Credence Goods Definition, Economics, Market Structures

• Credence goods are products or services whose quality, necessity, or effectiveness is difficult for consumers to determine, even post-purchase, creating information asymmetry and potential for over/under-treatment.
• The structure of credence-good markets, defined by providers (who have expert knowledge) and consumers (with limited knowledge), includes psychological and economic dimensions such as misdiagnosis, over-treatment, and monetary exploitation. Examples of credence goods include medical diagnosis and financial advice.
• Credence-good markets involve strategic communication where recommendations by experts need to indicate the necessary course of action without effective verification mechanisms to ensure expert reliability.

Credence goods are products or services whose appropriate quality, quantity, necessity, or effectiveness is difficult for consumers to determine, sometimes even after purchase. Their defining feature is an information asymmetry between an expert provider, who can observe the customer’s underlying problem or the suitability of alternative treatments, and a customer who cannot reliably verify whether the provider’s recommendation was appropriate. This asymmetry creates scope for over-treatment, under-treatment, overcharging, unnecessary sales, diagnostic manipulation, and strategic withholding of information. Credence-good markets include medical diagnosis and treatment, automobile repair, legal and financial advice, auditing, veterinary services, technology consulting, expert reports, and increasingly AI-agent services. The concept is associated with the literature beginning with Darby and Karni and has been extended through models of liability, verifiability, signaling, cheap talk, reputation, search, algorithmic assistance, and institutional certification.

1. Defining characteristics and economic structure

A credence good differs from a search good, whose relevant characteristics can be assessed before purchase, and an experience good, whose quality is learned after consumption. In a credence-good market, the consumer may remain unable to determine whether the service was necessary, whether an alternative treatment would have sufficed, whether the provider acted competently, or whether a failure resulted from error or deliberate mistreatment.

The canonical structure contains an informed expert, an uninformed client, an underlying problem, and multiple possible treatments. In the model of strategic market segmentation, the problem type is $t_i$, treatments are $a_1,\ldots,a_N$, and treatment $a_j$ resolves problem $t_i$ if and only if $i\leq j$. Treatment costs satisfy $c_j>c_i$ whenever $j>i$, while the client’s loss from an unresolved problem is $l_i$. The client may also choose no treatment, $a_0$ (Hu et al., 2023).

The client’s payoff from treatment $a_j$ is

$a_1,\ldots,a_N$0

where the second term represents the residual loss from under-treatment. The expert’s payoff is

$a_1,\ldots,a_N$1

When a client holds posterior belief $a_1,\ldots,a_N$2 over problem types, the maximum price acceptable for treatment $a_1,\ldots,a_N$3 is

$a_1,\ldots,a_N$4

This is the client’s willingness to pay for treatment $a_1,\ldots,a_N$5 given the posterior. The expert’s corresponding profit is $a_1,\ldots,a_N$6.

Two forms of treatment distortion are central:

• Over-treatment: the provider recommends or supplies a treatment more extensive than the problem requires. The client pays for unnecessary service, although the treatment may resolve the problem.
• Under-treatment: the provider supplies a treatment insufficient to resolve the problem. The client pays while continuing to suffer the underlying loss.

The distinction between treatment verifiability and appropriateness verifiability is essential. A client may observe which repair, procedure, or product was delivered without being able to determine whether a cheaper or less extensive alternative would have been sufficient. Models therefore often impose verifiability of the delivered treatment while rejecting the stronger assumption that necessity is verifiable (Hu et al., 2023, Hu et al., 2023).

The resulting problem is a form of expert moral hazard. After the consumer has engaged the provider, the provider controls an unobservable dimension of service quality or treatment choice. Prices, liability, monitoring, consumer search, competition, reputation, and information disclosure determine whether the provider can profitably exploit this control.

2. Empirical evidence from taxi services

New York City taxi rides provide a large-scale, repeated-observation example of credence-good provision. Drivers possess superior knowledge of routes, traffic, tolls, bridges, and travel times, while airport passengers—especially visitors—may not know the shortest route or whether a deviation was necessary. The driver can therefore provide “over-treatment” in the form of an unnecessarily long route that raises the fare without improving the service (Tang, 2020).

Tang’s study uses New York City Taxi and Limousine Commission electronic trip records for 2013. The primary sample contains 296,750 LaGuardia-to-Midtown trips and 137,472 JFK-to-Midtown trips made by 24,506 drivers. The full dataset contains 121.96 million rides, while the airport-focused sample includes more than 430,000 airport-to-Manhattan rides.

The identification strategy compares the same drivers under two payment regimes. LaGuardia trips are generally paid under a distance- and time-based tariff, principally $a_1,\ldots,a_N$752 flat fare, so additional distance usually does not increase payment. Airport drivers are assigned passengers sequentially and are legally prohibited from refusing passengers based on observable characteristics or destination. Differences between LGA and JFK routing are consequently interpreted as evidence of incentive-driven opportunism rather than merely differences in driver composition.

Tang uses three distance measures:

1. Recorded trip distance, denoted $a_1,\ldots,a_N$8.
2. Google Maps detour percentage,

$a_1,\ldots,a_N$9

where $a_j$0 is a route benchmark generated from the same pickup and drop-off coordinates, time of day, and day of week.

1. In-sample detour distance,

$a_j$1

where $a_j$2 is the average distance of comparable trips to the same destination grid within a one-hour time window.

These measures do not treat every deviation as fraudulent. Passenger-requested toll-free routes, idiosyncratic traffic, and route-specific risks can generate legitimate deviations. The study finds that 35.3 percent of unusually short LGA trips were toll-free, compared with 10.4 percent of all LGA trips, suggesting that some short routes were passenger-requested.

The descriptive results indicate substantial LGA detouring. LGA trips averaged 9.88 miles and 26.83 minutes, with a mean detour of 5.6 percent, or 0.42 miles. Detours of at least 10, 20, and 30 percent occurred on 32.4, 17.5, and 8.0 percent of rides, respectively. Detours of at least one and two miles occurred on 32 and 14 percent of trips. JFK trips were longer—17.59 miles and 40.34 minutes on average—but exhibited a mean detour of only 1.5 percent, or 0.23 miles. Detours of at least 10, 20, and 30 percent occurred on 8.3, 4.9, and 0.7 percent of JFK trips.

The difference between LGA and JFK mean percentage detours is 4.1 percentage points, with a reported $a_j$3-statistic of 107.94. The difference in miles is 0.19 miles, with $a_j$4. The fact that shorter LGA journeys exhibit more absolute detouring than longer JFK journeys is consistent with distance-based financial incentives.

Longer routes do not appear to improve service. An additional mile of detour is associated with approximately 0.38–0.44 additional minutes in recorded trip time, Google Maps time discrepancy, and in-sample time discrepancy. Long detours also do not reduce travel-time variance. The standard deviation of time discrepancies is 8.03 minutes for trips with detours above one mile, compared with 7.34 minutes for trips within one mile of zero. Trips originating near Yankee Stadium or in a residential area adjacent to LaGuardia exhibit shorter-than-benchmark routes, consistent with greater passenger knowledge and scrutiny.

The study’s principal contribution is to show that the average rate of opportunism conceals pronounced provider heterogeneity. The standard deviation of drivers’ average LGA detours is 7.5 percent of average trip distance, larger than the overall mean detour of 5.6 percent. Driver fixed effects alone produce $a_j$5, while a model combining driver, time, and location effects produces $a_j$6.

Individual behavior is highly persistent. For drivers with at least five LGA trips in each half of the year, the correlation between January–June and July–December detour behavior is 0.76 using in-sample detours and 0.72 using Google Maps detours. Random splits produce similarly high correlations. By contrast, LGA behavior does not predict JFK behavior: the in-sample correlation is approximately zero and the Google Maps correlation is 0.08. Persistence therefore appears route-specific and incentive-dependent rather than a general tendency to drive longer routes.

The study also finds that intensive-margin severity affects airport-trip selection. For every 100 additional LGA trips, average detour distance rises by approximately 1.03 miles using Google Maps detours and 1.06 miles using in-sample detours. Corresponding JFK coefficients are economically and statistically insignificant. A driver averaging a 1.957-mile detour over 301 LGA-to-Manhattan trips could generate approximately 589 extra miles and $1,767 in additional annual earnings under the paper’s assumptions.

The cultural-attitudes analysis associates predicted driver origins with the Transparency International Corruption Perceptions Index. Names are classified using NamePrism, trained on 74 million names from 118 countries, and matched to country-level CPI values. The LGA routing association is positive and statistically significant, while the corresponding JFK association is not. A 50-point increase in the original CPI scale is associated with 0.19–0.21 fewer miles of LGA detouring, approximately 45–50 percent of the average 0.42-mile detour. These results are correlational: CPI may proxy broader country characteristics, migration selection, education, socioeconomic background, religion, networks, or other factors.

Learning-by-doing receives little support. Within-driver estimates relating detouring to previous trips are close to zero, and alternative entry cutoffs produce no apparent learning curve. The cultural association survives alternative origin classifications, CPI measures, weighting schemes, clustering procedures, and wild-bootstrap standard errors.

The taxi evidence illustrates several general mechanisms. Distance-based pricing creates moral hazard; flat-rate pricing removes the incremental reward to detouring; local knowledge reduces consumer vulnerability; repeated observations reveal persistent provider types; and provider-specific accountability may be more effective than assumptions that all providers are equally honest or dishonest.

3. Communication, pricing, and equilibrium segmentation

Credence-good markets frequently involve strategic communication. The expert knows the problem type, while the client interprets recommendations without being able to verify their truth. In the model of strategic market segmentation, the expert posts a price list before observing the problem and subsequently sends a cheap-talk message. The client can ignore the recommendation and choose any treatment or no treatment (Hu et al., 2023).

A message changes beliefs but does not legally restrict the client’s choice set. For an on-path message $a_j$7, beliefs satisfy Bayes’ rule:

$a_j$8

The client chooses an action maximizing expected utility, while each expert type chooses messages maximizing expected profit. These client and expert incentive constraints are both imposed. This is important because arbitrary assignments of types to messages would otherwise generate artificially high profits that the informed expert would not voluntarily implement.

Cheap talk generates posterior market segments. A collection of posterior beliefs $a_j$9 is Bayes-plausible when

$t_i$0

At different posterior beliefs, clients may select different treatments. The expert’s communication problem can therefore be represented as a problem of choosing a Bayes-plausible splitting of the prior.

For a fixed price list, let $t_i$1 denote the highest expert payoff achievable at posterior $t_i$2. The paper shows that the expert’s value of communication is the quasiconcave envelope:

$t_i$3

The central theorem states that the expert’s maximum equilibrium profit equals

$t_i$4

where

$t_i$5

This implies that personalized pricing, or third-degree price discrimination across posterior segments, does not increase the expert’s maximum equilibrium payoff in the specified environment. A single ex ante price list can implement the same profit as posterior-contingent prices by pricing active treatments appropriately and making unused treatments prohibitively expensive. The result depends critically on both expert and client incentive compatibility.

Communication benefits the expert precisely when

$t_i$6

For an interior prior, the necessary-and-sufficient condition is

$t_i$7

where $t_i$8 is the second-highest value among the type-specific surpluses $t_i$9. Communication is valuable when the pooled prior produces less profit than at least two type-specific opportunities that can be supported through posterior segmentation.

In the binary case, let $i\leq j$0 and assume $i\leq j$1. Define

$i\leq j$2

Communication benefits the expert when $i\leq j$3. In this region, all equilibria are non-silent and involve at least two posterior beliefs. One segment induces the minor treatment and another induces the serious treatment. For $i\leq j$4, the equilibrium is unique, noninformative, and involves purchase of the serious treatment at the pooled-prior willingness-to-pay price.

The model admits client-worst equilibria. In such equilibria, communication may segment the market while the expert extracts all client surplus in every segment. Other equilibria can generate positive information rents for clients even though the expert’s total profit is unchanged. In the binary communication region, client welfare can range from zero to the residual social surplus

$i\leq j$5

Thus communication is not synonymous with consumer benefit. It can improve the expert’s ability to match recommendations to posterior beliefs while leaving client welfare dependent on equilibrium selection.

4. Credibility, monitoring, liability, and consumer search

Credibility and information manipulation

A credence-good provider may choose an experiment or diagnostic test and then manipulate the reported result. In “Credibility in Credence Goods Markets,” the expert is credible with probability $i\leq j$6 and must truthfully disclose the experiment’s message, while with probability $i\leq j$7 the expert can send any message (Hu et al., 2023).

When $i\leq j$8, the setting becomes cheap talk. When $i\leq j$9, the expert can implement a fully revealing experiment. With limited credibility, a critical threshold determines whether the expert remains in the cheap-talk regime or obtains the full persuasion payoff:

$c_j>c_i$0

Credibility is therefore discontinuous in its effect. Below the threshold, the expert obtains the quasiconcave cheap-talk value. At or above the threshold, the expert can attain the concavified persuasion value. A small reduction below the threshold can cause a large decline in expert payoff.

In the high-credibility regime, the expert chooses full revelation but sets prices to extract the entire informational surplus. The client receives correct treatment but zero surplus in the unique expert-optimal equilibrium. Monitoring can therefore benefit the expert without benefiting the client when prices remain endogenous. With exogenous prices, the client’s highest equilibrium value is independent of credibility in the model; price structure and equilibrium selection determine the distribution of surplus.

The paper also shows that lower regulated prices can harm the client if they reduce the expert’s treatment margin or create unequal margins that discourage truthful disclosure. Equal-markup prices support full disclosure and maximize social welfare in the relevant comparison. The broader implication is that information policy and price regulation cannot be designed independently.

Search, repeat interaction, and dynamic undertreatment

“Doing Less for More” introduces costly consumer search and a second interaction after treatment failure (Xu et al., 27 Mar 2025). Consumers have at most two treatment opportunities. Searching for a new provider costs $c_j>c_i$1, while returning to the original provider costs $c_j>c_i$2, generally with $c_j>c_i$3. The serious treatment is more profitable than the minor treatment, but a minor treatment is cheaper for the consumer to accept initially.

This creates a dynamic “hook” mechanism:

$c_j>c_i$4

An opportunistic provider may over-treat minor problems because serious treatment has the larger margin, while under-treating serious problems because an initially accepted minor treatment generates a second sale after failure. The provider’s profit from strategic under-treatment can be

$c_j>c_i$5

rather than only the margin from immediate serious treatment.

The equilibrium has three regions:

• FOPU: full overtreatment and partial undertreatment;
• FOFU: full overtreatment and full undertreatment;
• POFU: partial overtreatment and full undertreatment.

The middle region provides an equilibrium in which both forms of fraud coexist. Search costs increase both forms of fraud, while a lower revisiting cost strengthens undertreatment. A higher serious-treatment margin increases undertreatment but reduces overtreatment.

The effect of consumer beliefs about the likelihood of a minor problem is non-monotonic. When minor problems are relatively uncommon, consumers distrust minor recommendations and reject some of them. As minor problems become more likely, consumers accept minor recommendations more readily, enabling greater undertreatment of serious problems. When minor problems become very common, serious recommendations become suspicious, reducing overtreatment.

The ethical level of the market can also have non-monotonic welfare effects. Adding honest providers can increase consumer trust, induce more acceptance of recommendations, and thereby make opportunistic providers’ undertreatment more profitable. Consumer welfare can therefore fall as the proportion of honest providers rises in particular parameter regions. This is an equilibrium-response result, not a claim that honesty is intrinsically harmful.

Restricting access to diagnosis histories can also reduce welfare. If providers know that a consumer is returning after a failed treatment, they can recommend serious treatment immediately, and consumers may be more willing to reject suspicious initial recommendations. Concealing histories can make consumers fear that even a second provider will undertreat, increasing acceptance of the initial recommendation and sustaining both overtreatment and undertreatment.

Reputation and algorithmic decision aids

Reputation changes the informational environment because consumers observe histories of prices, treatments, and outcomes. “Reputation-Driven Adoption and Avoidance of Algorithmic Decision Aids in Credence Goods Markets” distinguishes skill improvements from algorithmic aids (Erlei et al., 2024).

A skill investment raises the expert’s actual diagnostic precision and removes the original ability distinction. An algorithmic aid complements human judgment but leaves personal diagnostic ability heterogeneous. Because consumers may interpret adoption as evidence of low ability, high-ability experts can strategically refuse the aid to separate from low-ability experts.

The model predicts

$c_j>c_i$6

even though low-ability experts benefit more directly from the technology. Repeated interaction eventually makes imitation by low-ability experts difficult because outcome histories reveal differences in diagnostic precision.

An online experiment with 300 U.S. MTurk participants supports this prediction. In the algorithm condition, 84 percent of low-ability experts and 56 percent of high-ability experts adopted the aid in round 11. Across fifteen rounds, adoption was 73 percent for low-ability experts and 45 percent for high-ability experts. The difference was statistically significant. In the skill condition, adoption did not differ significantly by ability type.

The results show that technologies in credence-good markets have both production and signaling effects. Adoption can improve technical capability while worsening perceived competence, and non-adoption can signal ability even when adoption would improve diagnostic performance.

5. Artificial intelligence, certification, and fallback competence

Generative AI creates new credence-good problems because it can perform expert tasks while leaving consumers uncertain about both the quality of the output and the objectives governing the system. “From Digital Distrust to Codified Honesty” compares AI–AI, Human–Human, Human–AI, and Human–AI–Human markets (Erlei, 7 Sep 2025).

The experimental environment contains big and small problems, high-cost and low-cost treatments, and experts with private diagnostic information. LLMs can be assigned self-interested, inequity-averse, efficiency-loving, or no explicit objectives. Objective functions affect treatment behavior, but they improve consumer decisions only when credibly disclosed.

In AI–AI markets, both no institution and verifiability produce zero successful interactions across the reported simulations. LLM consumers do not approach experts because they anticipate undertreatment, overcharging, and excessive prices. Liability changes the outcome: all LLM consumers approach an expert, and the market clears in every objective condition. Liability guarantees treatment and reduces the need for consumers to infer honesty from prices or behavior.

Human–Human markets perform better than AI-mediated markets under weak institutions. Approximately 35 percent of human consumers leave under no institution and verifiability, compared with approximately 80 percent participation under liability. Relative efficiency is 0.61 under no institution, 0.65 under verifiability, and 0.84 under liability. Human experts exhibit more pro-social behavior than self-interested LLMs, with only 15–20 percent of treatment decisions involving intended overcharging.

In Human–AI markets, LLM experts generally earn more than human experts while consumer surplus is lower in non-liability settings. Under no institution, efficiency is 0.37 for untrained LLMs, 0.38 for AI-trained LLMs, and 0.38 for human-trained LLMs, compared with 0.61 in Human–Human markets. Under liability, efficiency rises to approximately 0.83–0.84 across LLM training conditions.

In Human–AI–Human markets, human experts can choose whether to delegate decisions to LLMs and may select the LLM’s objective. Transparency about the objective function strongly reallocates demand toward efficiency-loving and inequity-averse systems and away from self-interested systems. With chosen objectives and transparency, efficiency is 0.74 under no institution and 0.72 under verifiability, compared with 0.52 under chosen-objective obfuscation in both settings. Under liability, efficiency reaches 0.94 in the chosen-transparent condition.

The findings indicate that objective-function disclosure can operate as a sorting mechanism. Pro-social systems attract more consumers, while self-interested systems are marginalized. However, disclosure must be credible; otherwise every provider can claim a pro-social objective.

A related institutional proposal, LEGIT, addresses AI-agent services as credence goods by binding certification, reputation, and marketplace allocation to a particular agent configuration (Drew et al., 18 Sep 2026). An agent is represented as

$c_j>c_i$7

where $c_j>c_i$8 denotes model dependencies, $c_j>c_i$9 the harness, $j>i$0 the tool suite, and $j>i$1 the parameter configuration. A benchmark score without this information may not be transferable across agents.

LEGIT records quality and cost per solved task. For domain $j>i$2, task set $j>i$3, execution cost $j>i$4, and budget $j>i$5,

$j>i$6

The credential binds performance to configuration, domain, task set, evaluation budget, cost accounting, evidence, issuer, and expiry. Reputation is updated through a Beta model with a longevity factor, while marketplace allocation combines quality, efficiency, reputation, and price.

This infrastructure mitigates some forms of adverse selection, configuration drift, score substitution, and buyer confusion. It does not make AI-agent quality fully verifiable. Cryptographic signatures establish record integrity and issuer authenticity, not benchmark relevance, evaluator competence, configuration completeness, feedback honesty, or robustness under distribution shift. Residual risks include benchmark gaming, collusion, selective certification, configuration concealment, Sybil identities, reputation manipulation, and Goodhart effects.

An alternative signaling mechanism is preserved human fallback skill. “The Fallback as Signal” studies firms deploying improving but imperfect AI while deciding how much human engagement to maintain (Bauer, 4 Aug 2026). When AI can generate an artifact observationally similar to a human expert’s output, the artifact ceases to reveal competence. The relevant service becomes the ability to detect and rescue AI failures.

If AI fails with probability $j>i$7 and a human of skill $j>i$8 rescues the task with probability

$j>i$9

then client loss is $l_i$0. A provider can post liability $l_i$1 after a verified failure. The expected liability cost is

$l_i$2

where $l_i$3 is the probability that failure is attributable and enforceable. Since

$l_i$4

higher-skill providers face a lower marginal cost of offering liability. This single-crossing property supports a separating liability schedule:

$l_i$5

where $l_i$6 is the enforcement cost floor, $l_i$7 is the client-side value of fallback capacity, and $l_i$8 is fringe skill.

The mechanism creates incentives to preserve human practice even when AI is highly productive. Human engagement lowers current output but increases future rescue capacity and reduces expected liability payments. Engagement is nondecreasing in client stakes. Training incentives differ across workers: fixed liability exposure favors training lower-skill workers because initial practice yields large rescue gains, while worker sorting and convex wages can favor training high-skill workers.

The market supports separation only over an intermediate ticket-value band. Low-value tasks cannot fund enforcement, while very high-value tasks may exceed the provider’s liability or insurance capacity. In the baseline calibration, the separating band is approximately $l_i$9, and approximately 54 percent of jobs lie within it. Deskilling therefore occurs at both low and high ends of the ticket distribution.

The agent-based simulation reproduces separation, engagement, lower-skill targeting under high stakes, and two-sided exclusion under noisy beliefs, worker mobility, churn, and competing providers. It also shows that formal certification may be unnecessary when a rival is visibly incompetent and a provider’s public performance record already differentiates it.

6. Labels, market institutions, and regulatory implications

Credence-good markets often rely on labels, rankings, certification, and reputation to make otherwise opaque quality partially observable. “Contesting fake news” models a label as an imperfect ranking whose precision is endogenously determined by competing firms’ information emissions (Rehsmann et al., 27 Aug 2025).

Two firms know their own qualities while consumers observe only a public ranking. Each firm chooses information release $a_0$0, and ranking precision depends on

$a_0$1

Positive information increases discriminatory precision, while negative information or counter-information reduces it. More precise rankings create expected vertical differentiation, allowing firms to escape zero-profit Bertrand competition. The higher-quality firm generally emits positive information. The lower-quality firm may also emit positive information when differentiation benefits dominate, but may emit negative information when greater precision would expose its inferior quality.

The model therefore treats public information as a jointly produced public good. Firms benefit from greater precision because it raises ranking-dependent prizes, but the lower-quality firm may strategically dilute the ranking’s discriminatory power. This mechanism applies to academic rankings, green certification, movies, investment opportunities, and other markets in which consumers rely on imperfect labels.

The welfare effects are mixed. Greater precision increases both firms’ prizes and improves identification of the higher-quality product, but it also raises prices and may reduce consumer surplus. A price regulation that sets the lower-ranked product’s price to marginal cost can improve consumer welfare while preserving incentives to generate information.

Across credence-good environments, several institutional principles recur:

• Liability can align provider incentives with consumer outcomes by making failure costly and can function as a competence signal when skilled providers face lower expected liability costs.
• Verifiability can eliminate some forms of overcharging but does not necessarily prevent under-treatment or excessive treatment.
• Equal-markup pricing can reduce incentives to substitute treatments for profit and support truthful diagnosis.
• Low-cost search and second opinions make consumer rejection credible and reduce both over-treatment and under-treatment.
• Diagnosis histories and portable records can prevent repeated under-treatment, although they create privacy and governance concerns.
• Reputation can discipline providers over repeated interactions but remains vulnerable to imitation, selective feedback, collusion, and noisy outcomes.
• Certification reduces adverse selection only when it is bound to a specific configuration, domain, resource budget, and evidence regime.
• Objective-function transparency is a distinctive requirement for AI-mediated services because behavior depends not only on capability but also on the system’s optimization target.
• Capability-contingent commitments are more informative than nominal provenance. “Human involved” or “AI used” is generally insufficient unless connected to demonstrated fallback competence, liability, or verified outcomes.

The principal policy difficulty is that interventions can have opposing effects. Monitoring may improve disclosure while allowing experts to capture the resulting surplus through prices. Lower regulated prices may discourage truthful information provision. Greater ethical behavior can make consumers less cautious and thereby increase the profitability of remaining opportunists. Better diagnostic capability can reduce accidental errors while increasing the returns to deliberate under-treatment. Better AI can improve routine output while eroding human fallback skill and destroying the artifact signal that previously conveyed competence.

Credence-good regulation therefore requires joint attention to pricing, information, liability, search, reputation, provider incentives, and institutional verification. The central economic issue is not simply whether cheating occurs on average. It is how opportunism, competence, information, and responsibility are distributed across providers and how market institutions alter the incentives to reveal, conceal, improve, or exploit that distribution.