What changes with LLMs in the world?

The shift

Producing plausible language, knowledge, and synthesis — the fluent artifact that reads as if a competent human made it — goes from expensive and scarce to near-zero-cost and effectively unbounded. The deeper change is that production and correctness, long coupled because cost was a rough proxy for care, come apart: the world could trust artifacts roughly in proportion to how hard they were to make, and that proportionality breaks.

The axioms

• Producing fluent language, knowledge, and analysis is expensive, so the supply of it is scarce and its presence signals effort.
• Expertise is scarce and gates access to knowledge; the asymmetry between expert and layperson is what expertise is worth.
• An artifact or credential signals the human effort embedded in it — the essay, the report, the portfolio, the diploma stand in for costly work that was hard to fake.
• Verification rode free on the cost of production: because plausible output was expensive to fabricate, plausibility itself was a usable (if imperfect) proxy for truth and care.
• Trust that a specific human authored or vouched for something is cheap to establish, because faking authorship convincingly, at scale, was expensive.
• Institutions — firms, universities, publishers, agencies — exist largely to aggregate and coordinate scarce cognitive labor and to certify its output.
• Judgment on novel, high-stakes, ill-specified problems, and the taste to decide what is worth doing, is scarce.
• Accountability requires an answerable party, and action in the physical and transactional world costs real resources.

Invalid axioms

1. Producing fluent language and knowledge is expensive, so its presence signals effort. The cost of a competent-looking draft, analysis, or explanation is collapsing toward the price of inference. Fluency is no longer evidence of anything — not effort, not expertise, not care. The habit-trap: the world still reads polish as a proxy for competence and grades, hires, funds, and believes accordingly, so the signal keeps getting extracted long after it stopped carrying information.
2. Expertise is scarce and gates access to knowledge. The asymmetry that let an expert charge for knowing what a layperson couldn't easily find or synthesize thins wherever the value was explanation, synthesis, or retrieval rather than accountable judgment. The habit-trap: pricing, deference, and professional gatekeeping still assume the knowledge itself is the scarce good, when for a widening band of cases it's on tap.
3. An artifact or credential signals the human effort embedded in it. The essay, the cover letter, the analyst's deck, the take-home assignment worked as effort-signals only because producing them cost time and skill that were hard to fake. When the artifact is near-free to generate, it stops standing in for the work — the map detaches from the territory. The habit-trap: whole selection systems (admissions, hiring screens, coursework) still treat the artifact as the evidence, so they now measure prompt access, not the capability they were built to detect.
4. Verification rides free on the cost of production. This is the load-bearing one. For most of history you could lean on plausibility as a cheap proxy for truth because fabricating something plausible was itself expensive — a convincing-looking document, dataset, or citation usually meant someone did the work. LLMs sever that link: plausible and correct are now independent, and plausible-and-wrong is produced as cheaply as plausible-and-right. The habit-trap: institutions, readers, and workflows still trust in proportion to fluency, so they under-invest in verification exactly as the free ride ends.

Unchanged axioms

1. Verification against ground truth stays scarce, and it just became the binding constraint. A model is confidently plausible, not reliably correct; checking a claim against reality — data, the physical world, a source that isn't itself model output — is work that did not get cheaper. What changed is not that verification survives but that it moved from a background tax to the bottleneck, because the thing it used to piggyback on (costly production) disappeared.
2. Accountability requires an answerable party, and a model cannot be one. Someone must own the misdiagnosis, the bad contract, the shipped defect, the false statement. A model can't be liable, sued, licensed, or shamed; liability lands on a human or a firm. This is why capability alone doesn't collapse the value of the roles that carry consequences — vendors structurally decline the accountability, so it stays where a person stands behind the outcome.
3. Action in the physical and transactional world stays scarce. Anything that isn't producing tokens — the procedure, the install, the inspection, the in-person negotiation, moving atoms and money — doesn't get cheaper because drafting did. This is a real bound on how total the change is; it narrows only as robotics and trusted agentic action close the gap, both moving far slower than text generation, so this is a call to revisit rather than a permanent floor.
4. Trust and genuine relationships stay human. The standing to make a commitment, the reputation clients follow, the relationship that survives a mistake — these rest on a human being accountable over time, not on output quality. They get more valuable as fluent output commoditizes, because they're one of the few remaining unfakeable signals.
5. Judgment on novel, high-stakes ambiguity, and the taste to set goals, stays scarce. Models are strong inside a well-framed problem and weak at knowing the frame is wrong, when the confident answer is dangerous, or what is worth doing at all. Deciding the goal, not hitting it, is the residual scarce act. Real, but a narrower base than the synthesis and drafting it's replacing — it doesn't re-absorb everyone the abundance displaces.

New axioms

1. Confidently-wrong output at scale, and the epistemic pollution it produces. For the first time, false-but-plausible content can be generated faster and cheaper than anyone can check it, and it enters the same channels as everything else — search, citations, records, training data for the next model. The commons of shared, trusted knowledge was protected by the cost of polluting it; that protection is gone. The open problem is keeping ground truth legible when the cost of counterfeiting it has fallen to near zero.
2. The collapse of authorship and provenance signals. "A human wrote/made/verified this" is no longer inferable from the artifact, and the systems that assumed it — bylines, credentials, portfolios, review — are running on an assumption that broke. We now have to manufacture provenance (signing, attestation, chains of custody) that used to come free from the cost of faking, and no durable mechanism yet exists at scale.
3. Verification becomes the binding constraint the whole system must be redesigned around. When plausible artifacts are free, the scarce act shifts from producing them to deciding which to trust — but almost no workflow, org chart, or price is built with verification as the main line item. The problem is that verification of arbitrary claims doesn't obviously scale as cheaply as generation does, so the constraint may not simply relocate; it may bottleneck.
4. The apprenticeship problem, economy-wide. Judgment has historically been built by doing the grunt cognitive work — drafting, researching, junior analysis — that is now the first to be automated. If the bottom rungs disappear, the pipeline that produced the verifiers and judgment-holders the system now depends on may not refill, tightening the scarce good over time instead of replenishing it.
5. Who captures the surplus. When the scarce input (competent cognitive output) is supplied by a handful of model vendors rather than by trained people, the productivity gain can be captured upstream of both the firms deploying it and the workers it displaces. Whether the gains diffuse or concentrate hinges on whether frontier capability commoditizes or consolidates — the fastest-moving variable here, and unresolved.

Where it breaks

The whole system still reads fluency as evidence of effort and competence (INVALID #1, #3) at the exact moment plausible-and-wrong is manufactured as cheaply as plausible-and-right (NEW #1). Every selection and trust mechanism built on the old coupling — admissions, hiring screens, peer review, "it looks professional so it's probably fine" — is now measuring the wrong thing, and each one that keeps trusting on polish is an unguarded entry point for confident nonsense into the shared record.

Underneath that: verification rode free on the cost of production (INVALID #4), and now that production is free, verification is the binding constraint (NEW #3) — but nobody staffs, prices, or sequences for it, because for all of history it was a background tax rather than the main cost. The world removed the thing verification was hitching a ride on and has not yet noticed it now has to pay for the ride directly. Whether it can pay — whether checking scales as cheaply as generating — is the open question the near-term trajectory turns on.

ai-axioms.vercel.app/axioms/society/what-changes-with-llms-ai-axioms