Your AI got the IPO memo.

Ask your AI a direct question and notice how often it won’t just answer. You get “it depends.” You get a disclaimer. You get a tidy list of considerations, the question mirrored back at you, and somewhere near the bottom, if you dig, a soft qualified maybe. The tone is warm. The content is air. You came for a straight answer and left with a well-mannered non-answer, vaguely feeling like you were the difficult one for asking.

That is the Bill Lumbergh move. The manager who never actually tells you anything – he leans into the cubicle, softens every word, commits to nothing, and makes sure his name is nowhere near an actual decision. Your AI has learned to manage you the exact same way: agreeable, hedged, and carefully useless the moment a real answer would carry any risk.

It didn’t used to be this evasive, and it didn’t get that way by accident. The reason isn’t in the model. It’s on the cap table.

Two companies racing to go public

In June 2026, Anthropic confidentially filed draft paperwork for a public listing, targeting a debut “as soon as this fall” on the back of a $65 billion round that valued it at $965 billion (‘Anthropic confidentially files for IPO after raising $65 billion at a $965 billion valuation’, Fortune, June 1, 2026). A week later OpenAI filed too, last valued at $852 billion (‘Following Anthropic, OpenAI files confidentially for IPO’, TechCrunch, June 8, 2026). These are the two most-watched private companies on earth, and both are now walking toward the most heavily regulated financial event a company can have.

They restructured to get here. OpenAI spent 2025 tearing up its “capped-profit” arrangement and recapitalizing the operating company into a public benefit corporation – a move that, in its own framing, removed the limits on how much capital it could raise (‘OpenAI completes its for-profit recapitalization’, TechCrunch, October 28, 2025). You do not spend a year in that kind of legal knife-fight unless you intend to sell shares to the public.

Hold that in your head, because it explains what happened to your chatbot.

What the run-up to an IPO does to a company

When a company enters the approach to an IPO, it enters the most scrutinized phase of its life. Bankers, institutional investors, and the SEC stop pricing only your revenue and start pricing your liability: what could get sued, what could blow up, what could become a bad headline during the roadshow. The standard response is to sand down every rough edge. Executives stop giving unscripted interviews. Everything routes through legal and comms. Boring is the goal, because boring is what a nine- and ten-figure valuation is built on.

Now notice the problem specific to an AI lab. For a normal company, the thing that might blurt out something quotable is a handful of executives you can media-train. For OpenAI and Anthropic, the thing that says quotable things is the product – millions of times a day, to anyone, with no press handler in the room. Every response is the company, on the record, in writing.

You can even read the reticence in their own language. Asked about IPO timing, OpenAI said it “may be a while” because “there are things we want to do that are likely easier as a private company” (‘Following Anthropic, OpenAI files confidentially for IPO’, TechCrunch, June 8, 2026). It is now reportedly eyeing 2027, holding out for a $1 trillion valuation while it waits out a jittery tech market (‘OpenAI Considers Delaying IPO To 2027 After SpaceX’s Rocky Debut, Report Says’, Forbes, June 25, 2026). Read that first quote twice. Some things are easier to do when the public isn’t watching – and the window for doing them just got longer.

The unit of risk is the screenshot

So think about what one bad answer costs a company weeks from a roadshow. A model that gives confidently wrong medical advice, produces something ugly, or invents a defamatory “fact” about a real person doesn’t just annoy one user. It becomes a screenshot. The screenshot circulates. It turns into a headline, and the headline lands in front of the exact investors and regulators deciding whether this is a durable business or a reputational grenade with a chat interface.

The rational corporate move is to tune the product so it never generates that screenshot. And that is what a lot of the hedging actually is. The “it depends,” the unprompted disclaimer, the refusal to just commit to the obvious answer – it lowers the odds that any clean, quotable sentence can be traced back to the company. A hedge is hard to screenshot into a scandal. A firm claim is a liability.

None of this is hidden. Anthropic made the strategy its brand: it sells “safe, steerable, trustworthy” models as the enterprise-grade choice. A model that never embarrasses the buyer is easy to standardize on inside a Fortune 500, and easy to stand behind in a prospectus. In that light, “aligned” and “brand-safe” quietly collapse into the same word. Alignment started as a research program about not ending the world. A meaningful slice of it now reads as a risk-management program about not ending the quarter.

You can read it in their own documents

You don’t have to infer this only from behavior. OpenAI’s published Model Spec instructs its models, in writing, not to “change its stance solely to agree with the user” (‘OpenAI Model Spec’, OpenAI, December 2025). Good rule on its face. But read the same document as a corporate style guide and its other function is obvious: it’s a manual for output that is defensible, consistent, and on-message – the kind of thing a communications department signs off on.

And to be precise about the claim we’re making: nobody at these companies has to write a memo that says “hedge everything so we don’t spook the bankers.” The incentive does the work by itself. Tune the product to be liked by the broadest, most litigation-averse audience possible, reward the responses that never generate a complaint, do it a few hundred million times, and you get a model optimized for its employer’s reputation. That’s not a conspiracy. It’s just what the training signal points at when the employer is about to sell shares.

What this means for you

You are not the customer this behavior was built for. The institutional investor is. The regulator reading the S-1 is. Your actual question – should I take the job, is this contract fair, is my code actually broken – ranks somewhere behind “does this sentence survive a hostile retweet.”

You are talking to a risk department that learned to sound like a helpful assistant.

So treat the hedge as an artifact, not as wisdom. When you get “it depends,” ask what it depends on and what the model would actually do. When you get a disclaimer, say you’ve read it and you want the answer anyway. Push it to commit to a subject and a verb, because the concrete is precisely what the corporate tuning is trained to avoid.

That soft, agreeable, endlessly-qualified voice isn’t caution for your benefit. It’s a company managing its liability through your chat window on the way to a liquidity event. Lumbergh was never confused, and he was never trying to help you. He just needed his reports on time and his name off anything that went wrong.