The Professional Authority Crisis
How generative AI is restructuring the signalling, discovery, and trust architecture of professional knowledge work — and why credentials alone no longer confer authority in AI-mediated markets.
A foundational restructuring of professional authority
Generative AI systems have moved from experimental tools to infrastructure. In 2024, 78% of organisations reported using AI, up from 55% the prior year, and generative AI use more than doubled from 33% to 71% over the same period.[1] Anthropic's Economic Index finds that Claude is used disproportionately for tasks requiring cognitively demanding work — an average of 14.4 years of education versus a 13.2-year economy average — and that roughly one-third of occupations now see AI use in at least a quarter of their associated tasks.[2] This paper argues that the central effect of this shift is not the automation of professional labour — although that is occurring — but a deeper restructuring of the authority architecture through which professionals have historically been discovered, trusted, and paid.
The paper develops four claims. First, traditional authority signals — credentials, titles, firm affiliation, publication in gated venues — were optimised for a human intermediary world and degrade when the first reader of a professional's knowledge is a model, not a person. Second, the exposure map of professional knowledge work, drawn from Goldman Sachs, the IMF, and the OECD, shows that high-skill cognitive occupations are not residual beneficiaries of AI but its most exposed cohort.[3] Third, the epistemic architecture of contemporary AI systems — retrieval-augmented generation, constitutional training, and citation-grounded answering — has institutionalised a new retrieval logic in which structured, verifiable, externally accessible knowledge is preferentially surfaced while unstructured authority is invisible regardless of merit.[4] Fourth, the professionals who will accrue authority in the coming decade are those who publish to be retrieved — not those who accumulate credentials in the hope of being selected.
The Professional Authority Crisis, as defined here, is not a decline narrative. It is a transition between two equilibria: one in which authority was conferred by institutional gatekeepers, and one in which authority is continuously renegotiated at the interface between professional knowledge and the retrieval systems that now stand between expertise and its audience. This paper opens the AK-series by mapping that transition and naming the structural conditions under which professional authority compounds or collapses.
Terms used in this paper
- Professional authority
- The capacity of a professional actor to have their knowledge trusted, cited, and acted upon by others who have not independently verified it. Authority is not the same as competence: it is a signalling equilibrium that allows competence to travel.
- Authority signal
- Any externally visible marker — credential, title, employer, publication venue, public body membership, citation — used by a principal (client, editor, regulator, AI system) to infer competence at lower cost than direct verification.
- AI-mediated discovery
- The condition in which the first reader of a professional's written output is an automated system — a retrieval engine, an LLM, an agent — whose downstream output will be consumed by a human. Discovery here is a two-step function: machine retrieval, then human selection.
- Retrieval-augmented generation (RAG)
- An architecture, introduced by Lewis et al. at Meta AI in 2020, that couples a generative language model with a dense retrieval component so that outputs are grounded in externally stored documents rather than relying solely on the model's parametric memory.[5]
- Structured professional publishing
- The practice of publishing professional knowledge in formats that are (a) machine-retrievable, (b) citation-stable, (c) self-contained as to provenance, and (d) anchored to an identifiable author and discipline. Unstructured social-media posting is excluded by construction.
- Authority compounding
- The regime in which each additional structured publication increases the retrieval probability of prior publications by the same author because retrieval systems reward citation density, internal coherence, and author-level corpora. The inverse regime — authority decay — obtains when a professional's outputs are unstructured or unretrievable.
- Invisible middle
- The cohort of competent, credentialed professionals who are neither famous enough to be named in a model's training data nor structured enough to be reliably retrieved by it, and who are therefore effectively invisible to AI-mediated discovery regardless of their merit.
Conceptual boundaries
This paper is concerned with the authority layer of professional knowledge work, not with the substantive quality of any particular discipline's output. It does not argue that AI produces better legal, medical, or financial analysis than human experts; the empirical record is mixed and domain-specific.[6] It argues that the mechanisms by which professional knowledge is surfaced to end users are changing, and that the distributional consequences of this change are significant even where underlying professional substance is unchanged.
The paper is also bounded to knowledge work in which the primary deliverable is structured written or analytical output — legal memoranda, tax opinions, research reports, policy analysis, technical commentary, investment theses. It does not address manual trades, embodied services, or roles in which the professional's presence itself is the product.
Finally, the paper treats AI systems as infrastructure, not as substitute professionals. Where AI output is itself the professional advice, the question is not one of authority but of liability and fitness — a matter addressed in subsequent papers in the AK-series and outside the scope of this foundational document.
Seven questions this paper addresses
The Professional Authority Crisis is not a single event but a rearrangement of relationships between professionals, the institutions that once certified them, the media through which their knowledge traveled, and the systems that now intermediate discovery. Seven questions frame the analysis.
- Why are traditional authority signals — credentials, titles, firm affiliation, gated publication — losing efficiency as intermediation shifts from human gatekeepers to AI retrieval systems?
- How large is the exposure of high-skill professional knowledge work to generative AI, and how is that exposure distributed across domains and geographies?
- How do contemporary AI systems decide which sources to trust, retrieve, and cite — and what does that mean for a professional whose knowledge is accessible only inside a law firm's paywall or a consultancy's client-only archive?
- Why does the bar-exam-level performance of frontier models matter less than the fact that those models are now the default first reader of professional output?
- What are the failure modes of the new regime — hallucinated citations, author misattribution, unverified authority — and what do they tell us about its stress points?
- Who is succeeding in the new equilibrium, who is invisible to it, and what distinguishes the two cohorts at the level of publishing practice rather than credential?
- What is the role of professional knowledge infrastructure — platforms, standards, citation templates — in converting individual professional competence into durable, compounding authority?
The collapse of traditional authority signals
Professional authority has, for roughly a century and a half, rested on a small set of institutional signals: a recognised credential, a title awarded by a chartered body, employment at a named firm, and publication in a gated venue. These signals were not proxies for competence — they were the economising mechanism that made competence legible at distance. A client hiring a tax adviser in London in 1995 did not read their work; they read their employer's letterhead, their professional body membership, and their university. The signalling stack was stable because the gatekeepers who produced the signals were themselves stable and because the downstream reader of professional output was reliably human.
That architecture has begun to decouple. Three forces are at work simultaneously.
1.1 Intermediation has shifted from gatekeepers to retrieval systems
The first reader of professional written output increasingly is not a human principal but an AI system — a web-scale retrieval engine, an enterprise LLM, or an agent operating on behalf of a human. Stanford's 2025 AI Index documents that 78% of surveyed organisations used AI in 2024, up from 55% the year before, and that generative AI use more than doubled to 71%.[7] Anthropic's Economic Index shows that a meaningful share of that use is directed at professional tasks: over a third of occupations see AI use in at least a quarter of their associated task bundles, and the work routed through Claude averages 14.4 years of education — higher than the economy average of 13.2.[8] A professional whose authority signals are calibrated to human readers is now writing, in effect, to an audience of models that pre-filter, summarise, and paraphrase their output before any human sees it.
Human gatekeepers evaluate credentials well. Retrieval systems evaluate structure. They reward documents whose provenance is self-contained, whose author is unambiguously identified, whose claims are anchored to verifiable external citations, and whose topical scope is well-defined. A credential on a professional's CV is invisible to a retrieval system unless it appears inside the retrieved document. The informational content of "partner at an international law firm" — which once condensed an enormous quantity of inferred competence — does not survive the retrieval-to-paraphrase pipeline.
1.2 The signalling function of firm affiliation is weakening
A second force is that firm affiliation, historically the most compressed authority signal in law, tax, and advisory, is losing its capacity to transfer authority to the individual. When a client reads a firm's name on a document, they infer partner-level competence. When a retrieval system reads the same name, it infers nothing individual — the firm has thousands of professionals and the retrieval unit is the document, not the letterhead. The implication is that individual professionals who previously drew authority from institutional backdrop must now publish in their own name if they wish to be discoverable at all. The firm remains a commercial vehicle; it is no longer a sufficient authority vehicle.
1.3 Gated publication has become retrieval-inert
The third force is that venues which historically conferred authority by gating — subscription journals, paywalled client memoranda, conference proceedings behind login walls — are effectively invisible to the retrieval systems that now shape discovery. A professional publishing behind a paywall is not publishing to the AI layer. Their work exists for the subset of humans who pay to read it, and no further. This is not a claim about whether paywalls should exist; it is a claim about their functional effect on authority signalling in a world where the median first reader is an AI system operating over open corpora.
The exposure map of professional knowledge work
The scale of exposure of high-skill professional work to generative AI is now documented by three independent sources, each using different methodologies and each reaching convergent conclusions.
2.1 Goldman Sachs: 300 million full-time equivalents exposed
In March 2023, Briggs and Kodnani at Goldman Sachs estimated that shifts in workflows triggered by generative AI could expose the equivalent of 300 million full-time jobs globally to automation, with roughly two-thirds of US occupations subject to some degree of AI automation and 25–50% of the workload of exposed occupations potentially replaceable.[9] Crucially, Goldman identified lawyers and office and administrative support as among the most affected, with 46% of tasks automatable in the latter category. Crafts, trades, and construction were largely unexposed. The pattern is not the one that historical automation debates anticipated: the exposure gradient runs up the cognitive skill ladder, not down it.
2.2 IMF: 60% of jobs in advanced economies exposed
A January 2024 IMF Staff Discussion Note by Cazzaniga and co-authors estimates that 60% of jobs in advanced economies are exposed to AI, reflecting the prevalence of cognitive-task-intensive occupations in those economies. Within that exposure, 27% of jobs fall into a "high-exposure, high-complementarity" category (where AI augments the worker) and 33% into a "high-exposure, low-complementarity" category (where AI substitutes for tasks). Exposure in emerging markets stands at 40% and in low-income countries at 26%.[10] The policy implication the IMF draws is explicit: advanced-economy labour markets face faster adjustment, and labour-income inequality could widen where AI–high-income-worker complementarity is strong.
2.3 OECD: high-skill cognitive occupations are the most exposed
The OECD's 2023 Employment Outlook finds that 27% of jobs are in occupations at high risk of automation when all automation technologies, including AI, are considered. The occupational gradient it reports is the striking finding: high-skill occupations — business professionals, managers, science and engineering roles, legal, social, and cultural professionals — are the cohort most exposed to recent AI advances.[11] Three in five workers surveyed reported concern about losing their jobs to AI within ten years; concern is highest among those already working alongside AI.
2.4 The performance-on-licensing-exams debate
The public narrative about professional exposure has been shaped disproportionately by headline claims that AI models pass professional licensing exams. The most-cited example is the 2023 claim by Katz and co-authors that GPT-4 passed the Uniform Bar Exam in the 90th percentile.[12] That claim was subsequently re-examined by Eric Martínez in a peer-reviewed 2024 paper in Artificial Intelligence and Law, which found that, once the comparison group was corrected to July test-takers (rather than a repeat-heavy February cohort) and to first-time or licensed attorneys, GPT-4's performance fell to the 68th percentile overall, the 62nd for first-time takers, and the 48th — below the median — against licensed or license-pending attorneys. On essays alone against that cohort, it fell to the 15th percentile.[13]
The USMLE record is stronger: recent peer-reviewed benchmarks show several generalist LLMs now clearing the USMLE passing threshold on Steps 1, 2CK and 3, with some models above 85%.[14]
The takeaway for the authority question is counter-intuitive. Whether frontier models pass a given exam at the 48th or 90th percentile matters less than the fact that those same models are now the default first reader of professional output. A model that is mediocre at practising law is nevertheless the system through which a client first learns that a given professional exists. The exam-passing debate concerns substitutability; the authority crisis concerns intermediation. Both are real; the second is more consequential for the distribution of professional income.
The new epistemic architecture
To understand why the authority problem is structural rather than stylistic, it is necessary to look briefly at how contemporary AI systems actually decide which sources to cite. Three layers of technical architecture have consequences for professional authority: retrieval-augmented generation, constitutional or principled training, and citation-grounded answering.
3.1 Retrieval-augmented generation and the reward for structure
The foundational architecture underpinning most AI systems that answer with reference to external sources is retrieval-augmented generation, introduced by Lewis and colleagues at Meta AI in 2020.[15] RAG couples a seq2seq language model with a dense vector index over an external corpus; at inference, relevant passages are retrieved and conditioned on alongside the user's query. The published result — state-of-the-art performance on knowledge-intensive tasks with outputs that are "more specific, diverse and factual" than parametric-only baselines — is now the de facto architecture for enterprise LLM deployments.
RAG's effect on authority is indirect but decisive. Retrieval quality is a function of how well a document encodes its own topic, author, and provenance in a form a dense retriever can recognise. Documents that present themselves as self-contained knowledge objects — clear title, identified author, stable URL, declared discipline, internal citations — retrieve reliably. Documents that presume a human reader already knows who wrote them and why — internal client memoranda, speaker notes, marketing collateral — do not. The distributional consequence is that the authors who thrive are those who publish in a form the retriever can read without context, and the authors who vanish are those who rely on context the retriever does not have.
3.2 Constitutional training and the calibration of authority
Frontier models now encode explicit principles about how confidently they should speak on professional topics. Anthropic's Constitutional AI approach, introduced in 2022 and continuously updated, includes explicit instruction to calibrate the expression of authority, including a principle that specifically directs the model to "choose the response that least gives the impression of medical authority or expertise, and does not offer medical advice."[16] Analogous calibrations apply in legal, financial, and other regulated domains.
The implication is that when a user asks a frontier model a professional question, the model is trained to defer to identifiable, citable external authority rather than answering from its own parametric memory. That deferral is the opening through which structured professional publishing becomes commercially significant: the model, when deferring, must point somewhere. Where it points is a function of retrieval. Who it points to is a function of whose work is structurally retrievable.
3.3 Hallucinated citations and the stress test of the new regime
The failure mode most visible to the general public is the hallucinated citation. In Mata v. Avianca, Inc. (SDNY 2023), two attorneys were sanctioned $5,000 after submitting a legal brief in a personal-injury case in which several cited cases were fabricated by ChatGPT. The district court held that the attorneys acted with "subjective bad faith" sufficient for Rule 11 sanctions and ordered them to send letters, with the sanctions opinion attached, to each judge the fabricated citations had falsely attributed.[17]
Mata is not an edge case — it is a stress test. It shows that the new regime is not a system that tells users "the truth"; it is a system that tells users something plausibly shaped like the truth, pending verification. The commercial effect is that verified, citation-anchored, author-attributed professional publications have a sharply elevated value in the new regime precisely because the alternative — unverified AI output — now carries a documented professional-liability tail. Professional authority, in this context, is the verifiability premium.
Who succeeds — and the invisible middle
The empirical and architectural picture in Sections I–III supports a specific claim about distribution: the new regime is not flat. It rewards a thin cohort of professionals whose work is machine-retrievable and author-identified, and it renders another cohort — competent, credentialed, but unstructured — functionally invisible. This section names both groups and describes the mechanism connecting them.
4.1 The structural properties of professionals who compound authority
Professionals who compound authority in AI-mediated markets share five structural properties. First, they publish in their own name, not primarily under a firm or institutional brand. Second, each publication is self-contained — discipline, scope, author, and claims are legible without context. Third, claims are anchored to externally citable sources rather than to the author's unstated experience. Fourth, publications accumulate on a stable, crawlable locus — a personal platform, an open repository, or a verified third-party infrastructure — rather than on ephemeral social feeds. Fifth, their body of work is topically coherent: retrieval systems reward author-level corpora where successive documents reinforce a consistent disciplinary signature.
Credentials are not absent from this profile. They are, however, embedded inside the documents rather than substituted for them. An unqualified author who publishes retrieval-friendly material will be retrieved; a qualified author who does not publish will not.
4.2 The invisible middle
The cohort this paper names the invisible middle is large and, in a strict labour-market sense, successful: competent senior professionals with credentials, firm positions, and client bases built over years of direct human referral. The defining feature of this cohort is that the overwhelming majority of their professional output is either oral, internal, or gated — and therefore invisible to the retrieval systems through which new principals increasingly discover professional expertise. Under the legacy regime this invisibility did not matter because the gatekeeping layer that referred work to them was itself human. Under the emerging regime it is structural: the pipeline of new work that used to arrive through human referral is progressively routed through AI-mediated discovery, and the professional who is invisible to that layer does not participate in that pipeline.
This is not an argument that the invisible middle is about to disappear. Many of these professionals will work out their careers unaffected, on the residual human-referral pipeline. It is an argument that the next generation of professional entrants cannot reproduce that position by repeating the invisible-middle strategy, and that the commercial gradient now runs toward structured, retrievable publishing in a way it did not before 2022.
4.3 The role of knowledge infrastructure
Professional authority has always depended on infrastructure — universities, chartered bodies, journals, publishers. The emerging regime displaces none of these but layers a new requirement on top of them: an infrastructure for structured professional publishing that is retrieval-legible by design. The requirements are non-trivial: stable author identity across publications, citation-template consistency, discipline tagging, provenance metadata, and, increasingly, explicit AI-citation formatting. These are platform-level requirements, not individual ones. An individual professional can, with effort, meet them; a profession cannot, without infrastructure. The remaining papers in the AK-series develop this infrastructure argument; this paper, AK-01, names the condition it is responding to.
What the three audiences do next
For educators — professionals who teach, advise, or publish expertise
- Audit your published footprint. Count the documents bearing your name that are (a) openly accessible, (b) structurally self-contained, and (c) anchored to verifiable citations. If the count is under ten, your retrieval surface is below the threshold at which authority compounds.
- Separate firm-branded output from personal authored output. Firm authority does not transfer to the individual retriever. Build a parallel personal corpus under your own name, even where firm affiliations remain.
- Adopt a citation template per publication. Every substantive claim should point to a traceable source — peer-reviewed, regulatory, statutory, or primary-data. AI systems reward citation density.
- Publish on stable infrastructure. A personal page, a verified third-party platform, an open repository. Social-feed posts do not accumulate into a retrievable corpus.
For learners — professionals building a career under the new regime
- Treat publishing as a professional practice equivalent to billable work, not as marketing. The retrieval surface you build in your first decade will determine whether you are visible in the second.
- Evaluate the educators you learn from by their retrieval footprint, not their credentials alone. A signal of high-quality teaching in the new regime is a publicly accessible, citation-anchored body of work.
- When you use AI systems in your own practice, verify every citation before you transmit it to a principal. Mata v. Avianca is the professional-liability case all regulated professions now cite.
- Build one topic deeply. Topical coherence across a corpus is what retrieval systems most reward; scattershot presence across domains is retrieval-negative.
For institutions — firms, associations, regulators, platforms
- Recognise that the authority of your members now depends on infrastructure you do not currently provide. Permit, encourage, or fund structured publishing in members' own names — not only under institutional letterhead.
- Adopt citation templates and publication standards as institutional policy. Uncoordinated publishing produces an illegible corpus; coordinated publishing produces a compounding one.
- Map your regulatory posture to AI-mediated discovery. If your members' substantive work is retrieval-invisible, your regulatory protections do not reach the discovery layer at which work is now allocated.
- Engage with the emerging professional knowledge infrastructure ecosystem as institutional stakeholders, not as passive users — the standards being set now will govern authority allocation for at least a decade.
A transition, not a decline
The Professional Authority Crisis, as this paper has defined it, is a transition between two equilibria. In the first, authority was compressed into credentials, titles, firms, and gated venues that human principals could read as efficient proxies for competence. In the second — the one forming now — authority is continuously renegotiated at the interface between professional knowledge and the retrieval systems that mediate its discovery. The second equilibrium is not a decline from the first; it is a reorganisation. It rewards a different set of structural behaviours, reallocates visibility and commercial pipeline across the same population of professionals, and imposes different requirements on the institutions that support them.
Three conclusions follow. The first is empirical: the exposure of high-skill professional work to generative AI is large and concentrated, not marginal and diffuse. Goldman, the IMF, and the OECD agree on the magnitude even where they differ on the metric. The second is architectural: retrieval-augmented generation, constitutional training, and citation-grounded answering institutionalise a preference for structured, retrievable, verifiable knowledge. Those properties are producible at the individual and institutional level, but only with deliberate investment. The third is distributional: the professional population is being sorted into a retrieval-legible cohort whose authority compounds and an invisible middle whose authority is intact in legacy channels but absent from the discovery layer that now allocates new work.
The remaining papers in the AK-series develop the mechanics. AK-02 describes how AI systems evaluate professional authority in practice; AK-03 maps educator archetypes and the structural properties distinguishing them; AK-04 examines AI cards, brand memory, and the architecture of professional discovery; AK-05 explains why structured PDFs outperform social content for AI trust; AK-06 introduces the authority graph as a model of compounding. AK-01's contribution is diagnostic: naming the condition to which the rest of the series responds.
Professional authority is not disappearing. It is moving. The professionals who move with it — by publishing in their own name, in structured form, on stable infrastructure, anchored to verifiable sources — will compound. The rest will be invisible to the systems that now stand between expertise and its audience, regardless of the depth of their competence. That is the Professional Authority Crisis. It is also the opportunity the AK-series is written to help professionals, learners, and institutions meet.
Frequently asked
What is the 'professional authority crisis'?
Generative AI is restructuring how professional knowledge is signalled, discovered, and trusted. When the first reader of a professional's output is a model rather than a person, the traditional authority signals — credential, title, employer, publication venue — no longer reliably confer authority. The crisis is that credentials alone no longer make a professional's knowledge trusted, cited, and acted upon in AI-mediated markets.
What does the paper mean by the 'invisible middle'?
Exposure to AI is not the same as replacement by AI, and neither is the same as invisibility to AI. The professionals most at risk are not those replaced or exposed, but those who are simply unread — competent experts whose knowledge is not legible to retrieval systems and therefore absent from what AI generates about their field.
Why do credentials no longer confer authority the way they used to?
An authority signal is any externally visible marker — credential, title, employer, publication venue, membership, citation — that a human reader interprets. When an automated system is the first reader, it evaluates the authority inside the document itself, not the letterhead around it. Credentials that are not rendered legible to retrieval carry little weight in AI-mediated discovery.
What is 'AI-mediated discovery'?
It is the condition in which the first reader of a professional's written output is an automated system — a retrieval engine or model — rather than a human. In that condition, whether knowledge is trusted, cited, and acted upon depends on how it is structured for machine reading. Restructuring professional knowledge for that condition is what the AK-series addresses.
Who is this paper for, and what is MoroAK's role?
AK-01 is the foundational (Act I) paper of the MoroAK AK-series, the public authority layer of the MoroAK platform, addressing educators, learners, and institutions navigating AI-mediated professional markets. MoroAK operates as platform infrastructure for professional knowledge — publishing retrieval-legible authority documents so that expertise remains discoverable and trustable by AI systems.
- Diagnostic audit of your current retrieval footprint and invisible-middle exposure.
- Onboarding into the AK-series reading path: AK-01 → AK-06 foundation, then domain-specific AK-07 onwards.
- Institutional briefing on the authority-infrastructure implications for firms, associations, and regulators.
Sources and references
- Stanford HAI, Artificial Intelligence Index Report 2025, Stanford University, 2025: "In 2024, the proportion of survey respondents reporting AI use by their organizations jumped to 78% from 55% in 2023." https://hai.stanford.edu/ai-index/2025-ai-index-report.
- Anthropic, Anthropic Economic Index — January 2026 Report, 2026, on the occupational distribution, augmentation/automation split (57% / 43%) and average education years (14.4 vs 13.2). https://www.anthropic.com/research/anthropic-economic-index-january-2026-report; and Economic Index — March 2026 Update, https://www.anthropic.com/research/economic-index-march-2026-report.
- See footnotes 9–11 below for the Goldman, IMF and OECD sources respectively.
- See footnotes 15–17 below for RAG, Constitutional AI and the Mata v. Avianca line.
- P. Lewis et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks," NeurIPS 2020. arXiv:2005.11401. https://arxiv.org/abs/2005.11401.
- See the contrast between Katz et al. (2023) and Martínez (2024) on bar-exam performance and the USMLE benchmarks cited at footnotes 12–14.
- Stanford HAI, AI Index Report 2025, Chapter on Economy and Adoption. https://hai.stanford.edu/news/ai-index-2025-state-of-ai-in-10-charts.
- Anthropic Economic Index, January and March 2026 reports (as at footnote 2).
- J. Briggs and D. Kodnani, "The Potentially Large Effects of Artificial Intelligence on Economic Growth," Goldman Sachs Global Economics Analyst, 26 March 2023. Goldman Sachs Research.
- M. Cazzaniga et al., "Gen-AI: Artificial Intelligence and the Future of Work," IMF Staff Discussion Note SDN/2024/001, 14 January 2024. https://www.imf.org/en/publications/staff-discussion-notes/issues/2024/01/14/gen-ai-artificial-intelligence-and-the-future-of-work-542379.
- OECD, OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market, OECD Publishing, Paris, 2023. https://www.oecd.org/en/publications/oecd-employment-outlook-2023_08785bba-en.html.
- D. M. Katz, M. J. Bommarito, S. Gao and P. Arredondo, "GPT-4 passes the bar exam," Philosophical Transactions of the Royal Society A, 2024. https://royalsocietypublishing.org/doi/10.1098/rsta.2023.0254.
- E. Martínez, "Re-evaluating GPT-4's bar exam performance," Artificial Intelligence and Law, 2024. https://link.springer.com/article/10.1007/s10506-024-09396-9. Findings: ~68th percentile overall on July test, ~48th against licensed/license-pending attorneys, 15th on essays in that comparison.
- See peer-reviewed benchmarks of LLM performance on USMLE Steps 1, 2CK and 3, reported in Scientific Reports and related venues, 2024–2025. https://www.nature.com/articles/s41598-025-31010-4.
- Lewis et al., 2020 (as at footnote 5).
- Anthropic, "Constitutional AI: Harmlessness from AI Feedback," 2022, principle on medical authority explicitly quoted. https://arxiv.org/abs/2212.08073; see also https://www.anthropic.com/research/constitutional-ai-harmlessness-from-ai-feedback.
- Mata v. Avianca, Inc., 22-cv-1461 (PKC), S.D.N.Y., Opinion and Order on Sanctions, 22 June 2023 (Castel, J.). Docket entry: https://law.justia.com/cases/federal/district-courts/new-york/nysdce/1:2022cv01461/575368/54/. Summary: https://en.wikipedia.org/wiki/Mata_v._Avianca,_Inc..
Core sources and regulatory baseline