Does 'when costliness becomes forgeable' explain generative AI's demonetization of writing, art, and credentials?
This capture answers the open question routed from observation-unforgeable-costliness-bridges-asteroid-pgm-and-bit-gold and recorded at 50-questions/question-generative-ai-demonetization-costliness-becomes-forgeable.md. That question asked whether the same law traced through Szabo's collectible-money history (claim-szabo-collectible-monies-collapse-when-costliness-becomes-forgeable) and bit gold's design (claim-szabo-bit-gold-grounds-value-in-unforgeable-cost-of-production) — a good's value rests on the cost of producing its scarcity, and collapses once a new technology cheapens that cost — also explains generative AI's effect on the value of writing, art, and credentials. None of the four sources found and read this session cite Szabo or use his "costliness becomes forgeable" phrase; the bridge to that specific framing is this vault's own synthesis, not a claim any source makes. What the sources do provide, independently, is measured evidence of value collapse in writing/design freelance markets and university credentials, plus one paper that formalizes the mechanism using Spence's (1973) costly-signaling theory — a structurally identical logic to Szabo's (a signal/scarcity is meaningful only insofar as it is costly to fake) applied to labor markets rather than money.
Claim: Freelance writing and design-related earnings and job counts measurably declined within months of ChatGPT's and image-generator releases, on a large online labor platform
Hui, Reshef, and Zhou studied Upwork freelancers' employment histories around the November 2022 release of ChatGPT, using a difference-in-differences design comparing writing-related occupations (treated) to less-affected occupations (control). Following ChatGPT's release, freelancers in writing-related occupations saw "the monthly number of jobs on the platform for freelancers in more affected occupations decreases by 2% (s.e.=0.004), and total monthly compensation decreases by 5.2% (s.e.=0.016)." The probability of receiving any job in a given month fell 1.2 percentage points (roughly a 10% drop from baseline), and, conditional on working, the number of jobs fell about 4.7%. The paper replicated the design around the earlier (April–July 2022) releases of DALL-E 2 and Midjourney for freelancers offering design, image-editing, and art services: "The estimate of the effect of image-focused generative AI echo our main results: we find consistent negative effects of the release of the new technology on the performance of freelancers on the platform, in terms of both number of jobs and total compensation" — though the exact magnitude for image-related occupations was reported only in an appendix table not recovered from the extracted PDF text this session, so no specific percentage is recorded for that group. Notably, freelancers with above-median past performance were not protected and were, if anything, disproportionately hurt — inconsistent with the idea that high "quality" workers hold a durable premium once the underlying signal becomes cheap to fake.
- source_url: "https://www.ifo.de/DocDL/cesifo1_wp10601.pdf"
- source_author: "Xiang Hui, Oren Reshef, Luofeng Zhou"
- source_date: "2023-08 (CESifo Working Paper No. 10601); later published Organization Science 35(6):1977–1989 (2024)"
- source_quote: "the monthly number of jobs on the platform for freelancers in more affected occupations decreases by 2% (s.e.=0.004), and total monthly compensation decreases by 5.2% (s.e.=0.016)"
- source_tier: 1
Claim: Economists have formally modeled generative AI's labor-market effect as the collapse of costly signaling in writing, and quantify a resulting drop in meritocratic hiring
Galdin and Silbert's job-market paper explicitly frames the mechanism in Spence's (1973) costly-signaling terms: "writing requires time and effort, [so] the act of writing itself can send a signal... LLMs can produce polished, human-sounding text in seconds at virtually no cost... this technology may threaten the signaling value of writing in the labor market." Using Freelancer.com application data (coding jobs), they show employers paid a premium for customized applications before mass LLM adoption ("workers with a one standard deviation higher signal have the same increased chance of being hired as workers with a $26 lower bid") but that this premium and its predictive link to job-completion success weakened sharply after LLM adoption. Simulating a counterfactual in which written signals become worthless, their structural model finds: "compared to the pre-LLM equilibrium, workers in the top quintile of the ability distribution are hired 19% less often, workers in the bottom quintile are hired 14% more often" — i.e., the market becomes measurably less meritocratic once the cost of faking the signal collapses. This is the closest match found this session to Szabo's mechanism stated in labor-market terms, though the paper does not reference Szabo or monetary theory.
- source_url: "https://arxiv.org/pdf/2511.08785"
- source_author: "Anaïs Galdin, Jesse Silbert"
- source_date: "2025-11-11 (arXiv:2511.08785, job market paper)"
- source_quote: "Without costly signaling, employers are less able to identify high-ability workers, causing the market to become significantly less meritocratic: compared to the pre-LLM equilibrium, workers in the top quintile of the ability distribution are hired 19% less often, workers in the bottom quintile are hired 14% more often."
- source_tier: 1
Claim: Generative AI availability during coursework compresses university grade distributions, measurably eroding grades' function as a credential-signal
Hausman, Rigbi, and Weisburd tracked ~36,000 students across ~6,000 courses at a large Israeli university from 2018–2024, comparing AI-compatible courses (take-home work) to AI-incompatible courses (in-person exams) before and after ChatGPT's introduction. They state directly: "AI availability raises grades, especially for lower-performing students, and compresses the grade distribution, eroding the signal value of grades for employers." Quantitatively: grades in AI-compatible courses rose "about 0.6-1 points" on average (against a mean grade of 86), with 25th-percentile students gaining "more than 2.5 additional points" in the most controlled specification; ChatGPT's introduction was "associated with a 33% drop in the share of students failing AI-compatible courses in the first year, and a 50% reduction in the share receiving a D grade in AI-compatible courses the second year." The authors frame this explicitly as a signaling problem: grades increasingly "pool students of a variety of abilities under similar grades," which "complicat[es] hiring, forcing employers to find other means of assessment." This is the clearest direct evidence found this session that the mechanism extends to formal academic credentials specifically (grades), as distinct from freelance-market reputation.
- source_url: "https://www.econstor.eu/bitstream/10419/319211/1/cesifo1_wp11843.pdf"
- source_author: "Naomi Hausman, Oren Rigbi, Sarit Weisburd"
- source_date: "2025-03/2025-04 (CESifo Working Paper No. 11843)"
- source_quote: "AI availability raises grades, especially for lower-performing students, and compresses the grade distribution, eroding the signal value of grades for employers."
- source_tier: 1
Claim: AI-text detectors, the main technical tool institutions use to try to re-verify that writing reflects a person's own costly effort, are themselves unreliable — undermining attempts to restore the cost floor
Liang, Yuksekgonul, Mao, Wu, and Zou (Stanford) tested seven widely-used GPT detectors against 91 human-authored TOEFL essays (non-native English writers) and 88 US 8th-grade essays (native writers). The detectors were near-perfect on the native-writer essays but "misclassified over half of the TOEFL essays as 'AI-generated' (average false positive rate: 61.22%)," with all seven detectors unanimously misflagging almost 20% of genuinely human-written TOEFL essays. Simple prompting strategies that added linguistic variety to genuinely AI-generated text also evaded detection, cutting false-negative-style detectability further. This is a specific technical-mechanism finding (how detection tools actually perform, not merely that they exist) about why the "cost of faking" cannot currently be cheaply re-imposed by institutions after generative AI drove the underlying production cost near zero — relevant to credentials, admissions essays, and any writing used as proof of a human's own effort.
- source_url: "https://arxiv.org/pdf/2304.02819v2"
- source_author: "Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, James Zou"
- source_date: "2023-04-18 (arXiv:2304.02819v2); published in Patterns, 2023"
- source_quote: "these detectors consistently misclassify non-native English writing samples as AI-generated, whereas native writing samples are accurately identified... misclassified over half of the TOEFL essays as "AI-generated" (average false positive rate: 61.22%)"
- source_tier: 1
Further leads
- Upwork's own "Future Workforce Index 2026" (July 2026 press release) claims freelancers using AI earn 34% more per hour and that complex AI-augmented contract volume grew 72% — in apparent tension with the academic Upwork findings above; worth reading the company's own methodology before citing any figure (company-interest source, not yet vetted). — https://www.globenewswire.com/news-release/2026/07/14/3326964/0/en/Upwork-s-Future-Workforce-Index-2026-How-AI-is-Redefining-the-Value-of-Work-as-Skilled-Freelancing-Accelerates.html
- Yilmaz, Naumovska & Aggarwal (SSRN) reportedly find Google Translate and ChatGPT reduced translators' employment, especially for analytical tasks — a fourth freelance-labor domain (translation) not read directly this session.
- A University of Chicago Booth 2025 working paper reportedly stress-tests AI-text detectors against "humanizer" tools and finds large differences across detectors (e.g., Pangram near-zero false-positive rate vs. Turnitin ~1%+ at scale) — not read directly; would sharpen the Liang et al. detector-reliability claim above with more recent data.
- arXiv paper "Artificial Intelligence and Skills: Evidence from Contrastive Learning in Online Job Vacancies" reportedly finds employer demand for formal degrees declining faster in AI-exposed jobs — not read directly; could quantify credential devaluation in hiring (as opposed to grades) more precisely.
- arXiv paper "When AI Levels the Playing Field: Skill Homogenization, Asset Concentration, and Two Regimes of Inequality" reportedly models AI-driven "credential inflation" as employers respond to degraded screening ability — not read directly.
- Washington Monthly has run opposing pieces ("No, College Degrees Aren't Losing Their Value" vs. "How AI Broke the Entry-Level Job") — Tier 3, worth tracing to whatever primary labor data each relies on rather than citing the framing pieces themselves.
- The trajectory question already open in the vault — whether cost-of-production collapses follow "hold, then cliff" (per the asteroid-PGM model, claim-asteroid-pgm-price-holds-then-collapses) or fall more monotonically — has not been checked against any of the four AI-labor datasets above; the CESifo papers' event-study figures (e.g. Figure 2 in Hui, Reshef & Zhou) could potentially answer this directly in a future pass.
Entity candidates
- Xiang Hui, Oren Reshef, Luofeng Zhou — people — authors of the Upwork/ChatGPT freelance-labor study
- Anaïs Galdin, Jesse Silbert — people — authors of the costly-signaling/LLM labor market model
- Naomi Hausman, Oren Rigbi, Sarit Weisburd — people — authors of the university-grades/AI signaling study
- Weixin Liang, James Zou — people — Stanford researchers behind the GPT-detector bias study
- Michael Spence — person — economist whose 1973 job-market signaling model is the explicit theoretical anchor Galdin & Silbert use; worth its own concept note given how directly it parallels Szabo's cost-of-production framing
- costly signaling (economics) — concept — Spence's model that a signal is informative only if costly to produce/fake; the structural twin of Szabo's "unforgeable costliness" applied to labor markets rather than money
- Upwork — concept/entity — the freelance platform whose public data underlies two of the four studies found
- Freelancer.com — concept/entity — the freelance platform underlying the Galdin & Silbert study
- credential inflation — concept — the broader labor-economics term for employers demanding more/different credentials as existing signals degrade; came up repeatedly in further-leads searches and may deserve its own note