One safety story a week, pressure-tested for what's actually happening underneath the surface.

By Lindsay Gross

On July 1, the Federal Trade Commission (FTC), the agency that goes after false advertising, scams, and hidden fees, proposed a policy statement about AI. A policy statement is not a law. It is the agency saying how it reads the rules and how it plans to enforce them. It carries out a White House executive order directing the FTC to treat "steering" a model's outputs as possible deception. Chairman Andrew Ferguson's FTC voted 2 to 0 to open it for public comment.

The FTC's argument is that AI companies sell their tools as giving you the best, most accurate answer they can, so a reasonable person expects a truthful one. If a company secretly shapes its model to steer you toward an agenda you never signed up for, the space between the honest answer you expected and the steered one you got can be deceptive under Section 5 of the FTC Act, the same law behind false advertising. The FTC leans on Anthropic's own finding that people accept AI answers without checking them 91% of the time. When trust runs that high, a quiet nudge does real work.

Here is the hard part for anyone who builds. Every model is already shaped. After training, companies decide how it behaves through human feedback, refusals, and a system prompt, and almost all of that is ordinary design, not a conspiracy. The FTC is careful. It exempts honest mistakes, and it says a company can disclose its priorities up front, as long as it does not bury that in the terms of service. But it doesn’t draw the line between making a model safe and "distorting" it, and that line is the whole game. It also calls Colorado's AI Act a state law that federal law can override, which sets up a fight over who gets to regulate AI.

Comments are open until July 31 at docket FTC-2026-0859. Watch whether "accuracy" ever gets a definition you could test a model against. This is a rare week your comment lands on the same docket as OpenAI's and Anthropic's, so if you build on these models, use it.

A LessWrong post by darshanav makes an argument that should bother anyone shipping on top of "passed our safety evals." The evals themselves have never been validated. Nobody knows their detection thresholds. The standout evidence, each item cited to its source:

The proposed fix borrows from how engineers validate any measuring instrument: seed models with known, controlled misalignment and measure how much the evals catch, then map detection rates against severity. Until someone does that, a passing eval tells you the model passed the eval, and not much else.

sysls, who ran systematic investment processes at large hedge funds and is now building OpenForage, argues that since roughly the GPT 5.5 generation, agents can correctly implement just about anything you ask. So any failure of creation is a failure to ask. What separates the outliers, in his framing:

His closing line is the piece in one sentence: "The question of the future is not whether you can build it, but whether you know what you actually want to build."

Our partners and sponsors get exclusive placements across the newsletter and access to AIC's in-person network — demo nights, dinners, hackathons, and forums across 180+ chapters.

For all inquiries, send us a note at [email protected].

The AI Collective is built by volunteers across 180+ chapters in 40 countries.

Thank you to the thousands of volunteers around the world who make this work possible. We truly could not do this without you.