Daily AI brief · October 6, 2026
AI Labs Can't Put a Number on the Worst Case
Anthropic, OpenAI, Google and Meta testified under oath to New York City's council but could not size their worst-case risk, ChatGPT will show ads while it makes images, and founders hedge across models.
AI companies testify under oath in New York and can't size the worst case (CNBC, City & State New York, amNY, CBS New York) Yesterday's hearing went ahead with all 51 council members. Anthropic sent Logan Graham, who leads its Frontier Red Team; OpenAI sent policy lead Morgan Dwyer; Google and Meta sent their AI policy directors. When Speaker Julie Menin asked each of them to put a number on the risk in a worst-case scenario, none did, and she called that "troubling at best." A former Anthropic employee testifying as a whistleblower warned that the industry may be building "our own adversary," and Elon Musk's SpaceXAI did not show up despite a subpoena. City & State's verdict was that the companies failed to impress. For a small business, the useful takeaway is narrower than the headlines: if the builders can't describe their worst case in plain terms, don't expect your vendors to, so the job of deciding what an AI tool is allowed to touch in your business stays with you.
ChatGPT will show ads while it makes your images (The Decoder, Android Headlines, Yahoo Finance) OpenAI says it will start testing visual ads in the US later this month, shown while ChatGPT generates an image, with an initial group of advertisers. The ads are labeled, kept separate from the finished picture, and only appear on free accounts; Plus, Pro and business plans don't get them. OpenAI also said ChatGPT now reaches 1.2 billion people a week. If your staff use free personal ChatGPT accounts for work images, this is one more reason to move them onto a paid business plan: no ads, and your company controls the account instead of an employee's personal login. It also tells you where marketing is heading, because a chatbot that sells ad space will soon be a place your customers see your competitors' products.
Founders are hedging across AI models instead of picking one (TechCrunch) Ahead of its Disrupt conference, TechCrunch looked at how startups decide between open models they can run themselves and closed models from the big labs, and found some building products that use several models rather than committing to one. The practical part: the companies whose whole business depends on AI are treating the model as a part they can swap, and a small business can borrow that habit cheaply by keeping its prompts, instructions and examples in its own documents instead of locked inside one vendor's tool.
Move work images off free accounts. If anyone on the team makes flyers, social posts or product mockups in a free ChatGPT account, have your AI champion list who does it and move that work onto the company's paid plan, or onto whichever image tool you already pay for. It takes an afternoon and closes off both the ads and the personal-account problem.
Write down what each AI tool may touch. Take five minutes with the team and add one line per tool to your staff AI rules: what data it can see, what it can change, and what it must never do on its own (send email, pay a bill, delete files). Yesterday's hearing showed that nobody else is going to draw those lines for you.
Be wary of "AI risk scores" from sellers. After a hearing where the biggest labs couldn't quantify their own worst case, treat any vendor that hands you a neat risk number for its AI product with suspicion. Ask instead what the tool can access, what it logs, and how you turn it off.
Where it's going: Lawmakers are now questioning AI companies under oath, and the answers they get will shape rules aimed at the businesses that deploy AI, so the businesses with written rules and a clear list of what their tools can do will have the least to change.