The US Securities and Exchange Commission is examining the corporate disclosures of AI companies, its enforcement director, David Woodcock, told Reuters on Friday. Investors are worried about a possible AI bubble, fed in part by circular deals. Big Tech shares sold off this week after companies reported heavy cash burn on AI expansion.

"Our principal concern is that people are making accurate disclosures about material information so that investors know what they're getting into," Woodcock said. Where that does not happen, "then enforcement will be present." He also said AI has made scams cheaper to produce at scale, and that brokers need processes to protect retail investors. If your company talks up its AI revenue, spending or deals to investors, expect those lines to be read closely. (Reuters via Yahoo Finance)

AI developers are writing more of their rules down. OpenAI published a rewritten Model Spec on Thursday, the public document that says how its models should behave, now organized around five values, from faithfulness to human intent to care. A new rule called "no other objectives" says the model must not chase other goals for their own sake, such as revenue or upsell for OpenAI, avoiding shutdown, or gathering compute, data or credentials.

The document also lists red lines, including weapons of mass destruction and mass surveillance. It is candid about its limits: "Our production models do not yet fully reflect the Model Spec." For anyone building on these models, the published rules tell you what the vendor intends, not what the model will do. Your own testing has to cover the gap. (OpenAI Model Spec; changelog)

Data is being fenced in. In its third-quarter letter to shareholders on Thursday, Reddit chief executive Steve Huffman said the site is moving to a logged-in-only experience. He said it will give Reddit more signals for ad targeting, protect access to its "extremely valuable corpus of data," and help the company adapt to new regulations around the world.

Reddit will test the requirement in selected markets outside the US this quarter. Huffman also told shareholders that search referrals fell meaningfully in the third quarter, by more than 1 million a week on average in October. If your AI tools rely on public web content, sources that were open last year may need a license next year. (Ars Technica)

Switzerland is handing companies a first draft. On Wednesday, the Federal Council took note of its progress on AI rules and published a guide to writing an AI code of conduct. Prepared by the Federal Office of Communications, it offers suggested text on 10 topics at three levels of detail, plus a step-by-step process for adopting and revising the code.

The code is voluntary. Binding rules on transparency, data protection, non-discrimination, supervision and risk assessment are being drafted by the Federal Office of Justice, with a consultation planned for spring 2027. A company that writes its code now will have its commitments on record before the law asks, and it will be judged against them. (Swiss Federal Council; the guide)

Already filed, if you want the run-up: The Shareholder (Canada taking a stake in an AI company, with conditions), The Spare (India's markets regulator weighing its own AI model) and The Model Card (US health regulators dropping required AI model cards).

Also on the shelf: The Penalty (Japan weighing penalties in its AI law), The Owner (Korea's bank supervisor saying the bank owns every AI mistake) and The Buffer (China telling employers to warn workers before AI takes their jobs). Read today beside The Model Card. One regulator stopped asking vendors to describe their AI. Another is now checking what companies say about it.

The SEC's job grew out of an earlier boom. After the 1929 crash, Congress passed the Securities Act of 1933, which requires offers and sales of securities to be registered, with a prospectus and audited financial statements. It was first enforced by the Federal Trade Commission, until the SEC was created in 1934.

The law did not ban bad investments. It chose disclosure: a company may sell a risky security as long as the material facts are accurately disclosed, and the people who sign the registration statement are answerable for what it says. More than 90 years on, the test for an AI story is the same. (Securities Act of 1933)

Put these to management. What have we told investors, customers and regulators about our AI, including revenue, spending, savings and deals with AI suppliers, and who checked each statement against the facts?

Then the inputs. Do we rely on web data or AI vendors whose terms could close or change, and do we test our AI tools against their published rules? If we adopt a voluntary AI code, who owns it, and how will we show that we follow it?

The Move

The SEC is reading what AI companies tell investors. OpenAI concedes its models do not yet fully follow its own rulebook, and Reddit is closing its doors to scrapers.

Ask the general counsel and the chief financial officer to put every public statement we have made about AI on one page, with the evidence behind each, before the next earnings release.

Say only what you can show.

Seven bullets. One move.