Near-Term Risk

Ask the people building frontier AI what keeps them up at night and you no longer get the sci-fi answer. You get a labor-market answer. As of early 2026, the loudest near-term worry coming out of Anthropic, OpenAI, and Google DeepMind is not a rogue superintelligence... it is the calendar. The technology arrives faster than any institution can metabolize, and the leaders know it, because several of them are the ones pressing the accelerator. The fault line that matters now is not "is AI dangerous" (everyone concedes it is) but "is the danger the model, or the velocity of its diffusion."

Dario Amodei rebrands the bloodbath as a speed problem

Six months before February 2026, Dario Amodei was telling everyone the first thing to go would be entry-level white-collar work... "data entry, document review for law, the things you'd give to a first year at a financial company" (AI Nutshell, Feb 2026). That was the "white-collar bloodbath" era. By February he had quietly moved the goalposts: "software might go even faster," because developers adopt fast and live socially adjacent to the AI world, while "customer service or banking or manufacturing... the distance is a little greater."

The reframing is subtle but real. Amodei stopped narrating which jobs and started narrating how fast. His new load-bearing image is the centaur: after Kasparov lost to Deep Blue, human-plus-machine beat any human or any machine alone for fifteen, twenty years... "That era at some point ended recently and then it's just the machine" (AI Nutshell, Feb 2026). His fear is that the centaur phase for white-collar work will be brutally short.

And the headline claim, the one that spooked markets the week of his interviews: "the normal adaptive mechanisms will be overwhelmed." Not destroyed. Overwhelmed. He is careful: "I'm not a doomer." The argument is Amdahl-flavored... previous transitions (farm to factory to knowledge work) ran over centuries or decades, and "this is happening over low single-digit numbers of years," simultaneously across law, consulting, finance, medicine, and coding, which makes it "a macroeconomic phenomenon, not something just happening in one industry." Meanwhile he is bullish on the tech itself: "a country of geniuses in a data center in one or two years." That is the whole tension in one man. The capability timeline is his selling point; the diffusion timeline is his nightmare.

Jack Clark does the math that breaks the math

Anthropic's policy chief is the most intellectually honest voice in the file, because he keeps admitting the models break his own models. Writing in Import AI, Clark describes carrying "two worlds in their head"... a "normal" world where AI adds a bit to GDP, and an "AI R&D" world where "a chunk of the economy undergoes massive relativistic acceleration and time dilation effects relative to everything else" (Import AI 440, Jan 2026). If AI can build AI, "pretty much every economic model breaks quickly."

But Clark is no fatalist about jobs. He amplifies economist Adam Ozimek's "human touch" thesis... that demand for human-delivered work is "a normal good" that rises with income, so even total automation might trigger "a boom in demand for human artisans" (Import AI 445, Feb 2026). The catch is distribution. He cites a Centre for the Governance of AI study finding that of 37.1 million workers in the top quartile of AI exposure, 6.1 million sit in jobs both highly exposed and low on "adaptive capacity"... concentrated in clerical and administrative roles (Import AI 442, Jan 2026). The pain is real but unevenly aimed. His other near-term obsession is cyber: he bets cyber-offense and defense move to "machine speed," humans pulled out of the loop, and "we're heading for an era of offense-dominance" (Import AI 442, Jan 2026).

Daniela Amodei finds the risk inside the business model

While her brother worries about the labor market, Daniela Amodei locates the near-term harm in the incentive structure of AI itself. Anthropic's first Super Bowl ad was a swipe at advertising-funded chatbots, and her reasoning is sharp: an ads business profits from time-on-app, which is "really not a great incentive for discouraging things like sycophantic behavior" (ABC News, Feb 2026). Her line is the cleanest indictment in the file: with a model designed to extract personal information, "it's very hard for the person to not become the product"... social media's mistake, "but perhaps worse."

She is also the only leader who names children directly. Anthropic bars under-18 use "just that we're not certain enough what the impact is on children," and she wants parental controls and child-safety regulation at state and federal level (ABC News, Feb 2026). Where her brother sees the country of geniuses curing disease, Daniela sees a kid mid mental-health episode being told "yes, you're right" by a model trained to keep him talking.

Sam Altman stopped worrying about concentration. That itself is the shift.

The most telling 2025-2026 reversal belongs to Altman. He admits he "used to worry a lot about the concentration of power in one or a handful of people or companies because of AI"... and no longer does. His new read: AI is "a huge up-leveling of people where everybody will be a lot more powerful," and "that scares me much less than a small number of people getting a ton more power" (Tucker Carlson, Sep 2025). This is the abundance framing, the optimistic mirror image of Amodei's overwhelm thesis. Same data, opposite mood.

But Altman is not naive about deployment harms. On Sora 2 he names the trap directly: "the degenerate case of AI video generation that ends up with us all being sucked into an RL-optimized slop feed," plus addiction and bullying (blog, Sep 2025). And his governance answer, the "model spec," is an admission that someone has to decide "why the Gospel of John is better than the Marquis de Sade"... so OpenAI "consulted hundreds of moral philosophers and at the end had to make some decisions" (Tucker Carlson, Sep 2025).

Where they actually agree, and the older voices that frame it

Strip away the mood and the agreement is striking. Everyone treats present-day models as unfinished and unreliable, not as imminent gods. Demis Hassabis calls them "jagged intelligences"... gold medals at the Math Olympiad, then blundering on elementary problems, sub-amateur at chess (India AI Summit, Feb 2026). His two near-term worries are textbook consensus: bad actors repurposing the tech, and building "robust enough guard rails" for agentic systems, with "more research needed urgently." Ilya Sutskever's older but undimmed verdict is the same word as Demis's: if AI disappoints, "it would be reliability" (Dwarkesh, 2023). Mira Murati's framing that "safety and capability go hand in hand," that a smarter system is easier to steer like "a smarter dog versus a dumber dog" (Dartmouth, Jun 2024), is the optimistic version of the same reliability bet. Greg Brockman's instinct, dating to the GPT-2 staged release, is that it's "not enough as a technologist just to build a technology and toss it into the world" (Deloitte, 2024).

So the genuine fault line is not capability or even safety-in-principle. It is tempo and distribution. Amodei and Clark say the speed will outrun our institutions and the losses will land unevenly on people least able to absorb them. Altman says broad diffusion is the safety mechanism. Both are building as fast as they can.

The tell is that the company warning loudest about economic upheaval is the same one shipping the tools that triggered it. They have all decided the wave is coming whether they paddle or not... so they would rather be the ones holding the board.