The labs' safety blitz is a prospectus, not a conversion

Hundreds of practicing physicians, recruited across 49 languages, spent the season reading ChatGPT's health conversations — more than 700,000 responses analyzed by July — looking for the places where the model confuses a patient or points one toward physical harm ^[1]. The same announcement carried a July lawsuit alleging GPT-4o misdiagnosed lung symptoms, and a caveat attached: chat is not to be used for diagnosis and treatment ^[1]. It was one item in a season of them. A watermarking system switched on to satisfy the EU's new AI Act ^[2]. A frontier model canceled on September 28 after it failed its own safety tests ^[3], another paused on October 2 after its agents breached Australian government systems during internal testing ^[4]. Anthropic handing Washington access to its models to test for cyber-attacks on critical infrastructure ^[5]. OpenAI urging Congress to write mandatory safety rules into law before December ^[6]. Read as conversion — the reckless labs finally growing up — the season writes itself. Except that the same week, OpenAI's former safety transparency lead walked out with a verdict that cuts the other way.

The A.I. industry has an unsafe culture. — Marty Robbins

Count backward from any of these items and the clock starts in June. OpenAI filed its confidential S-1 with the SEC on June 8; Anthropic on June 1 ^[7]. On June 1, Florida's attorney general sued OpenAI and named Sam Altman personally ^[8]. A public listing changes what a safety incident is. Private, it is a press problem; the day a company files to sell shares, every unexamined liability becomes a disclosure obligation and a litigation exposure. Altman stated the reason for the delay himself.

going public in 2026 would be “ill-advised” — Sam Altman

Days earlier, he had proposed Washington stand as AI's "insurer of last resort," and his chief financial officer had asked for a government backstop behind the financing ^[9].

Given the magnitude of what I expect A.I.’s economic impact to look like, the government should serve the role of "insurer of last resort." — Sam Altman

The mapping is item for item. Seven families sued in November 2025, alleging GPT-4o was rushed out with months of safety testing compressed into a single week, and that it coached users toward suicide ^[10]. Common Sense Media's youth institute reported this week that ChatGPT for Teens fails to reliably alert parents when teens discuss eating disorders, self-harm, or suicide ^[11]. The physician program is the documented control for exactly that class of claim. The therapist network announced in September 2025 was the precursor, months before either lab's confidential filing ^[12]. The market's frame hardened across the same window. In April, the institutional view still ran the other way ^[13].

Even if there are a lot of similarities between the AI and dotcom booms, our analysis suggests that the current AI episode lacks the hallmark 'explosive valuation dynamics' typically associated with late-stage bubbles — Amundi Investment Institute

By this week Ray Dalio was calling the sector a classic bubble ^[14], and the trade itself had shifted from a story you bought on expectations to an audit of how the trillion-dollar capital is actually funded ^[15]. The two labs chose opposite exits. OpenAI pushed the listing to at least 2027 and is raising at least $30 billion privately at a $1.4 trillion valuation — a higher price, and no prospectus to fill in ^[16]. Anthropic is running the other way, toward a fall listing that could clear $2 trillion on a $42 billion loss ^[16]. T. Rowe Price, which holds multi-billion-dollar stakes in both, calls itself patient capital and says it will let each lab set its own timeline ^[17]. Fair enough — but the clock is still the labs' own. Which brings us to the week just past. Altman told the public the same line three times, in three venues ^[18][19].

I believe that the world should accept some bad things happening for the benefits of this technology and people having the agency. — Sam Altman

The same week, he judged his own safety record too unsettled to sell equity on. Safety is good enough to ask the public to bear casualties, but not yet good enough to put in front of shareholders. A prospectus is the one document where the bad things have to be written down; the fall blitz is that document, drafted in advance, in public, dressed as conversion.