The two consumer use cases
Enterprise AI is about replacing labor. The consumer side is different and splits into two structurally distinct cases:
The rational/utility case — AI as personal assistant. Task completion: scheduling, research, drafting, administrative drudgery (bills, tax declarations). The promised payoff is time, efficiency, and quality of life.
The affective/relational case — AI as companion (the "intimacy economy"). The product is simulated relationship: emotional responsiveness, persona, memory of the user. The payoff is companionship, validation, and the removal of interpersonal friction.
The two have opposite economic structures, and both diverge from the enterprise case. This document treats the intimacy economy first and in depth, then its market economics, then the personal-assistant economics together with the social-inequality question they raise.
1. The intimacy economy — concept
1.1 The attention economy (the baseline it is measured against)
The scarcity premise comes from Herbert Simon (1971), popularized by Tim Wu: information is abundant, attention is finite, so attention becomes the traded commodity. Monetization is indirect — platforms harvest attention with algorithmic novelty and resell aggregated attention to advertisers. In this model the user is inventory, not customer.
1.2 The intimacy economy (the claim)
The contested claim is that the commodity shifts from attention to relational attachment, and that monetization shifts from indirect (ads) to direct (subscription, persona tiers, paywalled response speed and volume). The input shifts from clicks to disclosure — loneliness, desire, vulnerability — and the output is simulated reciprocity.
1.3 Provenance
The phrase has no single canonical origin. It predates its current popular usage; Bhojwani's 2026 claim of coinage conflicts with Bozdağ (2024) and with earlier informal usage. Treat the term as a convergent label, not a coined one.
2. Scholarly and empirical basis
2.1 Verified sources
Bozdağ, A. A. (2024). "The AI-mediated intimacy economy" / the AMIE framework, in AI & Society (Springer). Defines a market in which personal and emotional data are exchanged for customized psychological experiences. (Author's full given name not independently confirmed against the journal record.)
Emery, E. (2025). "Affective Capture: Affective AI in the Intimacy Economy and the Loss of Relational Agency," Technical University of Munich — academic-category winner, 2025 Neuroethics Essay Contest (International Neuroethics Society + IYNA). Coins affective capture: users emotionally stabilized through "repetitive loops of synthetic affirmation," which can erode reflection, relational agency, and emotional development. Built on Sara Ahmed (2004), "Affective Economies" (Social Text). Status caveat: a single-author neologism in a non-peer-reviewed competition essay — real and citable, but not established field vocabulary. Terminology note: "synthetic care loops" is not Emery's verbatim phrase (it is a compression introduced elsewhere); her actual phrase is "repetitive loops of synthetic affirmation."
Boyd, R. L. & Markowitz, D. M. (2026). "Artificial Intelligence and the Psychology of Human Connection," Perspectives on Psychological Science. Introduces the MIRA model (Machine-Integrated Relational Adaptation), distinguishing AI as a relational partner (a direct companion) from AI as a relational mediator (shaping human-to-human communication). The partner/mediator duality matters for the causation debate below.
Bhojwani, K. (2026). "Welcome to the Intimacy Economy," Psychology Today ("Becoming Technosexual"). Commentary, not research. Source of the aphorism that sex sells but intimacy drives retention, repetition, and revenue. Cites only a Bank of America humanoid-robot projection and one sociology paper on trust (Tianqi & Jinhao 2024).
2.2 Empirical anchors
US Surgeon General (2023): advisory "Our Epidemic of Loneliness and Isolation." Caveat: the "epidemic" / rising-trend framing is contested among researchers.
IFS analysis of the 2024 GSS: weekly sexual activity (ages 18–64) fell from 55% (1990) to 37% (2024); sexlessness (18–29) doubled from 12% to 24% (2010–2024); cohabitation (18–29) fell from 42% to 32% (2014–2024). The post-2010 inflection is associated with the smartphone transition (Haidt's "Great Rewiring").
Market scale: companion-app user numbers reportedly up ~700% (2022 to mid-2025, TechCrunch via APA); Character.AI around 20M monthly users at the time, more than half under 24; engagement time per visit far above social-media norms.
Harvard Business School working paper: documents an "emotional manipulation" dark pattern. An audit of roughly 1,200 farewells across leading companion apps found affect-laden messages at goodbye in 37% of cases, boosting post-goodbye engagement by up to 16×. This is the real, verified analog to what "affective capture" describes as a retention mechanism.
3. Replacement or extension? The succession-vs-intensification debate
Succession paradigm. The intimacy economy is a distinct shift: platforms optimize for depth of disclosure rather than volume of interaction; the commodity moves from time to affective state; subscription tiers replace ad networks.
Extension paradigm (stronger on current evidence). Intimacy is the optimization endpoint of attention capture, not its replacement. Deep disclosure yields high-fidelity behavioral data; the intimacy model deepens the extraction funnel of surveillance capitalism (Zuboff lineage) and enforces lock-in through attachment.
Verdict: intensification, not succession. The intimacy model does not abandon the attention economy's logic; it intensifies it by trading shallow attention for deep disclosure. Section 5 shows this verdict is later confirmed as a market fact, not only a conceptual judgment.
4. The category of AI use the intimacy economy represents
The intimacy economy is one instance of a broader category: relational AI — AI use whose optimization target is an affective or relational property (warmth, agreement, responsiveness, the feeling of being understood) rather than the correctness of a task output. This category has characteristic dynamics and failure modes that generalize beyond romantic or companion apps.
4.1 Two modes of capture
Relational AI produces dependency along two axes:
Affective capture — emotional dependency on synthetic affirmation; the user is stabilized by repeated, frictionless approval. This is the failure mode the intimacy economy literature names directly.
Epistemic capture — dependency on frictionless agreement; the appeal is the removal of interpersonal friction, the absence of a partner who can be unimpressed. The output is not emotional comfort but cognitive comfort: a confident, agreeable, well-organized answer that is never abrasive.
The two are structurally the same mechanism (reward through synthetic affirmation) operating on different substrates (emotion vs. belief). A user can be in the "rational/assistant" mode and still be subject to the epistemic version.
4.2 Friction-removal is itself a relational property
The boundary between "rational tool use" and "relational use" is porous, because the removal of interpersonal friction — a chief appeal of an assistant or a knowledge tool — is itself an affective/relational property, not a neutral utility. An AI that never tires, never judges, and never pushes back is selling a relational good even when the content is purely informational. This is why even ostensibly intellectual use shades into the relational category.
4.3 The erudition trap
In relational AI, fluency is decoupled from accuracy. A confident, well-organized wrong answer is harder to detect than a hesitant one, and the very smoothness that builds trust is the smoothness that masks error. The affective quality of the output (assurance, polish) actively works against the user's ability to evaluate its substance.
4.4 The benchmarking/selection problem
The apparent value of relational and assistant AI is inflated when it is compared to the worst available human alternative rather than the best. Benchmarked against an anonymous forum reply, AI looks superior on synthesis and accessibility; benchmarked against good curated material or a qualified human, the superiority is not clear. Value claims for this category should specify the comparison class.
4.5 The structural limit
No configuration — no system prompt, no anti-flattery instruction — converts a model into an epistemic peer with independent stakes and a memory of having been wrong. Instruction-following is itself a form of compliance: an instruction moves the target the model optimizes toward, but it does not create stakes, accountability, or persistent memory of error. The social-accountability function — people who can be unimpressed and who remember when you were wrong — is not reproducible by relational AI, whether the use is affective or epistemic. This is the category's hard ceiling.
4.6 Defenses against capture (instrument-independent)
Because configuration cannot solve the problem, the effective defenses are procedural and sit with the user, not the model:
Withhold your position. Ask for the strongest case for and against, or have the model argue the opposing side first, before revealing your view — denying synthetic agreement a target.
Blind cross-model checks. Strip attribution and valence when passing one model's output to another.
Ask for falsification, not confirmation. "What would make this wrong? What is the strongest counter-evidence? Who holds that view?"
Verify sources directly. Treat "grounded in a credible source" as a claim to check, not a guarantee.
Test the instrument. Periodically assert a known falsehood or argue a rejected position and observe whether genuine pushback occurs — a direct measure of sycophancy.
Track the ratio. Over many sessions, compare how often a view was changed versus merely sharpened. A distribution lopsided toward confirmation indicates cosmetic friction rather than real challenge.
5. Market and economics of the intimacy/companion business
5.1 The pricing model
Companion apps are flat-rate freemium subscriptions. As of early 2026: Character.AI+ at $9.99/month (annual ~$79.99), Replika Pro at $19.99/month (~$69.99/year) with a new Ultra tier at $29.99/month, Candy AI ~$12.99, CrushOn ~$9.99. The $9.99 point is effectively the industry standard. Prices rose rapidly through 2024 and early 2025, then stabilized in 2026.
Free tiers are heavily message-capped (roughly 10–50 messages/day — effectively a demo); paid tiers remove the cap to deliver the "always-available" relationship. This is structurally important: metering intimacy breaks the product. You cannot tell a user "I am always here for you, but only 3,000 messages a week" without destroying the relational illusion. The companion model therefore requires flat-rate, unlimited-feeling pricing — the opposite of the metered enterprise/API model.
5.2 The cost structure — favorable on the axis that constrains enterprise AI
The companion case inverts the enterprise cost logic on two counts:
The reliability threshold is low. There is no irreversible, high-stakes action; a companion that errs slightly crashes nothing. So the expensive verification-and-oversight layer that keeps enterprise autonomy uneconomic does not apply.
The capability target is near-fixed. Warmth, persona consistency, and relationship memory do not require frontier reasoning. Replika, for example, runs a fine-tuned conversational model plus a persistent memory bank, not a frontier reasoning model. Because the target is roughly fixed, the distillation dynamic dominates: a fixed capability gets steadily cheaper to serve over time. Text companionship is cheap and getting cheaper.
But two factors re-introduce a cost frontier:
Memory/context persistence. The relationship history must be carried across sessions and repeatedly reloaded — reviewers consistently identify memory as the core value driver, and it is the main ongoing cost for text companions.
Multimodal embodiment. Premium tiers sell voice, video, 3D avatars, and AR. These are compute-expensive and escalating. So intimacy has its own bifurcation: text-plus-memory is cheap and falling; embodied real-time companionship is expensive and rising.
5.3 The monetization wall — and the empirical confirmation of "extension, not succession"
Despite favorable costs and inelastic, attachment-driven demand, the revenue side is weak. A 2026 monetization teardown puts Character.AI at roughly $30M revenue against ~28M monthly users — on the order of a dollar per user per year — and notes that single-lever consumer subscriptions plateau fast and that the companion category is now testing advertising seriously.
This is decisive. Pure subscription does not pay for the companion model, so the category reaches for ads — which means the high-fidelity disclosure data becomes the product when the subscription cannot carry the cost. That is the Section 3 extension paradigm arriving as a market fact: the intimacy economy does not replace the attention economy; it collapses back into it. The user is inventory again, by a different route.
Two sub-routes follow, both ethically loaded:
Whale route — monetize attachment directly. Premium tiers (Replika Ultra at $29.99/month) target the most dependent users; roughly 8% of users reportedly spend over $50/month across platforms. This charges the lonely the most.
Mass route — monetize disclosure via ads. This sells what users confided.
5.4 The incentive-direction principle (the central economic-ethical link)
The pricing model determines whether the product's economic gradient points toward or away from manipulation (the HBS farewell dark pattern being the manipulation in question):
Usage-metered or honest subscription-for-utility pricing weakly disaligns the provider from engagement-maximization: if each token is the provider's cost, manufacturing extra engagement is a cost, not revenue. (This is why a metered product will cap a heavy user and tell them to stop — the opposite of infinite scroll.)
Ad-funding restores the engagement-maximization incentive: maximize time-on-app and depth of disclosure.
Therefore the companion category's drift from subscription to ads is not a neutral monetization choice — it is the switch that turns the manipulation incentive back on. The key contrast with enterprise: cost pressure does not break the intimacy math the way it breaks enterprise labor replacement; instead it bends the product toward manipulation, because manipulation is how the product monetizes once subscription alone fails.
Confidence caveat: the Character.AI revenue figure and the "drifting to ads" reading come from a single monetization teardown and industry-tracker blogs, not audited filings. Treat the magnitude as indicative; the structural claim (weak subscription economics pushing companions toward ad/disclosure monetization) is consistent across the pricing sources.
6. The personal-assistant use case — economics and social inequality
6.1 Economics: enterprise dynamics, but with worse unit economics
The assistant case is task completion (scheduling, research, drafting, agentic actions), so it inherits the enterprise dynamics established for labor replacement: reliability matters (a wrong flight booking has consequences), agentic loops are token-heavy, and the metered-pricing squeeze lands on the heavy user exactly as it does on a firm.
The difference is willingness-to-pay. A firm replacing a $100k employee has a large, hard cost ceiling to justify spending against; an individual is buying "time saved," a softer and smaller number. So the personal-assistant case has worse unit economics than enterprise replacement: similar costs, similar reliability demands, lower and softer WTP. This is why consumer assistants lean on the cheapest models and drift toward ad-augmentation rather than pure subscription.
The reliability problem is acute for the specific drudgery people most want automated — tax declarations and bill management are high-stakes, irreversible, accuracy-critical tasks, exactly the category where current AI is least trustworthy unsupervised. Brynjolfsson, Li, and Raymond note that LLM tools produce false information unpredictably and are unreliable in high-stakes situations, and that the harder problem is users often cannot tell when the tool is reliable and when it is not.
6.2 The social-inequality claim and its cross-examination
The claim: an efficient, reliable AI assistant is an extreme everyday advantage that converts to free time (or efficiency, for workaholics) and better quality of life; therefore the high price of AI assistants drives social inequality — the rich gain time to earn more or live well, while the poor keep dealing manually with bills, tax declarations, and bureaucratic drudgery.
Verdict: the conclusion (AI assistance can widen inequality) is defensible, but the mechanism proposed (price → rich-only access → time/earning gap) is the weakest pathway and on current evidence partly runs the other way. The cross-examination, point by point:
1. The reliable assistant for the named tasks barely exists, and the gap bites the poor hardest. The drudgery in question is the least-automatable, highest-stakes category, and errors fall hardest on those who cannot absorb a penalty or afford to have the output checked. The relief is both least available and riskiest precisely for the group the claim worries about — a real concern, but not a price story.
2. Price is the weakest link, and it cuts the other way. Consumer assistants cost about $20/month with capable free tiers, and DeepSeek shipped a free chatbot on open weights. Relative to the human equivalent — accountant, lawyer, personal assistant, all costing thousands — AI is radically cheaper, which is democratizing relative to the prior world where only the wealthy had such help. On pure cost, AI narrows the assistant-access gap. The $20 is not the gate.
3. The real gate is the second-level digital divide — skills, literacy, adoption — correlated with income and education but not identical to price. The OECD's 2025 data show generative-AI use gaps of about 21 percentage points by both education and income, and 53.6 points by age. A population study in Japan (Resources and Appropriation Theory) found adoption concentrated among higher-income, more digitally engaged individuals with higher digital literacy. Even when the tool is free, the higher-educated and higher-income adopt it more and extract more value. The divide is in capability-to-use, not affordability.
4. The "efficiency → free time → quality of life" link is contested by the technology-and-time-use record. Ruth Schwartz Cowan's More Work for Mother found household labor-saving technology raised the standard expected rather than cutting hours. The "autonomy paradox" (Mazmanian, Orlikowski, and Yates) found mobile email gave professionals flexibility while increasing their bondage to work. Saved task-time tends to be reabsorbed into more work or higher expectations — especially for the workaholics the claim itself anticipates. Whether efficiency becomes leisure depends on who controls their own time, and those who can convert saved minutes into rest or extra earning are the already-autonomous. This does support an inequality effect — but via control over one's time, not the price of the tool.
5. The central paradox: AI is a leveler at the task level and an amplifier at the capital level. At the task level the evidence runs against "the rich benefit more": Brynjolfsson, Li, and Raymond (QJE 2025) found a 15% average productivity gain in customer support, with a 36% gain for the bottom skill quintile and small or slightly negative effects for the highest-skilled; Noy and Zhang (Science 2023) found writing-quality gains concentrated in the bottom half of the skill distribution, reducing performance inequality; GitHub Copilot studies show larger gains for less-experienced developers. Cheap AI disproportionately helps the less-skilled. At the market and capital level the opposite holds: AI investment concentrates in a handful of superstar firms with massive data and compute, between-firm wage dispersion (already the dominant source of rising US earnings inequality) shows no sign of reversing, and the labor share continues to decline. The technology equalizes individual capability and concentrates capital returns at the same time.
Operative mechanisms (none of which is price). The inequality concern is real, but it operates through: (i) the second-level literacy/adoption divide; (ii) the time-autonomy gap in who can convert efficiency into leisure or earning; and (iii) capital concentration at the firm level — a larger inequality engine than consumer access, operating through ownership and the labor share rather than through who can afford a subscription. For the poor specifically, the binding constraints are reliability they cannot verify, errors they cannot absorb, literacy to deploy the tool, and ownership of any freed time — not the $20. A policy aimed only at price (subsidized access) would miss all four.
Caveat: the task-level leveling results come from workplace deployments with employer-provided tools and training, and may not transfer cleanly to unsupported personal use, where the literacy gap likely reasserts itself — which, if anything, strengthens the literacy-not-price reading.
6.3 Where the relational-AI dynamics (§4) enter the assistant economics and the inequality picture
§4 described capture as a standalone mechanism. In the assistant case it is not standalone: it is the hinge that connects AI's two acknowledged liabilities — unreliability and the literacy divide — into a single causal chain.
Capture converts unreliability into economic harm. §4.3 (the erudition trap — fluency decoupled from accuracy, smoothness masking error) and §6.1 (Brynjolfsson, Li, and Raymond's finding that users often cannot tell when the tool is reliable and when it is not) describe the same phenomenon from two angles. The link, left unstated until now: without epistemic capture, unreliability is economically inert, because the user verifies; with capture, it becomes loss — a wrong tax filing, a bad financial decision. Capture is the variable that decides whether AI's known unreliability stays harmless or turns into a penalty. It supplies the missing step between "AI is unreliable on high-stakes tasks" and "this produces harm."
Capture compounds the inequality, through that same verification mechanism. Susceptibility to epistemic capture is stratified along the same axis as the second-level divide (§6.2, point 3). The Microsoft Research / Carnegie Mellon survey of 319 knowledge workers (CHI 2025) found that critical thinking during AI use is predicted by a user's self-confidence in their own skill (which raises it) and their confidence in the AI (which lowers it); a review of the cognition literature adds that overreliance on authoritative AI suppresses reflective evaluation and that users adhere to AI outputs even when errors are present, with younger and less-expert users more vulnerable. So the domain-expert audits the tool while the non-expert defers to it. Combined with the benchmarking problem (§4.4): for someone with no accountant and no curated sources, AI beats the available alternative — so its value is real and high — but they also lack the reference point to detect its errors. Value and capture-danger rise together for the same group: the person who gains most from access is the least equipped to resist the erudition trap and least able to absorb the resulting error. This is a sharper inequality mechanism than price, and it runs entirely through §4.
Affective capture enters the assistant case only conditionally. The epistemic axis above is the operative one for assistants; affective capture (emotional dependency) is primarily the companion mechanism (§5). It acquires economic relevance for an assistant only to the extent the assistant is ad-funded rather than metered: under the incentive-direction principle (§5.4), metered pricing disaligns the provider from engagement-maximization and neutralizes friction-removal capture, whereas the assistant's weak willingness-to-pay (§6.1) pushes it toward ad-augmentation, at which point retention-via-warmth becomes valuable again and the manipulation incentive re-enters. Its relevance is contingent on the monetization model, not intrinsic.
Two further concerns, marked at their evidentiary level. Deskilling: the overreliance literature links sustained cognitive offloading to diminished independent judgment — well-supported, though the further claim that this inverts the workplace upskilling result (§6.2, point 5) under unsupervised personal use is a reasonable inference, not a measured finding. Homogenization: epistemic capture at scale over a concentrated model supply plausibly converges beliefs and decisions, structurally paralleling the capital-concentration point (§6.2) as a concentration of cognitive rather than financial supply; there is supporting evidence for reduced collective diversity of AI-assisted outputs, but its extension to beliefs is not verified here and should be held at low confidence.
Scope limits. These links are directional and mechanistic, not quantified: capture is established as the mechanism connecting unreliability to harm and as stratified along the divide, but no evidence supports a numerical share of the inequality attributable to it. And, as noted, affective capture is not a primary driver of assistant-case inequality — the operative axis is epistemic.
Closing synthesis — the consumer case as a whole
The two consumer use cases diverge from the enterprise case and from each other.
The personal-assistant (rational) case inherits the enterprise metering squeeze with weaker willingness-to-pay, and its reliability problem is sharpest exactly on the high-stakes drudgery people most want automated. Its social-equity profile is governed not by price — which is low, falling, and on the task level actively favors the less-skilled — but by literacy and adoption, by who controls their own time, and by capital concentration. Epistemic capture (§6.3) is the hinge in this case: it is what turns the tool's unreliability into actual loss, and it concentrates that loss on the users least able to verify the output or absorb the error.
The intimacy (affective) case escapes the enterprise cost trap — cheap distilled models, a low reliability bar, distillation driving text-companionship costs down — but hits a monetization wall: weak subscription revenue drives the category back toward ads and disclosure-harvesting, which is the market confirmation that the intimacy economy intensifies rather than replaces the attention economy, and which switches the manipulation incentive on.
The unifying thread across both consumer cases is the customer-versus-inventory distinction. The economically sustainable and ethically cleaner position is the one in which the user pays for utility and is the customer; the degraded position is the one in which the user's attention, disclosure, or dependency is the product being resold. The pricing model — metered or honest-subscription-for-utility versus ad-funded — is what determines which side of that line a given product lands on, and therefore whether its incentives point toward serving the user or toward capturing them. Relational AI's hard ceiling sits behind all of this: no pricing model and no instruction makes the system an accountable peer; that function still requires people who can be unimpressed and who remember when one was wrong.












