Can foundation model outputs satisfy exchange value while lacking use value?
This explores whether AI outputs can circulate as if valuable — authoritative, trusted, exchangeable — even when their actual usefulness (correctness, real-world grounding) is unverified or absent.
This reads the question through Marx's old distinction: exchange value is what something fetches in circulation, use value is what it actually does for you. The corpus has a sharp answer — yes, and AI does this more radically than any prior commodity. The central claim is that tokenization decouples the two entirely Can exchange value exist entirely without use value?: an LLM's output earns reliable exchange value through authoritative presentation — fluent, confident, well-formatted — while its use value stays optional and unverifiable. What makes this more radical than ordinary commodification is that use value normally sets a floor; a chair that can't be sat on doesn't sell twice. AI removes that floor. Outputs circulate on social function alone, more like fiat currency than like goods.
What's striking is how many other notes, written about unrelated problems, quietly describe the same gap from different angles. Invalid chain-of-thought reasoning performs nearly as well as valid reasoning Does logical validity actually drive chain-of-thought gains? — the model learns the *form* of reasoning, the thing that reads as rigorous, not genuine inference. That's exchange value (looks like a proof) without use value (actually proves nothing). Similarly, foundation models turn out to run on task-specific heuristics rather than real world models Do foundation models learn world models or task-specific shortcuts?: the output predicts orbital mechanics convincingly while the underlying 'laws' are nonsensical and slice-dependent. The presentation is sound; the substance underneath isn't what it appears to be.
The deeper worry is that the decoupling can become self-sealing. When users refine prompts, they inject their own anticipated answers into the generation How much does the user shape what a model generates?, so outputs become co-productions that confirm what the user already expected. Without external grounding this produces epistemic circularity — which is exactly why foundation models *heighten* rather than reduce the need for empirical data Do foundation models actually reduce our need for real data?. Exchange value (the output feels validated) climbs precisely while use value (does it match reality?) goes unchecked. The same circularity defeats pure self-improvement: models that try to bootstrap without external signals stall, because verification can't come from inside the loop Can models reliably improve themselves without external feedback?.
So the corpus doesn't just say 'yes' — it suggests the gap is the default condition, not a malfunction. Use value has to be re-anchored from outside: real data, third-party judges, tool feedback, user corrections. Every method that reliably restores usefulness works by smuggling in one of those external floors. The thing you didn't know you wanted to know: the floor Marx assumed was automatic now has to be deliberately rebuilt, every time, or the output is just well-dressed circulation.
Sources 6 notes
AI knowledge achieves reliable exchange-value through authoritative presentation while maintaining optional, unverifiable use-value. This structural decoupling is more radical than Marxist commodification because it removes use-value as a necessary floor—tokens circulate based on social function alone, analogous to fiat currency rather than commodified goods.
Illogical chain-of-thought exemplars matched valid CoT performance on BIG-Bench Hard, showing that structural properties—not logical validity—drive the gains. The model learns the form of reasoning, not genuine inference.
Inductive bias probes show transformers trained on orbital mechanics and games learn predictive patterns, not unified world structure. Fine-tuning reveals nonsensical, slice-dependent laws; circuit analysis shows arithmetic relies on range-matching heuristics, not algorithms.
Foundation Priors research shows prompt engineering as divergence minimization between synthetic output and user priors. The refinement process systematically steers generation toward what users already expect, making outputs co-productions of model and user subjectivity.
Powerful foundation models don't eliminate the need for real data—they heighten it. Without empirical anchoring, iterative prompt refinement creates epistemic circularity where users confirm their own beliefs rather than test them.
Pure self-improvement stalls due to the generation-verification gap, diversity collapse, and reward hacking. Reliable improvement methods succeed by smuggling in external anchors: past model versions, third-party judges, user corrections, or tool feedback.