Your future job will be to keep AI on task

URL SCAN: Your future job will be to keep AI on task | Noah Smith
FIRST LINE: "The one thing humans will always be comparatively good at is knowing what we want."

THE DISSECTION

Smith is performing the textbook move of the intellectual class during systemic transition: finding a seat for humans at the table of the machine. The thesis: AI needs alignment, humans provide alignment, therefore humans remain relevant. He frames this as a "slightly different answer" to the problem of human uselessness, but it's identical copium dressed in technical vocabulary. The Lumbergh analogy is the tell—Smith literally name-checks the character whose entire satirical function was to demonstrate middle management's pointlessness, and reimagines him as humanity's salvation. The move is so audacious it borders on performance art.

The text does three things simultaneously:
1. Acknowledges that AI has found its "killer app" in agentic coding (the exact mechanism of cognitive automation P1)
2. Concedes that "some people in the AI industry still think that humans will be rendered economically irrelevant" - which he dismisses as "unsatisfying" rather than inaccurate
3. Proposes a human niche in "alignment work" and "verification" as if this job category survives intact while every other cognitive task collapses

The slop section is where the thesis breaks down visibly. If AI already generates 50%+ of internet traffic and "over one-third of new websites," and we're "being overwhelmed by a wave of AI output of questionable quality," theVERIFICATION WORKLOAD is not a stable job—it's a drowning person's flailing. Every unit of AI output requiring human verification means the verification labor scales with AI output. This is not a cure. It's a symptom.

THE CORE FALLACY

Humans don't actually know what they want. This is the Hume quote Smith cites as foundational, but he misreads it cruelly. Hume's point is that human preferences are irrational, inconsistent, evolving, and often unconscious. "Knowing what you want" is not a stable input for AI alignment—it's the exact problem that makes alignment hard. Humans can't reliably specify their own goals, and as AI systems become more powerful and general, the gap between "what humans say they want" and "what they actually want" widens catastrophically. Smith's Lumbergh fantasies assume humans have stable, legible goals to transmit. They don't.

Verification is not a durable human niche. Under the Discontinuity Thesis, the economic value of verification collapses the moment AI can verify AI. Smith implicitly assumes human verification remains cheaper or more accurate than algorithmic verification. This assumption collapses when:
- AI auditing tools mature (already happening)
- The verification burden becomes impossible at scale
- Liability shifts to AI-assured outputs with legal/regulatory frameworks

The "slop" problem Smith describes is solved by more sophisticated AI, not by armies of human quality controllers. This is exactly the trajectory of every previous automation " niche"—e.g., automated spell-check replaced human typists who caught spelling errors. Verification automation WILL follow the same path.

The "prestige" niche is a rounding error. Smith acknowledges disbelief in the "human status symbol" thesis but doesn't adequately address scale. Prestige economies serve <1% of economic activity. Post-WWII capitalism derives its stability from mass participation in wage labor. If the only remaining human economic niche is "having a real human make your sandwich as a status symbol," you've described an economy serving billionaire preferences and serving nothing else.

HIDDEN ASSUMPTIONS

1. Human goal stability: Assumes humans can reliably specify their own preferences at scale and over time. Historically false—all preference revelation mechanisms (markets, polls, contracts) are imperfect and gamed. As life stakes rise, this gets worse, not better.

2. Verification as labor-intensive constant: Implicitly assumes verification work scales linearly with AI output, and that human verifiers remain necessary. Disregards the certainty of verification automation.

3. "Alignment" as a permanent human task: Smuggles in that AI labs will require human alignment workers perpetually. In reality, alignment itself is being automated—AI-assisted interpretability, constitutional AI, automated red-teaming. The workers Smith envisions are being recruited to fill a job that will itself be automated.

4. AI remains a tool requiring human steering: The entire thesis assumes humans as essential overseers. But the trajectory of agentic AI is toward less human oversight, not more. Labs want autonomous AI. The "alignment worker" niche competes against the economic incentive to make AI operate with minimal human intervention.

5. Social order continuity: The entire framework assumes stable functioning institutions overseeing this transition. The Discontinuity Thesis addresses what happens when those institutions fragment—which they do when mass unemployment destabilizes consumption, tax bases, and political legitimacy.

SOCIAL FUNCTION

Transition Management Theater. This piece performs the crucial cultural work of reassuring the professional class that their relevance is not ending—merely shifting. It's designed for readers who want to believe AI doesn't threaten them specifically: engineers, managers, knowledge workers. It gives them a story where they graduate from "doing things" to "overseeing things," which is psychologically comfortable.

The Lumbeh analogy is the ideological key. By rehabilitating the film's symbol of uselessness, Smith performs the characteristic move of anxious professional class discourse: turning anxiety about obsolescence into a narrative of subtle indispensability. "Maybe the annoying middle manager WAS right all along" is a seduction for people who fear they're the next annoying middle manager.

This is prestige signaling dressed as structural analysis. It targets an audience that wants serious, technical-sounding reassurance rather than either true optimism or acknowledged despair.

THE VERDICT

Smith's "alignment worker" thesis is a sophisticated-seeming version of the same wrong answer given by everyone predicting human relevance: finding the specific task AI can't do YET, and declaring that task the secure niche. It's the same logic that told truckers in 2015 they'd remain employed because "you still need a human driver to make decisions." The decision-making job got automated before the trucks did.

The structural reality: When AI achieves durable cost and performance superiority in cognitive work (P1), the demand for human verification doesn't stabilize a human economic niche—it merely demonstrates that human verification is the current bottleneck to be eliminated. Every competent AI lab is already working to make human verification unnecessary. Smith is describing a temporary job category, not a permanent human function, and presenting it as if it were the latter.

Mass employment→wage→consumption capitalism cannot survive on "alignment workers" and "sandwich artisans to the rich." The Lumbeh fantasy is a way to not say this.

Copium with footnotes. Designed to comfort. Structurally inaccurate.