On Monday, CNBC’s Daily Open described a turn in how the big AI labs talk about themselves: after two years of racing each other on raw capability, the language has shifted toward caution, guardrails, and restraint. a shift from speed to seatbelts. The framing is neat. It also assumes something nobody has actually built: a way to measure whether any of that restraint is real.
Compare that to a tool the Guardian covered the same day, one that does the opposite of vague reassurance. Researchers have built a system that directly links a specific volume of carbon emissions to a dollar figure of environmental and economic damage. It takes something diffuse and unfelt, a ton of carbon released into a shared atmosphere, and turns it into a number a court, an insurer, or a shareholder can act on. That is what real accountability infrastructure looks like. It does not announce a value. It measures one.
Josh Kaufman names the reason restraint talk without a gauge is so easy to say and so hard to trust. In The Personal MBA, he calls it Absence Blindness: “we have a hard time seeing things that aren’t present.” A carbon tool solves this by making an invisible cost visible after the fact. AI safety commitments have no equivalent. Nobody outside the labs can see the model that wasn’t released, the capability that was held back, or the incident that restraint supposedly prevented. Kaufman’s broader point, that measuring a system in operation is how you catch problems before they arise, cuts the other way here: without a running measurement, there is nothing to catch anything with. A promise to slow down is not an instrument. It is a sentence.
That gap between promise and proof is exactly the space narrative fills. On Bluesky, the writer known as carnage4life pointed to two competing media narratives about AI agents playing out at once, one where agents are quietly taking over real work, one where they are mostly theater. Both stories can find supporting anecdotes because neither is built on a shared count of what agents actually do, how often they succeed, or what they replace. When there is no meter running, coverage reaches for archetype instead: the tireless helper or the thing that comes for your job. Readers pick a side based on which story matches their mood, and the industry gets to claim whichever one is convenient that week.
Tressie McMillan Cottom’s observation, also from Bluesky, points at who is actually having this argument: technology enthusiasts severely overestimate the average person’s desire or will to adopt new tools. Put that next to the seatbelts framing and the whole debate looks smaller than it sounds. Labs are publicly negotiating how fast to drive a car that most people have not gotten into yet. Adoption outside the industry tends to be slower, more skeptical, and more practical than the people building the thing expect, which means the speed-versus-caution fight is largely a conversation insiders are having with each other, dressed up as a conversation about the public.
None of this means the underlying capability race is fake, or that caution is theater. It means the caution is currently unverifiable, and unverifiable claims are cheap to make and cheap to walk back the moment competitive pressure returns. A lab that says it is slowing down can reverse that decision in a press cycle, because there is no external record showing what it actually did versus what it said. The carbon tool works precisely because it does not depend on anyone’s stated intentions. It runs on emissions data whether a company wants it measured or not.
Where the Real Pipeline Is Running
While this argument plays out in headlines that will be forgotten by winter, a much slower and more durable process is underway with almost no attention. IEEE’s TryEngineering program is launching a STEM book series aimed at tweens, introducing engineering concepts to kids years before they choose a career. Daniel Coyle’s research in The Talent Code explains why that timing matters more than it looks: “the vast majority of world-class experts start young,” because the neural wiring that supports deep skill is far easier to build in childhood than to retrofit later. The engineers, auditors, and regulators capable of actually building an AI accountability instrument in 2040 are, right now, ten-year-olds deciding whether a book about how machines work is worth reading on a Tuesday.
That is the real race, and it has no press cycle. It compounds in classrooms, not earnings calls, and nobody currently arguing about seatbelts will get credit for it. The industry that eventually gets measured honestly will not be the one that talked about caution the loudest. It will be the one somebody, eventually, built the instrument to check.
A car can announce that it is slowing down. Only a speedometer can prove that it did.
Leave a Reply