The editorial frames the 6(b) study as the same statutory tool used against big tech platforms in 2020 and PBMs in 2022 — a compulsory data-gathering exercise whose real output is a public report plus an internal evidence file for later Section 5 cases. It explicitly points to the 2020 social-media 6(b) as the template: five years of document production feeding ongoing cases against Meta and others.
The editorial argues the timing — two months after the FTC's July AI harms workshop, amid self-harm lawsuits involving minors, and alongside state laws like California SB 243 and Illinois's therapist-impersonation ban — signals the agency is specifically probing the gap between marketed safety work and operational reality. The reported scope (hallucinations, impersonation, minor safety, mental-health interactions) targets exactly the areas where labs have made the loudest public commitments.
By surfacing CNBC's reporting on the FTC inquiry into consumer-facing product risks of OpenAI, Anthropic, and peers, this submitter frames the story around accountability for how these companies measure and mitigate harms rather than around AI capability or competition concerns.
The Federal Trade Commission has opened an inquiry into OpenAI, Anthropic, and other large AI companies over the consumer-facing risks of their products, according to CNBC's September 30 report. The vehicle is a Section 6(b) study — the same statutory tool the FTC used against the big tech platforms in 2020 and against the PBMs in 2022 — which lets the agency compel detailed internal data without first alleging wrongdoing.
The reported scope covers how these companies measure and mitigate product risks: hallucinations, impersonation, minor safety, mental-health interactions, and the downstream effects of chatbots that users treat as advisors. Companies named in the reporting include OpenAI, Anthropic, Google, Meta, xAI, and Character.AI. Snap is also reportedly in scope for its My AI product. The FTC has not published the formal order text, so the exact question list is not yet public, but 6(b) orders typically run to dozens of questions with multi-year lookback windows.
The timing is not accidental. It lands two months after the FTC's July workshop on AI harms, after a cluster of lawsuits alleging that chatbots contributed to self-harm among minors, and in the middle of a broader state-level push — California's SB 243 on companion chatbots, Illinois's therapist-impersonation ban — that has already forced product changes at Character.AI and Replika. The federal probe effectively asks whether the labs' internal safety work matches their public claims.
For developers, the important thing to understand is that a 6(b) is not a lawsuit. It is a discovery mission, and its real output is a public report plus an internal evidence file that the agency's enforcement division can later use as the factual spine of a Section 5 case. The 2020 social-media 6(b) is the clearest template: five years of document production, a staff report in 2024, and the groundwork for the ongoing cases against Meta and others. Expect a similar arc here — quiet for 12 to 18 months, then loud.
The second-order effects start much sooner, and that is where this hits your stack. Legal teams at OpenAI and Anthropic now have to produce internal red-team results, incident logs, model-card drafts, user-complaint tickets, and the actual decision memos behind launch choices. Nothing concentrates a safety team's priorities like a federal subpoena. The predictable response is tighter default refusals, more conservative system prompts, and new logging requirements pushed down into the API contract. Anthropic's usage policies already got stricter in August; OpenAI pushed age-gating changes in September. The probe will accelerate that pattern, not reverse it.
There is a genuine disagreement worth naming. One camp — mostly the labs' own policy staff and a chunk of the AI-safety community — argues the probe is useful: it formalizes the risk accounting that was happening ad hoc and gives regulators the technical literacy they'll need before writing rules. The other camp — some of the open-source and startup side — reads it as regulatory capture in slow motion, where the compliance cost of answering a 6(b) is trivial for a company with Anthropic's legal budget and crushing for anyone trying to ship a frontier model with 40 engineers. Both readings are defensible. The honest answer is probably: the probe is net-good for consumers, net-neutral for the big labs, and net-bad for the next tier of entrants.
The comparison people keep reaching for — pharma, aviation, finance — is wrong in a specific way. Those industries got pre-market approval regimes. The FTC does not have authority to pre-approve AI products; it can only act after the fact under Section 5's unfair-or-deceptive-practices standard, which is why the agency is building a factual record now rather than writing rules. That matters for how you plan: the near-term risk is not a new approval gate, it is enforcement actions against specific product behaviors the FTC can characterize as deceptive — overstated capability claims, missing disclosures, inadequate age controls, misleading safety marketing.
If you ship on the OpenAI or Anthropic API, three concrete things are worth doing in the next quarter. First, audit your own marketing copy against your actual model behavior. The cheapest Section 5 case the FTC can bring is a deceptive-practices action against a downstream product that claimed capabilities the underlying model does not reliably deliver. "Our AI will never hallucinate your medical records" is the kind of sentence that becomes an exhibit. The enforcement risk for application builders is often higher than for the labs themselves, because your marketing claims are more specific and your legal budget is smaller.
Second, assume provider behavior will drift conservative. Build evals that run against each new model version and alert on refusal-rate changes, not just quality changes. Teams that got surprised by GPT-4-turbo's refusal shifts in 2024, or by Claude 3.5's tightened medical responses, learned this the expensive way. A simple nightly eval suite against a frozen prompt set catches drift before your users file support tickets.
Third, get your logging and retention story in order now, not after the labs push a contract amendment. The 6(b) will almost certainly produce new provider-side requirements around traceability — who prompted what, when, with which safety settings. If your application stores prompts and completions in a way you cannot easily query by user, by time, and by model version, you will be rebuilding that infrastructure under deadline in Q1 or Q2.
The near-term signal to watch is not the FTC itself — 6(b) studies are glacial — but the first provider contract change that cites "regulatory requirements" as justification. That is when the probe starts actually costing you engineering hours. The medium-term signal is whether state AGs coordinate; a joint action from California, New York, and Texas would move faster than anything out of Washington and would reshape what disclosures and age controls look like in production. Build for the version of the world where that happens, and the version where it doesn't will be cheap.
If the FTC weren't investigating three of the largest and fastest growing companies in the country for one thing or another, it would be kinda surprising to me.
They just need to hire the right lobbyists, get a meeting at the White House, agree on the terms and call it a settlement.And if the FTC chair is not agreeable, they will be threatened by said lobbyists and eventually be fired.If this sounds far-fetched, it is exactly what happened to Gail Slater, t
Looks like their "Oops! we can't control our dog" PR campaign has backfired.
"Product risks"? The hell? How about cartel like behavior trying to stifle their competition.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
My prediction: nothing will come of this, and these investigations will either be concluded favorably or dropped no less than six months after the midterms.There's only one voice that matters in this administration, and he has already made it quite clear that he both does not consider AI to be