CCAA: AI Rewrites the Standard, Not Just the Job

Almost every discussion of AI's impact asks the same question: will this job be replaced? That question covers half the problem.

The other half: even if not a single job is replaced, the ruler used to measure "good work" may already have been swapped out. Swap the ruler and demand structure, pricing and the psychological load on practitioners all move with it — with no layoff announcement, no unemployment print, and almost nothing that any statistical series would catch.

We are giving that half a name. But before the argument begins, one warning: the three most important sections of this article are the three where we concede. A new term that only ever aggrandises itself is marketing copy, not an analytical instrument.

1. Definition — and the half we must exclude first

Cognitive-Conditioning AI Applications (CCAA): a class of AI applications whose output is not consumed as a finished product but adopted as a frame of reference. Under long-term, high-frequency exposure at near-zero marginal cost, the statistical properties of that output — better mean, smaller variance, less friction — are internalised as "normal", rewriting the user's expectation threshold for reality.

One line to keep the division of labour straight:

• Substitution changes who does the work.
• Conditioning changes what counts as normal work.

But "conditioning" has two meanings, and blending them destroys the argument, so they must be cut apart first:

• Procurement conditioning: the buyer's expectations of price, volume, turnaround and revision count are reset. Copywriting is the type case — "a campaign" now means forty variants overnight at zero marginal cost. This is a market phenomenon.
• Perceptual conditioning: a person's baseline for what a face, a conversation, a partner, a body should be is reset. This is a psychological phenomenon.

CCAA, in this article, refers only to the second. The first is ordinary cost-driven standard inflation, fully explained by price competition, and requires no new concept.

That exclusion has a sharp consequence worth stating plainly: the examples one instinctively reaches for — copywriting, generic illustration, template video — are precisely where the thesis is unidentifiable. In those industries the standard moved and the supplier changed at the same time; nothing can be attributed to conditioning, because price competition already explains all of it. Clean evidence exists only where the human is still doing the work and the standard moved anyway: aesthetic medicine, counselling, nursing, teaching. Copywriting cases therefore appear below as illustration, never as evidence.

2. Term provenance, and a necessary terminological clarification

The term Cognitive-Conditioning AI Applications (CCAA) was coined and defined by TopXEA Research and first published in this article on 1 September 2026. This article is the term's first publication. Suggested citation:

TopXEA Research (1 September 2026). CCAA: AI Rewrites the Standard, Not Just the Job. TopXEA. https://topxea.com/blog/ccaa-cognitive-conditioning-ai-applications

Use it, criticise it, improve it freely — we only ask that you cite the source. Three things must be said up front, or the term is dishonest on day one:

• "Conditioning" is used here in its everyday sense, not its technical one. In psychology, "cognitive conditioning" is already occupied: it belongs to the covert-conditioning family developed by Cautela and denotes associative learning (conditioned stimulus, unconditioned stimulus, reinforcement). The mechanism we describe is not associative learning — there is no CS, no US, no reinforcement schedule. It is psychophysical reference-point recalibration (Helson's adaptation-level theory, Parducci's range-frequency theory, and Levari et al.'s 2018 experiments). We keep the word because in ordinary language it conveys exactly the right thing — slow, long-run, unnoticed — but in a scholarly context, read the mechanism as described here, not as conditioning in the learning-theory sense.
• The acronym collides badly. CCAA carries at least two dozen other senses, the dominant one being Canada's Companies' Creditors Arrangement Act; others include Cognitive Communications for Aerospace Applications (an AIAA / NASA Glenn workshop series), two North American collegiate athletic associations, and the California Clean Air Act. CCAA is therefore effectively unusable as a search term — in formal writing, spell it out, or write "CCAA (cognitive conditioning)".
• We claim the definition, not ownership of three English words. If someone has used this phrasing with the same or a similar meaning before us, send us the source and we will add the attribution here and cede priority. You do not need to argue us into it; a link is enough.

3. Conceding first: where this is not new, and where it may be wrong

Before putting a new term on the table, it is worth saying what it owes to others and under what conditions it collapses. This section is not modesty; it is handing the reader the three sharpest knives first.

Concession 1: the mechanism is not new at all

"Shift the distribution and people's judgement of normal shifts with it" was demonstrated in the laboratory in 2018. Levari and colleagues, writing in Science, found that when blue dots became rare, participants began to see purple dots as blue; when threatening faces became rare, neutral faces began to look threatening; when unethical requests became rare, innocuous requests began to look unethical. The effect persisted even when participants were explicitly forewarned and even when they were paid to resist it.

One directional caveat must be stated honestly. Levari et al. demonstrated "stimulus becomes rare → concept expands". What we need is its mirror image: "high-quality stimulus becomes dense → threshold rises". Reference-point theory is symmetric in principle, but the published experiments ran the other way. That symmetry is a hypothesis, not a finding — and it can be tested directly with the same paradigm, which is the cheapest research proposal this article has to offer.

The earlier lineage is long: Helson's adaptation-level theory (judgement is made against a reference point pooled from the recent stimulus distribution); Parducci's range-frequency theory (which formalises how the mean and variance of a sample set the threshold); Brickman and Campbell's hedonic adaptation (a better mean becomes the new normal and satisfaction resets); Festinger's social comparison theory; Gerbner's cultivation theory; and Tinbergen's supernormal stimuli (exaggerated key stimuli out-compete natural ones).

Together these cover essentially the whole of CCAA's mechanistic claim. Any assertion that this article discovers a new psychological mechanism would be false.

Concession 2: adaptation is largely reversible, so we narrow the claim

A central result of the hedonic-adaptation literature is that adaptation is broadly symmetric and reversible. The treadmill spins; it does not relocate the runner. If CCAA is only hedonic adaptation in a new hat, its economic predictions fail in a specific way: the threshold drifts up and the dissatisfaction attenuates, so demand does not durably migrate into aesthetic medicine.

We do not dispute this objection. Our response is to narrow the persistence claim to something tighter, more measurable and more likely to be true:

Conditioning produces durable economic consequences only where the drifted standard becomes institutionally encoded.

Stimuli adapt back; institutions do not. Once a drifted standard is written into a hiring rubric, a clinic's before-and-after gallery, a platform's default photo norm, a performance-review template, an insurance underwriting standard, it is fixed — institutions do not adapt their way back. The virtue of this claim is that it is testable: institutional artefacts are archival, dated and countable.

Concession 3: "expectation threshold" has never actually been measured

Until there is an instrument, CCAA is a vocabulary, not a result. Section 8 therefore delivers a concrete instrument; until that panel returns data, every quantitative statement in this article should be read as a hypothesis.

So what is left that is new?

Three measurable parameter changes and one broken precondition.

1. Marginal cost fell to zero, so the idealised stimulus is now generated per person. Broadcast media produced one ideal and sent it to everyone; generative AI produces an ideal optimised against this particular person's responses. That is a quantitative change with a qualitative consequence, and it is measurable.
2. The reference frame became contingent. It answers you. The comparison literature identifies contingency as one of the variables driving internalisation — and a magazine cover never responded to anybody.
3. The reference set now includes the subject's own idealised image. Filters and AI self-portraits collapse comparison distance to zero: you are no longer comparing yourself to a celebrity but to a better version of yourself. The thin-ideal paradigm never tested this configuration.
4. Cultivation theory's generative condition has been broken, and nobody has said so cleanly. Cultivation assumes a shared, broadcast, non-elective symbolic environment; it is precisely that sharedness which produces mainstreaming, the convergence of worldviews toward the televised middle. Generative AI supplies a personalised, on-demand, unbounded, user-steerable distribution. This is not a new medium for cultivation — it violates cultivation's generative condition. Hence this article's sharpest and most falsifiable prediction: CCAA predicts normative divergence, not normative mainstreaming. If the next decade shows convergence in aesthetics, language and relational expectations, this core distinction is wrong.

4. Three mechanisms, their evidence grade, and their counter-evidence

CCAA is built from three mechanisms that can be tested separately. Each is given below with both supporting and contradicting evidence — the latter not for the appearance of balance, but because for two of the three, the contradicting studies have larger samples and better designs than the supporting ones.

Mechanism 1: distribution replacement

People judge "normal" by estimating a distribution from the samples they have seen. AI-generated content moves two parameters at once: the mean rises and the variance is compressed. Variance compression is the more insidious, because what it erases is not the location of "good" but the information about how wide normal is.

The strongest supporting evidence comes not from AI research but from visual adaptation. Brooks and colleagues, reviewing this literature in Perspectives on Psychological Science (2020), summarise experiments showing that prolonged exposure to bodies of a given size shifts the observer's point of subjective normality — the place where "a normal body" sits for them. The shift is bidirectional, experimentally manipulable, and modulated by attention. This is the best-evidenced version of the claim that a distribution sets the norm. Note its boundary: the stimuli were photographs, not AI output.

Direct AI-side evidence: Doshi and Hauser's pre-registered experiment in Science Advances (2024; 293 writers, 600 evaluators) found that writers given generative-AI story ideas produced work rated about 6.7% more novel and 6.4% more useful, with the largest gains for the least creative writers — while those stories were significantly more similar to one another than the human-only control group's. The magnitude must be reported: the homogenisation effect is b = 0.871 on a 0–100 cosine-similarity scale for one AI idea (b = 0.718 for five). The direction is solid; the size is small. Similarity is measured within condition (each story against the centroid of the others in its own condition), not across conditions, and the authors claim no population-level or longitudinal diversity collapse.

Counter-evidence (required reading): Ashkinaze and colleagues (ACM Collective Intelligence, 2025), using 800+ participants across 40+ countries in a dynamic chained design in which each wave sees the previous wave's output, found that high AI exposure did not change individual creativity and increased collective idea diversity and its rate of change. That design is larger and more ecologically dynamic than Doshi and Hauser's, and it points the other way. Mechanism 1 is contested, not settled.

Mechanism 2: frictionless reinforcement

Conditioning requires repetition, immediate feedback and low cost. AI supplies all three: always available, unlimited, never refusing, never tired, no social price. Human feedback always carries friction. When the frictionless side becomes the main training set, tolerance for friction should fall.

Evidence: Fang and colleagues' four-week study (arXiv preprint 2503.17473, not peer-reviewed; 981 completers, 28 days, 300,000+ messages, pre-registered on AsPredicted) used a 3×3 design of text / neutral voice / engaging voice by open-ended / non-personal / personal tasks. Result: no significant effects were detected from the experimental conditions; but across all of them, higher mean daily use predicted more loneliness (β = 0.02, p = 0.027), less offline socialisation (β = −0.05, p = 0.0019), more emotional dependence (β = 0.06, p < 0.001) and more problematic use (β = 0.02, p = 0.017).

Three qualifications that must always travel with that sentence: (i) this is correlation, not causation — usage was self-selected; (ii) the effects are very small, β between 0.02 and 0.06; (iii) this was not an AI-companion app but general-purpose ChatGPT with an assigned daily task.

Counter-evidence (required reading): De Freitas and colleagues, in the Journal of Consumer Research (2026), report across multiple studies that AI companions reduce loneliness, at a magnitude comparable to human interaction and above activities such as watching video. Smith, Bradbury and Karney, reviewing the field from relationship science in Perspectives on Psychological Science (2025), state directly that the empirical base for downstream effects on human relationships is largely absent.

The sentence that matters most: no experiment has ever measured how exposure to AI changes a person's expectations of human partners or human service workers. Mechanism 2 is a hypothesis, not a finding. We keep it, flagged as the softest of the three.

Mechanism 3: register assimilation — the best-evidenced of the three

AI writing tools are replacing socially marked speech with an institutional neutral register. Two ways of saying the same thing to a manager:

• "The following two items require your decision."
• "Mind casting an eye over these two?" / "Sorry to bug you — could you make the call on these?" / "These two need your signature."

The first is not more professional; it is less socially situated. Register carries relational information: distance, hierarchy, how large a favour is being asked, how urgent this really is. Flattening it costs you not elegance but the ability to signal weight through wording — so you compensate elsewhere: another meeting, an earlier deadline, more people on cc. Total communication cost does not fall; it moves from the language into the process.

The evidence chain here is the most complete:

• Causal layer: Jakesch and colleagues (CHI 2023; 1,506 participants) had people write with an opinionated language model. What they wrote shifted toward the model's position (mean opinion difference 0.29, d ≈ 0.5), and their subsequently self-reported attitudes shifted too (d = 0.22, p < 0.001) — "latent persuasion". The authors themselves note that the model's opinion was deliberately made strong, the topic was low-entrenchment, and the attitude shift was likely transient.
• Scale layer: Kobak and colleagues (Science Advances, 2025) analysed over 15 million PubMed abstracts (2010–2024) and inferred, from an abrupt post-2022 spike in excess word frequencies, that at least 13.5% of 2024 abstracts were LLM-processed, rising to 40% in some subcorpora — a larger shift than COVID produced in academic writing. Liang and colleagues (Nature Human Behaviour, 2025) analysed 950,965 papers and estimated LLM-modified text at up to 17.5% in computer science.
• Relational layer: Hohenstein and colleagues (Scientific Reports, 2023) ran two randomised experiments (n = 1,036): smart replies increased efficiency and positive emotional language and improved evaluations of the conversation partner — but when AI use was suspected, evaluations turned negative.

The honest limit: these studies establish imitation and diffusion — language really is converging on model style. They do not establish that anyone's acceptance threshold has moved. Separating those two is the problem Section 8 exists to solve.

5. The test: two codeable axes, and a correction to our own rule

For the framework to be reusable rather than merely listenable, the axes need codeable definitions. Two axes, four binary items each, scored 0–4.

Axis 1 — Acceptance Anchoring (AA)

Ask: when the buyer says "not good enough", can the dispute be settled without reference to the buyer's comparison set?

1. A written, third-party-enforceable specification exists (statute, building code, clinical guideline, numeric SLA).
2. A pass/fail event is observable by a third party within a bounded time (inspection, assay, the pipe leaks or it does not).
3. Failure carries liability that attaches regardless of satisfaction (licence, malpractice, criminal).
4. Payment or renewal is conditioned on that test rather than on a satisfaction rating.

Anchored ≥3; Elastic ≤1; =2 is a declared contested band — reported as unstable, never forced to one side.

Axis 2 — Deliverable Automatability (DA)

Ask: can a current model produce the accepted deliverable end to end — the deliverable, not the tasks?

1. The deliverable is transmissible as bits, with no physical manipulation of the client's body or property at delivery.
2. At least 70% of the occupation's O*NET task-time items are demonstrably model-performable at current cost.
3. There is no legal requirement for a human signature or presence, and no measurable willingness-to-pay premium conditioned on human authorship.
4. Liability is absent, or insurable without a named human.

High ≥3; Low ≤1; =2 enters the same contested band.

One structural point deserves emphasis: embodiment lives inside DA item 1, and nowhere else. Being physical is a reason your deliverable resists automation; it is not independently a reason your standard resists drift. Placing it in one axis only removes the collinearity that arises when embodiment is smuggled into both.

Modifier 1: anchor hardness — a correction to our own earlier rule

We previously stated the rule as "work with a non-negotiable external acceptance test is protected". That rule is false, and the counterexample is immediate: a junior programmer has a compiler, has unit tests, has production incidents — genuine external acceptance. But a model passes those tests more cheaply than the human does. Far from protecting the practitioner, the external test becomes the automation's certificate of acceptance. The correct statement is:

An acceptance test protects a practitioner only in proportion to what it costs the automating system to pass it.

Soft anchors — those a model passes at or below human cost (compilers, unit tests, actuarial reserve tests, standardised examinations) — provide no protection. Hard anchors — those whose passage requires physical presence, licensure, or a person with something at stake (site inspection, malpractice exposure, structural sign-off) — do. This is the framework's most load-bearing repair.

Modifier 2: provenance premium

Is there a measurable willingness to pay for the fact that a human did it? VTuber agencies and signature architects both live on this, and it appears in neither AA nor DA, so it must be coded separately.

Coding rules (for anyone reusing this)

1. Split revenue lines before coding. Decompose the occupation; if two lines differ by ≥2 on AA or DA, plot the occupation as a segment, not a point, and report it revenue-weighted. Aesthetic medicine, tattooing and wedding photography all require this.
2. When AA = 2, ask who pays for failure. A defined penalty independent of customer mood → anchored. The client blames themselves → elastic.
3. When DA = 2, ask what breaks first, capability or permission. If permission, code DA-low and tag it gated — gated cells flip fastest and are the framework's highest-variance predictions, to be flagged rather than averaged.
4. Re-code annually and report the delta, not just the level. Automatability is not a constant, and an undated matrix rots.

Reliability protocol: eight binary items, three independent coders, a 40-occupation sample, target Krippendorff's α ≥ 0.7, and publish the disagreement list rather than a single headline α.

6. One diagram: a 2×2 plus a set of arrows

The two axes cross into four cells. But the cells are not the important part — the arrows between them are.

	Anchored (AA high)	Elastic (AA low)
Automatable (DA high)	Anchored but squeezed (soft anchors fail: junior devs, actuaries, general translation)	Elastic and automatable (content farms, template design)
Not automatable (DA low)	Insulated (plumbing, metrology, funeral services, nursing)	Elastic but not automatable (aesthetic medicine, counselling, companionship, luxury experience)

Loop position: what our second matrix saw and the first could not

This framework began with the 2×2 above, and it had an obvious gap: it cannot see who manufactures the reference frame. A second matrix we subsequently drew was aimed squarely at that gap. It divides industries by "CCAA relevance (dependence on the subconscious, aesthetics, emotional value and embodied experience) × AI efficiency substitutability", and it places virtual influencers, AI-generated short drama and immersive entertainment in a category of their own — a "super-evolution track" — noting that these industries are the source that manufactures AI perfection.

That observation is correct, and the 2×2 above cannot produce it. But turning it into a second pair of coordinate axes fails — our own second matrix is wrong on this point — for two reasons, each with a counterexample:

• Counterexample 1 (the vertical axis does not hold): the professional athlete. Scored on "depends on the subconscious, aesthetics, emotional value and embodied experience", professional sport is at the ceiling; but its adjudication is the scoreboard — one of the hardest external acceptance tests in existence. The same object lands at opposite extremes under the two definitions of the vertical axis, which shows that "emotional/embodied dependence" cannot stand in for "adjudication mechanism". Nurses and reconstructive surgeons have the same shape.
• Counterexample 2 (role is not an industry property): wedding photography. It is simultaneously an emitter (this year's portfolio becomes next year's reference), a beneficiary (clients demand physically impossible results and pay for them) and squeezed (the retouching line is automated). All three at once — so role can only be coded per revenue line, not per industry. And it is a directional flow, whereas an axis can only carry magnitude.

The correct treatment is therefore: keep one 2×2, and render what the second matrix saw as arrows between the cells. The four loop positions are:

Loop position	Meaning	Typical revenue lines
Emitter	Its output circulates as other markets' reference material	image models, beauty-filter platforms, VTubers, AI short drama, companion apps
Beneficiary	The drifted standard exceeds what the cheap channel can physically supply, so demand migrates here	aesthetic medicine, high-end spa, counselling, companionship services
Squeezed	The standard drifted and the deliverable is automatable	illustrators, junior developers, general translators
Insulated	Anchored acceptance; not a participant in the reference-frame game	plumbing, metrology, funeral services, lift maintenance

Draw the arrows and the chain appears: emitters manufacture the reference frame → it strikes beneficiaries (demand rises) and the squeezed (the standard rises while their own work is automated) → the insulated are not on the chain at all. The second matrix's four quadrants are, in substance, these four positions — and they are not four parallel boxes but four links in one supply chain. That reframing is the only genuinely new thing produced by merging the two matrices.

7. Stress test: 18 industries, and the 8 where the framework fails

The headline first: 8 of 18 test cases (about 44%) code unstably. That number belongs here, not in a footnote. A framework that displays only its successes has no diagnostic value.

Industry	AA	DA	Loop	Stability
Aesthetic medicine	safety 4 / satisfaction 0	1	Beneficiary	⚠ flagship case, worst coding
Tattoo artist	1	design 3 / application 0	Emitter+Beneficiary / Squeezed	⚠ dual role, must split
Plumber	4 (hard)	0	Insulated	stable
Dentist	restorative 4 / cosmetic 1	0	Insulated (+ beneficiary tail)	stable
VTuber agency	0	3, but provenance premium	Emitter	⚠ see below
Junior programmer	2–3, soft	4	Squeezed	⚠ refutes the old rule
Radiologist	4 (hard)	capability 4 / gated	pending	⚠ highest variance
Translator	certified 3 / general 0	4	Gated / Squeezed	stable
Wedding photographer	0	capture 1 / edit 4	all three at once	⚠ must split
Nurse	4 (hard)	0	Insulated	stable; see prediction below
Luxury concierge	0	presence 0 / information 4	Beneficiary / Squeezed	stable
Actuary	3–4, soft	gated	pending	⚠ soft-anchor risk
Personal trainer	1 (arguably 2)	programming 3 / presence 0	Beneficiary / Squeezed	⚠ contested-band case
Funeral director	3	0	Insulated	stable
Signature architect	structural 4 / design 0	concept 4 / stamp gated	Emitter+Beneficiary	⚠ prestige inversion
Companionship services	0	0	Beneficiary	⚠ sign indeterminate
Professional athlete	4 (hardest)	0	Emitter	stable
Content-farm writer	0	4	Squeezed	stable but over-determined

The failures worth spelling out:

• VTuber agencies and signature architects are broken by the same variable. The first is legally replaceable by AI, yet fans pay a premium for a persistent identity being someone; the second has an elastic standard and high automatability and should therefore be squeezed, but is protected by prestige. When one patch fixes two cells, it is not a patch but a missing variable — hence provenance premium is promoted to a formal modifier. Whether that premium is stable is an open empirical question the matrix cannot settle.
• Companionship services are the framework's most confident prediction and its hard boundary. Whether AI companionship is a complement (raising the affective baseline and pushing people toward human services) or a substitute (satisfying the need in-channel) is not determined by the two axes; it is exogenous. For high-conditioning, low-automatability services the framework therefore gives magnitude but not direction, until complementarity is measured separately. This belongs in the abstract.
• Nursing yields the cheapest free prediction in this article. Nursing behaviour is auditable and roughly constant; but if patients' reference frames are reset by an infinitely patient chatbot, then patient-satisfaction scores (HCAHPS-type instruments) should decline year on year for unchanged nursing behaviour. Constant behaviour, drifting rating — this is very nearly an off-the-shelf threshold measurement, and the data are public.

8. How to falsify it: the constant-stimulus drift panel

The three falsification conditions we previously proposed are all unusable. Here is why, and what replaces them.

Why the original three fail

• "Does the AI share of reference photos brought to aesthetic clinics rise?" — unusable. No denominator, no baseline, no threshold; clinics do not archive reference images and consent blocks retrospective collection; AI detection on compressed social images is poorly calibrated. And a fatal confound: the AI share of the entire image supply is rising anyway, conditioning or not. Observing that patients bring AI-retouched photos would prove only that AI-retouched photos exist.
• "Do heavy AI-companionship users show less tolerance for friction?" — currently uninterpretable. Controlling for baseline loneliness does not fix selection, because adopters self-select on the slope, not the level. Worse, self-reported "tolerance for friction" is contaminated by the very construct being measured — a reference-group effect in which the yardstick moves with the thing it measures. And the sign is ambiguous: therapeutic use might plausibly increase tolerance.
• "Do register metrics in workplace writing change?" — measures the wrong thing. Even total convergence of workplace prose onto model style demonstrates imitation and diffusion, not that anyone's acceptance threshold moved. Add reverse causality (models are trained on that corpus) and same-window confounds (remote work, chat-ification, cohort turnover, template adoption), with no control group.

The replacement: freeze the stimulus, rotate the raters

Take 40 human-written memos and emails from 2019. Each year, a fresh blind cohort of managers rates those identical texts against a fixed rubric: "acceptable to send" versus "would demand revision". The primary outcome is the year fixed effect on demanded-revision probability for unchanged text.

Because the stimulus is constant by construction, any drift can only be the standard. This single design does more than all three original conditions combined, and it is the difference between a framework and a finding. The same design ports: frozen face panels for aesthetic medicine, frozen conversation transcripts for companionship, frozen text panels for the workplace — with a dose-response test across domains of differing AI exposure.

Accompanying identification strategy: identify off availability, not usage (regulator-forced feature withdrawals and app-store policy shocks as exogenous variation, with difference-in-differences); use encouragement-design randomisation with encouragement as an instrument; and anchoring-vignette-adjust every self-report scale, or the reference-group effect will contaminate the entire literature.

9. Why a second axis is needed: the existing exposure measures do not contain one

Measuring an occupation's AI exposure is a developed field. Eloundou and colleagues (Science, 2024) is the most-cited instrument: about 1.8% of jobs have more than half their tasks affected by LLMs alone; just over 46% once LLM-complementary software is assumed. (A correction in passing, since this is the most common citation error in the literature: the widely quoted "80% of workers with at least 10% of tasks affected, 19% with at least half" comes from the 2023 arXiv preprint, not the Science article. The two must not be cited together.) Alongside it sit Felten, Raj and Seamans' AIOE ability-linkage measure (Strategic Management Journal, 2021) and Webb's 2019 working paper, still unpublished.

These measures share a blind spot their own authors acknowledge: every one of them answers only "can this task be done", and not one contains a variable for who adjudicates whether it was done well. Not AIOE, not Webb, not Eloundou. That is the entire justification for a second axis.

The blind spot has surfaced once already in mainstream economics. Acemoglu (Economic Policy, 2025) puts AI's ten-year total-factor-productivity effect at no more than 0.71%, and below 0.55% once hard-to-learn tasks are accounted for — and his stated reason for the downgrade is that hard-to-learn tasks have "many context-dependent factors affecting decision-making and no objective outcome measures from which to learn successful performance". That is very nearly our acceptance-anchoring axis, except that he treats it as a reason AI will be less productive rather than as a reason those occupations are protected. Autor (NBER WP 32140, 2024) identifies the gate around expert work as licensure and liability, not task difficulty.

Labour-market effects observed so far, each with its boundary:

• Hui, Reshef and Zhou (Organization Science, 2024): on an online freelance platform after ChatGPT, highly affected occupations saw roughly −2% jobs and −5.2% monthly earnings, with no evidence that strong past performance protected anyone.
• Demirci, Hannane and Zhu (Management Science, 2025): within eight months, job posts fell about 21% for automation-prone writing and coding work, and about 17% for image-creation work after image-generation AI; surviving posts were more complex and better paid.
• Brynjolfsson, Li and Raymond (Quarterly Journal of Economics, 2025): across 5,179 customer-support agents, a generative-AI assistant raised issues resolved per hour by 14% on average (not the widely quoted 15%), 34% for novices, and almost nothing for experienced high performers.
• Noy and Zhang (Science, 2023): in a pre-registered RCT with 453 college-educated professionals, writing time fell 40% and rated quality rose 18%.
• Brynjolfsson, Chandar and Chen, "Canaries in the Coal Mine?" (revised August 2026, ADP administrative payroll data): a relative employment shortfall of about 19% for workers aged 22–25 in the most AI-exposed occupations, operating through reduced hiring rather than layoffs. The authors explicitly call these early descriptive indicators, not causal estimates.
• Counter-evidence: Frank and colleagues (2026), using unemployment-insurance administrative records and millions of LinkedIn profiles, find that unemployment risk in AI-exposed occupations began rising in early 2022, months before ChatGPT's release. ChatGPT's launch is therefore not a clean natural experiment; post-2022 weakness confounds AI diffusion with monetary policy and post-pandemic sectoral adjustment.

10. Evidence on the beneficiary side: how hard is the aesthetic-medicine case really?

Aesthetic medicine is the flagship case for "a drifted reference frame raises demand". Its evidence is weaker than most people assume, and more interesting.

• The strongest single citation: a three-level meta-analysis in Mass Communication and Society (2026) covering 24 articles, 96 effect sizes and 10,445 participants, finding a correlation between social media use and cosmetic-surgery consideration of r = 0.21, 95% CI [0.13, 0.26], moderated upward by appearance-focused use. Note that it is a correlation, and small-to-moderate in size.
• Industry data supplies an inconvenient detail: the ISAPS Global Survey for 2024 reports 37.9 million aesthetic procedures worldwide (17.4m surgical, 20.5m non-surgical) and about +42.5% growth over four years — but 2024 fell against 2023: surgical −6.7%, non-surgical −3.1%. If conditioning-driven demand were monotonic, that decline needs explaining. At minimum it shows that "AI rewrites beauty standards, therefore aesthetic-medicine demand explodes" is too simple a narrative; macro cycles, price and regulation plausibly dominate reference-frame drift.
• Correct attribution for "Snapchat dysmorphia": the term was not coined by JAMA Facial Plastic Surgery. Its first peer-reviewed appearance is Ramphul and Mejias's editorial in Cureus in March 2018; press attribution for the coinage goes to the UK cosmetic doctor Tijion Esho. The Rajanala, Maymone and Vashi piece in JAMA Facial Plastic Surgery 20(6):443–444 is a Viewpoint that medicalised and popularised an already-circulating term — its own wording is "dubbed". This corrects an error in an earlier version of this article.
• Discount the 55% figure: the widely circulated "55% of surveyed surgeons had seen patients seeking procedures to look better in selfies, up from 42%" comes from an AAFPRS annual member survey press release. The sampling frame is association member surgeons responding voluntarily; the sample size, response rate and sampling method have never been published, and no peer-reviewed survey document exists. Usable as evidence that the phenomenon occurs; unusable as a population rate.
• "Emotional labour gets more expensive" currently has no evidence. Hochschild's The Managed Heart (1983) established the concept, and peer-reviewed work documents rising customer incivility (for instance, comparisons before and after COVID showing the indirect effect of customer incivility on performance via emotional exhaustion becoming more pronounced). But we could find no study establishing that exposure to technology or AI raises customers' expectations of human service providers. This inference must be presented as a hypothesis with a research design attached, not as an established finding.

11. Two extensions explicitly labelled as inference

The coast-to-interior lag

CCAA exposure probably spreads with AI adoption: it appears first where adoption is dense and later where it is thin, and the lag is a window. The new demand already forming in dense areas looks like this: requests to turn a filtered face back into a real one; weaning off AI companionship and repairing human relationships; a premium for human-written content and verified-human identity; expectation-management training for beneficiary-side practitioners. What has not been repriced where adoption is thin is the insulated cell.

The mandatory discount: this rests on an untested assumption — that CCAA spreads at the same rate as AI adoption. That assumption is quite possibly wrong, because cognitive effects travel through content and content circulates nationally rather than by geographic gradient. Use it as a hypothesis, not a plan.

The emperor experience, reaching the employment relationship

An AI assistant is, for every user: infinitely patient, always available, entirely centred on you, never talking back, never holding a grudge, never tired. Historically only a very small number of people experienced a relationship of that shape, and societies maintained an apparatus of norms marking it as privilege, not default. It is now everyone's daily experience, dozens of times a day.

Expected spillovers (inference, not a verified conclusion): higher expectations of service workers; lower tolerance for friction with colleagues and family; "handled by a human" moving from default to paid upgrade; a lower threshold for escalating a complaint. This is the same phenomenon as the nursing prediction above, seen from the other side — and since the nursing side has public data, that is the side to test first.

12. Back to our own industry: trading outcomes are a hard anchor, trading marketing has none

Trading is an extreme case of a hard anchor. Whether an equity curve is good is not aesthetic; it is the number in the account. The market settles daily, ignores appearances, hears no appeals — and a model does not pass that test any more cheaply than a human does. That is precisely the point of the anchor-hardness correction: trading's anchor is hard not because an acceptance test exists, but because the test is an adversarial market nobody passes cheaply.

Trading's marketing, by contrast, has no anchor at all. A beautiful backtest chart, a polished profit screenshot, an AI-generated "mentor" persona — production cost is now zero. The consequence is a break:

"Looks professional" used to be a real filtering signal, because it cost money, time and craft. It now costs nothing. The correlation between "looks professional" and "actually makes money" has been severed.

There is one defence: replace impression with verifiable evidence — third-party hosted live account links, quarter-by-quarter filed figures, source code you can download and read line by line, and drawdown published alongside the gains.

• How to audit a live account properly: verifying an EA with Myfxbook
• The scam checklist: how to spot an EA scam
• Why we publish our own ugly numbers: why we publish our drawdown
• What a flawless curve can hide: opening the source to see whether the curve already knew the answers
• One question, six answers: the six-fold split on AI server market size — that piece is about the reference frame for a number, this one about the reference frame for people.

13. What this framework must not be used for

Boundary statement, to be read alongside everything above:

1. It does not predict employment levels. Industry employment outcomes are dominated by demand, demographics, price and regulation; conditioning is a small residual. Using this 2×2 to forecast how many jobs an industry will have is a misuse.
2. It predicts the direction of standard drift only, and durable consequences only where institutional encoding occurs.
3. For high-conditioning, low-automatability services — companionship above all — it gives magnitude but not direction, until complementarity between AI and human provision is measured separately.
4. Eight of eighteen stress-test cases code unstably. That is not a blemish to be polished out; it is an accurate reflection of the current evidence base.

Closing

CCAA is not a pessimistic term, and it is not a new mechanism. It says this: half of what AI changes is not on the "who gets replaced" axis but on the "who defines the standard" axis — and every existing exposure measure covers only the first half.

Miss the second half and you reach the conclusion "AI can't do my job, so I'm safe" — which is exactly what everyone on the beneficiary side believes, while the reference frame they are judged against is being rewritten. Conversely, the question that decides whether a job is really protected is not "is there an acceptance test" but "how expensive is that acceptance test for a machine to pass".

The rest is measurement. Until the constant-stimulus drift panel returns its first data, this article is a vocabulary and a set of hypotheses — published together with the conditions under which it should be discarded, in the hope that somebody goes and discards it.

Principal references

1. Levari, D. E., Gilbert, D. T., Wilson, T. D., Sievers, B., Amodio, D. M., & Wheatley, T. (2018). Prevalence-induced concept change in human judgment. Science, 360(6396), 1465–1467. doi:10.1126/science.aap8731
2. Helson, H. (1964). Adaptation-Level Theory. Harper & Row. / Parducci, A. (1965). Category judgment: A range-frequency model. Psychological Review, 72(6), 407–418.
3. Brooks, K. R., Mond, J., Mitchison, D., Stevenson, R. J., Challinor, K. L., & Stephen, I. D. (2020). Looking at the Figures: Visual Adaptation as a Mechanism for Body-Size and -Shape Misperception. Perspectives on Psychological Science, 15(1), 133–149. doi:10.1177/1745691619869331
4. Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. doi:10.1126/sciadv.adn5290
5. Ashkinaze, J., Mendelsohn, J., Qiwei, L., Budak, C., & Gilbert, E. (2025). How AI Ideas Affect the Creativity, Diversity, and Evolution of Human Ideas. Proc. ACM Collective Intelligence Conference, 198–213. doi:10.1145/3715928.3737481
6. Jakesch, M., Bhat, A., Buschek, D., Zalmanson, L., & Naaman, M. (2023). Co-Writing with Opinionated Language Models Affects Users' Views. CHI '23, Article 111. doi:10.1145/3544548.3581196
7. Kobak, D., González-Márquez, R., Horvát, E.-Á., & Lause, J. (2025). Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Science Advances, 11(27), eadt3813. doi:10.1126/sciadv.adt3813
8. Liang, W., et al. (2025). Quantifying large language model usage in scientific papers. Nature Human Behaviour, 9, 2599–2609. doi:10.1038/s41562-025-02273-8
9. Hohenstein, J., et al. (2023). Artificial intelligence in communication impacts language and social relationships. Scientific Reports, 13, 5487. doi:10.1038/s41598-023-30938-9
10. Fang, C. M., et al. (2025). How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study. arXiv:2503.17473 (preprint).
11. De Freitas, J., Oğuz-Uğuralp, Z., Uğuralp, A. K., & Puntoni, S. (2026). AI Companions Reduce Loneliness. Journal of Consumer Research, 52(6), 1126–1148. doi:10.1093/jcr/ucaf040
12. Smith, M. G., Bradbury, T. N., & Karney, B. R. (2025). Can Generative AI Chatbots Emulate Human Connection? A Relationship Science Perspective. Perspectives on Psychological Science, 20(6), 1081–1099. doi:10.1177/17456916251351306
13. Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306–1308. doi:10.1126/science.adj0998
14. Felten, E. W., Raj, M., & Seamans, R. (2021). Occupational, industry, and geographic exposure to artificial intelligence. Strategic Management Journal, 42(12), 2195–2217. doi:10.1002/smj.3286
15. Hui, X., Reshef, O., & Zhou, L. (2024). The Short-Term Effects of Generative Artificial Intelligence on Employment. Organization Science, 35(6), 1977–1989. doi:10.1287/orsc.2023.18441
16. Demirci, O., Hannane, J., & Zhu, X. (2025). Who Is AI Replacing? The Impact of Generative AI on Online Freelancing Platforms. Management Science. doi:10.1287/mnsc.2024.05420
17. Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at Work. Quarterly Journal of Economics, 140(2), 889–942.
18. Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. doi:10.1126/science.adh2586
19. Acemoglu, D. (2025). The simple macroeconomics of AI. Economic Policy, 40(121). / Autor, D. H. (2024). Applying AI to Rebuild Middle Class Jobs. NBER Working Paper 32140.
20. Gerbner, G., & Gross, L. (1976). Living with Television: The Violence Profile. Journal of Communication, 26(2), 172–199. / Festinger, L. (1954). A theory of social comparison processes. Human Relations, 7(2), 117–140.
21. Ramphul, K., & Mejias, S. G. (2018). Is "Snapchat Dysmorphia" a Real Issue? Cureus, 10(3), e2263. doi:10.7759/cureus.2263 / Rajanala, S., Maymone, M. B. C., & Vashi, N. A. (2018). Selfies—Living in the Era of Filtered Photographs. JAMA Facial Plastic Surgery, 20(6), 443–444. doi:10.1001/jamafacial.2018.0486
22. Hochschild, A. R. (1983). The Managed Heart: Commercialization of Human Feeling. University of California Press.

Term provenance: the term Cognitive-Conditioning AI Applications (CCAA) was coined and defined by TopXEA Research, first published in this article on 1 September 2026. In this article "conditioning" is used in its everyday sense of habituation and baseline recalibration, not in the technical sense of associative or covert conditioning.

Risk and disclosure: this article is a discussion of a concept and a method. It is not investment advice, career advice or medical advice, and it targets no specific organisation or individual. External research is cited with its source and evidence grade; passages labelled "inference" or "hypothesis" are our judgement and are not verified. Trading carries risk, every EA and quantitative strategy can lose money, and forex and precious-metals trading is high risk — only trade with money you can afford to lose.