Verification loop that rewards deception
Singapore’s government‑run dating service asks users to submit detailed profiles and then computes compatibility scores from those self‑reported traits. The service’s designers wonder whether they can detect deceptive answering, feed back unsuccessful date outcomes, and adjust scores with a “Je n’sais quoi” factor that rewards or penalises users based on past behavior.
The mechanism at work is simple: the service relies on information that users control, users benefit from presenting themselves more favorably than they are, and the service tries to close the gap by checking the truth of those claims. When the service introduces a check — whether it flags inconsistent answers, weights down users whose dates repeatedly fail, or adds a hidden factor derived from feedback — users notice that the check creates a new dimension they can manipulate. They learn which aspects of their profile trigger the check and adjust their self‑presentation accordingly, often by refining the deception rather than eliminating it. The service’s attempt to improve accuracy therefore creates a new incentive to game the very metric it is trying to protect, and the signal that the score is based on becomes progressively less informative.
In the dating‑service thread, users ask whether the platform can tell when someone is trying to widen their net versus undergoing genuine growth, whether pairing detected deceivers with other deceivers is sensible, and whether a feedback‑derived factor can be used to bias future matches. These questions reveal that the service is contemplating a verification step that would look at historical outcomes and use them to reshape the compatibility calculation. If such a step were introduced, users would quickly realize that a pattern of unsuccessful dates could be avoided by deliberately selecting partners who are also likely to fail, or by altering profile answers in ways that do not affect the feedback loop but still improve the perceived score. The service would then face a choice: accept that the score is now reflecting users’ ability to game the feedback loop rather than their underlying compatibility, or add another layer of detection, which would again be met with new counter‑measures.
The same loop appears wherever a system distributes rewards based on self‑reported data and then tries to verify those reports. In consumer credit, lenders rely on applicants’ stated income and existing debt to calculate a risk score. Applicants who inflate income receive better loan terms, so they have a direct incentive to overstate earnings. Lenders respond by requesting pay‑stubs, contacting employers, or using third‑party income verification services. Applicants counter by producing forged pay‑stubs, providing fake employer references, or using shell companies that confirm the inflated figure. Each verification upgrade spawns a more sophisticated falsification technique, and the income signal that drives the score loses its predictive power.
Insurance underwriting works similarly. Insurers ask prospective policyholders about health status, occupation, and lifestyle to set premiums. Individuals who conceal a pre‑existing condition or a risky hobby obtain cheaper coverage. Insurers counter with medical examinations, prescription‑history checks, and employer questionnaires. Policyholders answer by visiting compliant doctors, timing examinations to hide transient symptoms, or purchasing short‑term policies that avoid scrutiny. The verification effort pushes the concealed information deeper into the supply chain of deception, making the original health signal increasingly noisy.
Online job markets exhibit the pattern as well. Candidates submit résumés that list education, experience, and skills. Recruiters use keyword‑matching software and automated scoring to rank applicants. Candidates who add buzzwords, exaggerate tenure, or claim unearned certifications rise in the rankings. Recruiters respond by running background checks, requesting work samples, or administering skill tests. Candidates counter by purchasing fabricated diplomas, hiring coaches to rehearse interview answers, or using deep‑fake video tools to simulate competence. The verification loop drives the signal toward a measure of how well a candidate can simulate the desired traits rather than how well they actually possess them.
Historical instances predate digital platforms. In the late‑nineteenth‑century United States, patent‑medicine manufacturers advertised curative properties for tonics, elixirs, and salves that contained little more than alcohol and sugar. Consumers relied on these claims to choose remedies. The Pure Food and Drug Act of 1906 forced manufacturers to list active ingredients, intending to make the claim verifiable. Manufacturers responded by renaming ingredients, using proprietary blends that evaded disclosure, or shifting to vague “natural” descriptors that sidestepped the regulation. The verification rule (ingredient labeling) prompted a new form of misrepresentation, and the efficacy signal that consumers used to judge products deteriorated.
A century earlier, the Chinese imperial examination system tested candidates on their ability to compose essays on Confucian classics. Success granted elite bureaucratic appointments, so candidates memorized stock passages and formulaic arguments. Examiners, noticing the uniformity, introduced surprise topics and prohibited the use of prepared texts. Candidates responded by developing mnemonic techniques to reconstruct essays on the fly, by employing hidden cheat sheets, or by bribing examiners to overlook infractions. Each verification tweak produced a more elaborate cheating strategy, and the examination score increasingly reflected skill at evasion rather than mastery of the classics.
Even earlier, medieval craft guilds stamped goods with a quality mark to assure buyers that a piece met the guild’s standards. Buyers trusted the mark and paid a premium for stamped items. Unscrupulous workshops began to forge the stamp or to apply it to sub‑standard work. Guilds hired inspectors to check workmanship and to punish fraudsters. Counterfeiters responded by producing near‑perfect imitations of the stamp, by marking goods in inconspicuous places, or by moving production outside guild jurisdiction. The verification act (the stamp) became a target for forgery, and the original guarantee of quality lost its force.
Across these cases the causal chain is identical: a distributor of advantage (a lender, insurer, employer, state, or guild) offers a benefit based on a signal that the recipient can influence; the recipient improves their outcome by distorting the signal; the distributor introduces a check to recover the true state of the signal; the recipient learns how to satisfy the check while preserving the advantage; the signal’s correlation with the underlying trait weakens. The loop does not require malice on either side; it follows from the asymmetry of payoff and the feasibility of observation.
What remains unresolved is whether any verification mechanism can ever restore a reliable signal in such environments, or whether the system settles into a perpetual arms race where the cost of verification rises without a corresponding gain in predictive accuracy. The Singapore dating service’s contemplation of a feedback‑derived “Je n’sais quoi” factor illustrates the next step in that arms race: once the platform begins to weight users by past date outcomes, users will learn which behaviors generate unfavorable feedback and adjust their profiles to avoid those triggers, not to become better partners. The service will then face the choice of accepting a score that reflects skill at avoiding negative feedback or adding yet another layer of scrutiny, which will again be met with new counter‑measures.
The essential point is that any attempt to improve a self‑reported metric by observing its consequences creates a new dimension for strategic manipulation, and the metric’s validity erodes as a direct result of the improvement effort. Unless the platform can eliminate the users’ ability to influence the observed outcome — something impossible in a setting where users control their own presentation — the verification loop will inevitably degrade the signal it seeks to protect.