Thank you for the opportunity to comment. Matt Shepard Shepard Consulting shepardconsulting.ai