Trying to shape AI tech to protect jobs is a bad idea since no one can predict which technologies will augment humans and which will simply replace their labor
Instead, we should empower workers and create mechanisms for redistribution. — I've been to a number of meetings and panels recently … Tweets: @jackstilgoe , @emmabgo , and @psythor Tweets: Jack Stilgoe / @jackstilgoe : I don't get this from @Noahpinion. Hinton's speculation about radiologists is not a good argument against anticipation or steering. It just means we should broaden expertise to include people who know about how technologies actually impact people https://www.noahpinion.blog/ ... Emma Goldberg / @emmabgo : interesting @Noahpinion post arguing that the answer to fears about AI and job displacement could be 1) union power and 2) universal cash benefits https://www.noahpinion.blog/ ... James O'Malley / @psythor : This @Noahpinion piece on “AI is going to destroy jobs” predictions and how to protect workers is great. Don't stand in the way of technology - have a strong welfare state! https://www.noahpinion.blog/ ... Expand More For Next Unexpand More For Next
Context & Ripple Effects
Noah Smith's argument lands in an already-contentious research field: even in mid-2023, coverage noted that employment researchers could not agree how many jobs AI would touch, partly because automation targets tasks rather than whole occupations. His claim that augmentation versus replacement is unpredictable is a direct attack on the premise of 'anticipation' approaches — shaping AI development itself to preserve work.
The counterposition was immediate: scholars like Jack Stilgoe argued Hinton-style speculation about radiologists is no reason to abandon steering, only to broaden who gets expertise over deployment effects. Smith's alternative — stronger unions, universal cash benefits — also echoes the wealth-concentration worry raised weeks earlier in New Yorker coverage of AI sharpening capitalism's knife blade.
First-order effects
- The immediate effect is on the AI-and-labor policy conversation: Smith forces advocates to choose between steering the technology (Stilgoe's position) and building redistribution mechanisms (unions, cash benefits), rather than pursuing both by default.
- Experts currently teaching AI systems their craft — radiologists, accountants, PhDs cited in his examples — become the test case for whether augmentation predictions can be made at all.
Second-order effects
- If governments accept Smith's unpredictability claim, pressure shifts off AI developers to prove their systems augment rather than replace labor, and onto labor institutions and benefit programs to absorb displacement — a redirection competitors in the policy arena like Stilgoe are already contesting.
- Labor advocates lose the leverage of demanding 'worker-friendly' design from specific deployers, since the argument holds that no one can specify ex ante which deployments protect jobs.
Third-order effects
- The pattern points toward policy convergence on distribution-over-direction: regulating AI's gains through unions and transfers rather than shaping what gets built — though whether that consensus holds depends on evidence still unresolved years later, when critiques of contrasting US employment studies and GovAI-Brookings findings on who finds new jobs show researchers still disagreeing about AI's net effect on work.
The trend: The AI-jobs debate is splitting between steering the technology itself and redistributing its gains, with the empirical record still too contested to settle which approach works.