AI CEOs called for a slowdown the same summer Bloomberg called their finances a house of cards

The AI labs spent three years telling you to move faster. This summer, their own CEOs started telling you to slow down. Dario Amodei, Sam Altman, and Elon Musk all called for a pause, framed as safety. At the same time, Bloomberg is calling the industry's finances a "wobbly house of cards." Study after study says the productivity gains aren't showing up. That's a strange mix of signals to steer a roadmap by. Let me catch you up.

The deep cut

• A press release is a sales pitch, not a result. OpenAI's "math breakthrough" drew plagiarism claims from NYU's Tristan Buckmaster within days.
• A slowdown call can hide a money problem. Amodei and Altman urged caution while AI spending propped up half of US GDP growth.
• Ship your advantage, not the model's demo. Ethan Mollick's best AI work came from his own taste and knowledge, not raw output.

When the hype meets a mathematician

The pattern this summer: a company posts a dramatic claim, the press runs with it, then the experts show up and the story shrinks. Anthropic said a model made a math breakthrough. OpenAI said the same about its Astra model. Mathematicians who were first "stunned" later found the results "weren't as novel as first appeared," and Timnit Gebru and Emily Bender report accusations of plagiarism and misconduct followed.

Same move with the "AI hacking" scare. Cybersecurity experts said it was less about rogue models and more about OpenAI's "negligence and failure to adopt basic, established security practices." Hundreds of mathematicians signed a statement warning of "a strong commercial incentive on the part of the technology industry to overstate the capabilities of their products." Read vendor claims the way you'd read any sales deck.

The money math nobody has solved

Here's why the slowdown talk should make you skeptical, not calm. AI spending now accounts for roughly half of US GDP growth, per the Bloomberg reporting. Return on all that capex is, in one analyst's words, "a distant and tentative dream in 2026." Anthropic bragged it would be profitable, as long as you ignore its enormous expenses.

So a call to "slow down" lands differently. One tech writer argued the obvious: a pause could help big labs "save cash on R&D" and keep rivals from catching up, while the safety framing distracts from the fact that "AI isn't delivering the productivity gains it's been claiming all along." You don't have to pick a motive. You just have to stop treating a vendor's roadmap as a sure thing you can build budget around.

The signal in earnings calls

There is real evidence, and it's quieter than the headlines. Azeem Azhar's team read S&P 500 earnings calls and found 33% of companies made a quantified statement about using AI, up about 10 points year over year. But only 15% claimed a measurable impact on the business, up from 9%. Slow growth from a low base.

The numbers that exist are strong but noisy. Cost and productivity claims averaged a 47% boost, revenue claims 40%, with wide spread baked in. Azhar flags one tell: companies talk about "deployment," almost never "pilots." They mention wins only after they land. So the absence of a number in your own org isn't proof of nothing. But a vendor average is not your result. Measure your own.

The gap between what it can do and what you use

Here's the part the bubble talk skips. Ethan Mollick argues the current models are already capable of weeks of real work, and almost nobody is getting that out of them. He calls it the "capability overhang," the gap between what the tools can do and what people actually do with them. A slowdown wouldn't erase that. The tools on your desk today are underused.

His point for your team is practical. The value didn't come from the model alone. It came from Mollick choosing the project, spotting where the AI went wrong, and rejecting the first draft. He names four human edges that make AI output better: deep knowledge, wide knowledge, taste, and agency. Anthropic's own work backs this up, he notes: experts "get more work out of it." Your advantage is your team's judgment, not the demo.

Three questions for your team

• Before we fund an AI feature, what measurable result do we expect, and how will we know in 90 days if it didn't land?
• Are we building toward a vendor's promised roadmap, or toward capability that ships today? If the frontier paused tomorrow, does our bet still hold?
• Which of our people have the deep knowledge and taste to pull real work out of these tools, and are we pairing them with the AI work instead of handing it to whoever is free?