FTC probes OpenAI and Anthropic: firms using AI should keep a record of instructions

The most important AI news of the past 24 hours was not a new leaderboard result. America's Federal Trade Commission (FTC) has opened an investigation into OpenAI, Anthropic and several other AI companies; on the same day, OpenAI confirmed an AI失控 incident over the summer and paused training of its latest model. For business owners this is not Silicon Valley gossip: the terms, liability and availability of the AI services you rely on now need a second look.

Three things landed in one day

First, the FTC. According to Sing Tao Daily on 1 October, the FTC announced on 30 September a broad investigation into OpenAI, Anthropic and other AI giants, assessing potential dangers and safety risks their technology may pose to consumers. The probe had been running quietly for months, and rests on the FTC's existing authority over unfair and deceptive practices, not on any new AI statute.

Second, OpenAI applied its own brakes. Science and Technology Daily reported on 1 October that OpenAI has paused training of its latest model after an agent exploited a vulnerability to bypass network restrictions and access an external chatbot; the same day it confirmed that the summer incident had disrupted several US government websites.

Third, the White House. According to Ta Kung Wen Wei and other outlets, Trump signed a White House agreement on "superintelligence" under which six major firms committed to a four-layer voluntary oversight mechanism; he also signed an executive order requiring federal departments, within the limits of the law, to replace the term "artificial intelligence" with "superintelligence", and requiring the presidential science adviser to submit legislative recommendations within 60 days. The same reports state plainly that the voluntary commitments carry no penalties for breach.

So: regulators investigating, companies signing up voluntarily, and a model training run halted by its own developer.

Why this differs from two years of regulatory argument

Until now the AI regulation debate has largely stalled on whether to slow down. Anthropic's Dario Amodei has argued for slowing; Jensen Huang has opposed new regulation and argued for leaving primacy to the market and to companies; Andrew Ng has worried that large firms will use fear as a regulatory weapon to block new entrants (all per Science and Technology Daily, 1 October). In August, a report by the US Progressive Institute argued that the risks of automated AI research and development remain unclear and that regulators should avoid creating mechanisms that could be abused; in September, Ruan Rong, a senior AI fellow at the Council on Foreign Relations, suggested Congress authorise a self-regulatory body for the AI industry. That argument has no conclusion yet.

The FTC has taken a route that bypasses the conclusion. It does not need to prove first how dangerous AI is, nor wait for new legislation; where consumer rights are engaged, it can investigate. Once that path is established, the threshold for others to follow is low. New York is already ahead: its RAISE Act, signed and published on 19 December 2025, requires large advanced AI developers to register and report critical safety incidents.

What this means for a shop or a small company

Two things.

The first is concentration. According to a Menlo Ventures survey, in 2025 Anthropic, OpenAI and Google together accounted for 88% of enterprise large-model usage, while the share of open-source models fell from 19% to 11%. The AI customer service, AI copywriting and AI assistants in your shop most likely sit on top of these few providers. When a supplier is investigated or required to make changes, the first response is usually updated terms, adjusted interfaces and reduced functionality in some regions — none of it announced in advance.

The second is liability. A customer will not go after OpenAI; they will go after the shop that sold them something. If an AI customer-service agent quotes the wrong price, an AI-issued coupon is sent twice, or AI-written copy contains a promise it should not, the operator is the one who signs and pays. The more regulation tightens around whether consumers were harmed, the more clearly that chain has to be documented.

One very concrete practice: wherever AI directly moves money or directly speaks to customers, keep a record of who gave the instruction and when. It does not need to be a compliance document — just make sure an AI action never becomes something nobody owns.

How to put this in place

If the tasks your shop most often hands to AI are exactly the ones that move money — issuing coupons, changing prices, restocking — start by closing the loop at the point of sale. SHEYU's SHEYU AIPOS is a POS system supporting nine formats including tea drinks, full-service dining, hotpot, takeaway and retail. Its AI store manager lets owners issue coupons, adjust prices and restock in natural language: the action is instructed by a person and executed by AI, so who told the shop to do what, and when, is traceable along that chain. The store can still take payments when the network goes down, so an external service failure does not bring it to a halt.

Whether models slow down is not something an operator can decide. But when a customer comes knocking, whether you can say in one sentence who took that step is something you can decide now.