The Securitization of Remote Work: Referral-Driven AI Interview Funnels as Bond-like Claims on Uncertain Work
Abstract
AI-run hiring platforms that supply experts to frontier AI labs, such as Mercor and micro1, now recruit through two linked devices: an automated AI interview that applicants generally must pass, and a referral programme that pays existing members for bringing in new applicants. Applicants who fail are often encouraged to apply for other listings. This paper argues that the resulting funnel is economically closer to a securitization chain than to a conventional vacancy. A loosely specified pool of client projects is the underlying asset; referrers act as originators paid on volume; the platform’s “opportunity” is the pooled vehicle; the AI interview and follow-on assessments tranche applicants by their probability of ever being paid; and applicants “buy” these claims with their time and personal data. We document the design of the referral contracts from the platforms’ own published terms, and assemble evidence from a research paper co-authored by micro1 staff, wage litigation against Scale AI’s Outlier, and regulator data on job-related fraud. A simple model shows that when a platform earns a positive data or option value on every applicant and pays referrers only on contingent events, it is privately optimal to issue far more claims than there are paid seats. Free entry then dissipates the entire wage rent of the jobs in screening costs, and redirecting failed applicants to new listings resembles re-securitizing a residual tranche whose defaults are highly correlated. We close with disclosure, risk-retention, and score- and data-portability proposals that borrow directly from post-2008 securitization reform.
Keywords: securitization; online labour markets; AI interviews; referral bonuses; rent dissipation; data labour; platform economics.
1. Introduction
A new kind of job solicitation now reaches skilled workers. An AI-run platform says it is hiring, but the offer is conditional on passing an automated interview, and a candidate who fails is typically encouraged to apply for other listings. This paper argues that such funnels behave less like a labour market clearing a vacancy and more like a financial intermediary issuing claims on a pool of uncertain future work. We call this the securitization of remote work.
The motivating artefact is an ordinary follow-up message received by the author from a recruiter at Alignerr, the expert-contractor platform operated by Labelbox:
“Just following up to see if any questions have come up regarding the software engineering projects or the onboarding process. We still have active opportunities for engineers with experience in Go, Solidity, and agentic AI systems, and completing your profile is the best way to get matched with projects that fit your background.”
The message does not identify a client, project, rate, hours or headcount; it offers only the chance to be considered for a pool of projects, conditional on completing a profile and passing a screen. The skills requested (a systems language, a smart-contract language and agentic AI) do not describe one job; they describe a bundle broad enough to cover many possible future projects. The message is also a follow-up: its purpose is to keep an applicant who has already entered the funnel from leaving it.
The analogy rests on three features of the funnel.
1. Pooling: many heterogeneous, short-lived client projects are bundled into a single advertised “opportunity,” just as many mortgages are pooled into one security.
2. Wrapping and tranching: the underlying work is wrapped in layers (referral link, profile, AI interview, skills assessment, project-level qualification), and each layer sorts applicants into groups with different probabilities of ever being paid.
3. Over-issuance: because marketing a listing is close to free and referrers are paid per sign-up or per hire, the number of claims issued can far exceed the paid hours that ever materialise.
Before 2008, mortgage originators were paid on volume and passed loans on to securitizers; Keys et al. (2010) show that loans that were easier to securitize were screened less carefully and defaulted more often. In the job funnels studied here, referrers originate applicants, platforms pool them, an AI system rates them, and the applicants themselves, not outside investors, hold the risk that the claim never pays.
Contributions. First, we document, from the platforms’ own published terms, how referral contracts at Mercor and micro1 are designed, and when they pay. Second, we give an explicit mapping from these funnels to the structure of asset-backed securities. Third, we write a simple model that shows when a platform rationally over-issues claims, how much of the jobs’ value is dissipated, and why redirecting failed applicants is weaker diversification than it looks. Fourth, we draw welfare and policy implications and set out an empirical agenda.
Throughout, “bond” is an analytical lens. We do not claim that any named firm acts unlawfully or deceptively. The platforms studied pay real money for real work, and many workers report good experiences. We argue only that a common business model creates incentives whose costs fall on applicants, and that tools from the analysis of securitized credit help measure them.
2. Institutional background
2.1 Expert-data platforms and AI interviewers
Training and evaluating large language models requires large volumes of expert human judgement: code reviewed by engineers, proofs checked by mathematicians, answers graded by doctors and lawyers. A set of platforms has grown up to supply this labour on a contract basis. Mercor raised US$350 million at a US$10 billion valuation in October 2025 and said it was paying more than US$1.5 million per day to more than 30,000 experts who “earn over $85 per hour on average” (TechCrunch, 2025b). micro1, Alignerr (Labelbox) and Outlier (Scale AI) operate similar models. Most of the work is remote, project-based and paid by the hour or by the task.
These platforms screen at scale with AI interviewers. micro1 describes its interviewer, Zara, as an agent that sources, interviews and scores candidates (micro1, 2026d); its referral page states that “the AI interview takes 20–40 minutes, depending on the skills required for a role,” with some roles requiring a further skills assessment (micro1, 2026a). Mercor’s documentation describes a matching system that weighs “your performance in previous interviews and assessments at Mercor,” and marks some candidates as “prequalified,” who “are treated as if they’ve already applied” (Mercor, 2026c). A third-party guide describes Alignerr’s onboarding as including “skill assessments and qualification tasks” (AIGigJobs, 2026).
2.2 Referral programmes
Employee referral is an old and generally efficient hiring channel: referred workers tend to be better matched and less likely to quit (Montgomery, 1991; Burks et al., 2015). The platforms studied here, however, open referral to anyone with an account, attach a cash bounty to each listing, and allow a single member to hold thousands of active referrals. The referrer thus becomes an independent marketing agent whose income depends on the volume and quality of the people they bring into the funnel. Section 4 documents the terms.
2.3 Related literature
This paper connects four literatures. The first is the economics of securitization and structured finance: originate-to-distribute incentives and lax screening (Keys et al., 2010; Ashcraft & Schuermann, 2008), the sensitivity of tranche values to correlation (Coval, Jurek & Stafford, 2009), and conflicts in ratings when the rater is paid by the issuer (Bolton, Freixas & Shapiro, 2012). The second is the economics of online labour markets and platform intermediation (Autor, 2001; Horton, 2010; Pallais, 2014; Kässi & Lehdonvirta, 2018; Rochet & Tirole, 2003) and the sociology of invisible platform work (Gray & Suri, 2019). The screening problem itself is a signalling and adverse-selection problem in the sense of Spence (1973) and Akerlof (1970). The third is the theory of rent seeking and lottery-like entry: when a fixed prize is allocated by a contest with free entry, entrants’ spending can dissipate the prize (Tullock, 1980), and Harris and Todaro’s (1970) model of migration toward scarce high-wage jobs rests on the same expected-wage entry condition. The fourth treats personal data as a form of labour (Arrieta-Ibarra et al., 2018): an AI interview is both a screen and a recorded work sample.
3. Conceptual framework: the job as a securitized claim
A conventional vacancy is a single, identified position with an employer, a wage and an expected start date. The applicant’s uncertainty is about whether they will be chosen. In the platform funnel, the applicant faces a second, larger uncertainty: whether paid work exists at all, how much of it there is, and for how long. The platform’s “opportunity” is a claim on that uncertain flow. Figure 1 sets the two chains side by side.
Figure 1. Mortgage securitization and the AI-interview job funnel share one structure. The returning arrow marks the redirection of failed applicants to new listings.
Table 1 makes the mapping explicit. The mapping is meant to borrow concepts and remedies developed for securitized credit, not to equate the two markets. Strictly, a referrer is closer to a broker who places claims with retail buyers than to a loan originator; the incentive problem, payment on volume without retained risk, is the same.
Table 1. A dictionary from securitized credit to the AI-interview job funnel
Because applicants pay in time and data rather than money, the arrangement is rarely seen as a sale. Yet a 20–40 minute AI interview, a profile, and a multi-hour assessment have a clear opportunity cost for a skilled engineer, and the recorded interview has value to the platform as a work sample and as an input to future matching. The rating agency and the issuer are also the same entity. The platform designs the interview, scores the applicant and decides whether to re-offer the applicant to another listing. In credit markets even the weaker version of this conflict, ratings paid for by the issuer, inflated ratings (Bolton, Freixas & Shapiro, 2012); here the rater and issuer are one firm.
4. Case evidence
We assemble evidence from the platforms’ own published terms, a research paper co-authored by micro1 staff, court filings reported in the press, and regulator data. Where we rely on secondary or unverified sources we say so. All web sources were accessed on 13 October 2026; terms change often, and the figures below describe the terms as published on that date.
4.1 Case A: the Alignerr follow-up message
The message quoted in Section 1 has the formal features of a pooled claim. It refers to “active opportunities” in the plural, without a count. It requests three unrelated specialisms, which is consistent with a pool of projects rather than one role. Its call to action, “complete your profile,” moves the applicant one layer deeper into the funnel and creates the data asset the platform needs to match the applicant later. A third-party guide reports that “some workers report waiting weeks after onboarding before receiving their first project” (AIGigJobs, 2026). Alignerr does not publish its referral terms on its public site, so we do not include it in the comparison of referral contracts in Section 4.5.
4.2 Case B: Mercor’s earnings-share referral contract
Mercor’s referral contract is the most outcome-linked of the three. According to its published policy, “referrers earn 20% of their referral’s earnings, including hourly pay and bonuses, up to an Earnings cap,” and the cap for hourly contracts is four times the referral’s hourly rate (Mercor, 2026a). Earnings-sharing begins only “when your referred candidate is hired and begins billable work.” Work trials do not count. Because the referrer is paid only out of the recruit’s earnings, the contract works much like risk retention. Mercor’s marketing page summarises the result as “$250–$15,000” per referral depending on role and seniority (Mercor, 2026b).
The same policy reveals the intended volume. A referral link “can be used up to 100 times per rolling 30 days,” and a member may hold up to 5,000 active referrals (Mercor, 2026a). These limits are high for an individual referrer and would accommodate members who recruit at high volume. Mercor’s “Instant Offers” mechanism adds the pooling feature: candidates can be marked prequalified on the basis of past interviews and assessments, and the documentation recommends “continuing to apply to roles on the Work Platform” (Mercor, 2026c). The candidate’s interview record becomes inventory that can be re-offered to future listings.
Pooling also concentrates data. Several proposed class actions were filed against Mercor in spring 2026 after an intrusion that the complaints date to about 24 March 2026. One of them, Ananthula et al. v. Mercor.io Corporation et al., №3:26-cv-03362 (N.D. Cal., filed 21 April 2026), alleges that the exposed data included recorded AI interviews, facial biometric data and government identification documents (Hausfeld, 2026). Mercor has denied the allegations, calling the claims “speculative” and saying it complies with applicable laws (Teiss, 2026). The claims are unproven. They illustrate a cost that applicants cannot price when they enter the pool: interview recordings and identity documents held in one place are a concentrated risk.
4.3 Case C: micro1, Zara and the volume of unsuccessful interviews
micro1’s terms are explicit about the referral design. Its referral page offers “up to $3,000 for every successful referral,” paid as “flat, fixed amounts, not percentages” after the referred candidate completes 10 hours of work; members “can refer as many candidates as [they’d] like,” and “referring a candidate does not guarantee placement” (micro1, 2026a). A sub-affiliate programme pays a member “an additional bonus equal to 20% of their payout” on the referral earnings of people they recruit as referrers, limited to one level: “there are no multi-level structures or payout chains” (micro1, 2026b). The referral terms, last updated on 8 October 2026, let micro1 “withhold, delay, reduce, condition, set off against, or refuse to make any Referral Payout … in its sole and absolute discretion” and change “any payout amounts, rates, tiers, or criteria” at any time, while prohibiting referrers from making “promises, guarantees, or representations regarding employment, project placement, compensation, [or] hiring outcomes” (micro1, 2026c).
This clause places the duty not to over-promise on the referrer, who is paid to promote the claim, while the platform keeps full discretion over payment. The arrangement resembles a prospectus with full risk disclosures distributed by a sales force paid on volume; we make no claim about how referrers actually behave.
The volume of unsuccessful interviews is reported in a paper by micro1 researchers. Yazdani, Mahajan and Ansari (2025) report that over a three-day period, micro1’s platform conducted 4,820 interviews with candidates who were unsuccessful, of whom 10.7% requested feedback. Candidates “who excel in their virtual interview with Zara, progress to subsequent human interview rounds” and their profiles are added to micro1’s talent pool. Annualised mechanically, 4,820 failed interviews in three days is roughly 590,000 unsuccessful interviews per year. At 30 minutes each, the midpoint of the range stated on micro1’s current page (which may differ from interview length in 2025), that is about 290,000 hours of applicant time per year spent on interviews that did not lead to the next round. The figure is an extrapolation from three days and should be read as an order of magnitude, but it shows that the unpaid input side of the funnel is large relative to the paid side.
4.4 Case D: Scale AI’s Outlier and the gap between coupon and yield
The advertised hourly rate of a project is the coupon of the claim. What the worker earns per hour actually spent is the realised yield. Litigation against Scale AI’s Outlier division illustrates the gap. In January 2025 a former contractor, Amber Rogowicz, filed a suit under California’s Private Attorneys General Act alleging that her pay effectively worked out to about US$15 per hour, below the state minimum, that she worked about ten hours a day but was paid for about five, and that time spent reviewing instructions and training was unpaid (TechCrunch, 2025a). This followed a proposed class action filed in December 2024 alleging wage theft and misclassification. Scale said it paid the plaintiff fairly, that she was removed from Outlier for violating community guidelines, and that it “fully complies with the law.” If substantiated, these allegations would illustrate the mechanism: unpaid onboarding, training and qualification time lowers the realised hourly yield below the advertised rate, as fees and defaults do for a bond.
4.5 Comparing the referral contracts
Table 2. Referral contract design at Mercor and micro1, with an onboarding-fee benchmark (terms as published, accessed 13 Oct 2026)
micro1’s page states a 10-hour trigger in its step-by-step text and in its FAQ, but 20 hours in one image description; we use the text. The benchmark row is a contract design analysed in Section 5, not a platform. Sources: Mercor (2026a, 2026b); micro1 (2026a, 2026b).
Table 2 orders the designs by how much risk the referrer retains, and Figure 2 plots the payouts. Mercor’s earnings share is the closest to risk retention; micro1’s flat fee is contingent on 10 paid hours but is not scaled to the value of the match; a fee paid on onboarding, before any paid work, would be a pure origination fee. The model in Section 5 predicts that the lower the retention, the larger the incentive to push applicants into the funnel regardless of the chance they will be paid.
Figure 2. Referral payouts computed from the published terms. (a) Mercor pays only as the referred person earns and reaches its cap of 4× the hourly rate after 20 paid hours at $85/h; micro1 pays one flat amount, up to $3,000, after 10 paid hours. (b) Mercor’s cap at the average rate sits at the low end of both platforms’ advertised maxima.
Under the hourly-contract cap, a referral of an expert paid the reported average of $85 per hour yields at most $340, near the bottom of the “$250–$15,000” range on Mercor’s marketing page. The upper part of that range would need hourly rates of several thousand dollars, so it can only come from deliverable-based contracts, whose cap is set in the contract terms (Mercor, 2026a). The marketing page does not say how referrals are distributed across that range.
4.6 Benchmarks: ghost jobs and task scams
Two comparisons help place these funnels. At the benign end, conventional employers also post positions they do not fill. Ashby’s analysis of more than 22,000 jobs opened between January 2021 and December 2024 found that 18% of jobs open in 2024 did not lead to a hire (Ashby, 2025). Ghost postings are an old problem; the funnel studied here differs in that the pool is explicitly open-ended and the applicant’s assessment is retained and re-used.
At the pathological end, fraudulent “task scams” make the applicant pay money to unlock earnings. The U.S. Federal Trade Commission reports that losses to job scams more than tripled from 2020 to 2023 and exceeded US$220 million in the first half of 2024 alone, with task-scam reports rising from about 5,000 in all of 2023 to about 20,000 in the first half of 2024 (FTC, 2024; Figure 3). Legitimate platforms do not charge applicants money, and they do employ many people. Task scams are the extreme case in which the work sold to the applicant does not exist at all.
Figure 3. Reported U.S. job-scam losses rose from $90 million in 2020 to $286 million in 2023, and task scams grew from none in the 2020 sample to 38.8% of job-scam reports in January–June 2024. Source: FTC (2024), Consumer Sentinel data.
5. A model of over-issuance
5.1 Set-up
A platform has S paid seats in a period, each delivering worker surplus W (wage times hours, net of the worker’s effort cost) and platform margin μ. The platform draws N applicants into a funnel. Each applicant bears a cost c of profile-building, interviewing and assessment. Each passes the screen independently with probability p. Seats are filled from the pool of passers, so the number of matches is
M(N) = min{ S, pN } (1)
and the probability that a given applicant’s claim pays is
q(N) = M(N) / N = p if N ≤ S/p, and q(N) = S/N if N > S/p. (2)
Each applicant also gives the platform a value δ ≥ 0: a recorded work sample, an entry in the talent pool that can be offered to future clients, and data that improve the interviewer. Referrers are paid b_h per hire or b_o per onboarding; w denotes the hourly wage and h paid hours.
5.2 Platform’s choice
The platform’s profit from a funnel of size N is
Π(N) = (μ − b_h)·M(N) + (δ − b_o)·N. (3)
Beyond N = S/p, extra applicants add no matches, so Π rises linearly in N with slope δ − b_o. If δ > b_o and the platform’s own cost per applicant is negligible, the platform has no reason to cap the funnel, and its size is set by applicants’ participation decisions (Section 5.3). We assume risk-neutral, homogeneous applicants, a large pool so that pN passers can be treated as deterministic, a constant δ, and zero platform cost per applicant; compute costs of AI interviews and a δ that falls with N would bound the funnel. The platform’s own cost of drawing in applicants is zero when referral fees are contingent on hiring (b_o = 0), because the cost of marketing is borne by referrers and the cost of applying is borne by applicants. As under originate-to-distribute, the party that sets volume does not bear the cost of unmatched volume.
5.3 Efficient and equilibrium funnel size
Social surplus counts applicants’ costs:
Get Jung-Hua Liu’s stories in your inbox
Join Medium for free to get updates from this writer.
V(N) = W·M(N) + δ·N − c·N. (4)
Below S/p the marginal applicant adds pW + δ − c, which is positive whenever pW > c − δ. Above S/p the marginal applicant adds only δ − c, negative whenever an applicant’s time is worth more than the data they leave behind. The efficient funnel is therefore
N* = S / p. (5)
Now suppose applicants enter until the expected value of a claim equals its cost, q(N)·W = c. Using (2) for N > S/p, the free-entry funnel is
N_RE = S·W / c, (6)
the same expected-wage condition that drives migration in Harris and Todaro (1970). It exceeds N* provided pW > c, so that free entry pushes the funnel past S/p; we assume this throughout. At N_RE, applicants’ aggregate surplus W·S − c·N_RE is exactly zero: the entire wage rent of the S seats is spent on screening. This is the full-dissipation benchmark of the rent-seeking literature, which obtains under free entry with risk-neutral entrants (Tullock, 1980). The excess loss of social surplus relative to the optimum is
ΔV = (c − δ)·(N_RE − N*). (7)
If, in addition, the listing is opaque so that applicants form a belief q̂ > q(N), entry continues past N_RE and applicants’ expected aggregate surplus becomes negative. An undisclosed headcount thus plays the role of an inflated rating. Adding the platform margin μ·M and client surplus to V would not change N or ΔV, since M = S at both N and N_RE.
A calibration from published figures. In (6), the free-entry number of applicants per paid seat, N_RE/S = W/c, does not depend on the unobserved pass rate p (given pW > c). Both W and c can be bounded from public sources. Table 3 uses the reported Mercor average pay of $85 per hour (TechCrunch, 2025b), micro1’s stated interview length of 20–40 minutes (micro1, 2026a), and the 10-hour threshold at which micro1 treats a referral as successful.
Table 3. Calibrating the free-entry funnel with published figures
Because W uses gross pay and c counts only the interview, the true W/c is likely lower for a 10-hour engagement and higher for long projects. Taken as orders of magnitude, these figures imply that if applicants enter until expected pay covers interview time, a 10-hour seat attracts 15 to 30 applicants, and each claim pays with probability of roughly 3% to 7%. The micro1 figure of 4,820 unsuccessful interviews in three days is consistent with funnels of this size, though the paper does not report the number of successful ones, so the actual ratio cannot be computed from public data. Applicants whose time is worth less than $85 per hour face a lower c, so for them W/c and the implied number of applicants per seat are higher.
5.4 Redirection after failure as re-securitization
When an applicant fails, the platform often invites them to another listing. If each listing were an independent draw with pass probability p̄, the chance of at least one match after k listings would be 1 − (1 − p̄)^k, and persistence would look like diversification. But listings on one platform share an interviewer, a scoring model and, in Mercor’s case, an explicit re-use of past interview performance (Mercor, 2026c). Outcomes are therefore correlated through a persistent applicant type. Let a share π of applicants be strong, passing each listing with probability p_g, and the rest weak, passing with p_b. Then
Pr(≥1 match in k) = π·[1 − (1 − p_g)^k] + (1 − π)·[1 − (1 − p_b)^k]. (8)
To gauge magnitudes, consider a hypothetical example, since no public data report pass rates by applicant type. With π = 0.3, p_g = 0.6 and p_b = 0.05, the average pass rate is p̄ = 0.215. An applicant who believes draws are independent expects a 70% chance of a match within five listings; under equation (8) the true figure is 46%. After ten listings, belief is 91% and the truth 58%. Coval, Jurek and Stafford (2009) show that the same mistake made CDO-squared products fragile: re-pooling mezzanine tranches whose defaults are correlated produces far less diversification than independent-default models imply. The prediction is testable by comparing the pass rates of reapplicants and first-time applicants (Section 7.2).
5.5 Referral design
Referrers choose effort to recruit applicants. A referrer paid b_o per onboarding gains b_o from each recruit regardless of q, so their recruiting is independent of whether seats exist. A referrer paid a share s of the recruit’s earnings, capped at four hours’ pay as at Mercor, gains q·min{s·w·h, 4w}, which falls as the funnel grows, so recruiting slows as q falls. Under a flat fee contingent on 10 hours, recruiting depends on q but not on the size of the match. The ordering in Table 2 is therefore also an ordering of how strongly each design restrains over-issuance. Side payments from referrers to recruits, if they occur, would further inflate entry by raising the applicant’s perceived payoff.
6. Welfare, information and risk-bearing
The funnel has real benefits. Pooled screening lowers the cost of matching specialists to short projects; one interview can qualify an applicant for many listings; and AI interviewers can operate in any time zone and, as micro1’s paper shows, can generate feedback for rejected candidates. For applicants with low outside options, a small probability of a well-paid remote contract can be worth a short interview. The welfare question is who bears the cost of the excess funnel and whether that cost is visible.
The first cost is rent dissipation. As Section 5 shows, the rent of scarce, well-paid remote work tends to be competed away in unpaid screening time; the losses are small per person and large in aggregate. Part of that time produces something of value to the platform: an AI interview is a recorded work sample that improves matching and the interviewer itself, and Arrieta-Ibarra et al. (2018) argue that such inputs should be treated as labour. Applicants receive in return a lottery ticket on a seat.
The second concerns rating and risk-bearing. The platform rates applicants and profits from both their data and their placement. Its score is not portable, so an applicant who fails on one platform must repeat a similar screen on another, multiplying c. Clients can cancel or pause projects and the platform keeps the pool, while the applicant, who holds one claim on one platform, bears the default. Risk thus sits with the party least able to diversify it.
The third concerns distribution and concentration. Applicants with low opportunity cost, often outside high-income countries, enter more readily, as in Harris and Todaro (1970), so the funnel draws a disproportionate share of its unpaid inputs from workers with the fewest alternatives. Large pools of identity documents and recorded interviews are also attractive targets; the reported 2026 litigation over Mercor contractor data (Section 4.2) illustrates a risk that applicants bear but cannot price.
7. Policy implications and empirical agenda
7.1 Lessons from post-2008 securitization reform
Several post-2008 securitization reforms have direct counterparts here.
1. Listings should state the number of paid seats or expected paid hours, and the number of active applicants or prequalified candidates. A “pool ratio” of applicants per paid seat is the analogue of loan-level disclosure and corrects q̂ toward q.
2. Referral fees should be paid out of worker earnings, as in Mercor’s design, and not on onboarding. Payment on sign-up or onboarding is an origination fee that rewards volume regardless of outcome.
3. Interview scores should be portable and available to the applicant, so that one screen can be used across platforms. Portability lowers c directly and weakens the platform’s incentive to re-screen.
4. Multi-hour skill tests that produce usable work samples should be paid, which raises the platform’s marginal cost of over-issuance from zero to something positive.
5. Applicants who are not placed within a stated period should have their recordings and identity documents deleted by default, extending the deletion-on-request right in Illinois law discussed below.
Several jurisdictions already regulate parts of this pipeline. The EU AI Act classifies AI systems used “for the recruitment or selection of natural persons,” including to filter applications and evaluate candidates, as high-risk (Annex III, point 4(a)), with obligations phased in over several years; New York City’s Local Law 144 requires bias audits of automated employment decision tools; and Illinois’s Artificial Intelligence Video Interview Act requires notice, explanation and consent for AI analysis of video interviews, and deletion of videos on request. None of these addresses the volume problem identified here, which is a question of disclosure and incentives rather than of bias.
7.2 Empirical agenda
• Scrape listings and referral bounties over time and estimate the elasticity of applicant inflow with respect to bounty size.
• Survey applicants on time spent, number of listings tried, and paid hours received, to estimate c, k and q directly.
• Use changes in referral terms (for example, a move from onboarding fees to earnings shares) as natural experiments in a difference-in-differences design.
• Test the correlated-failure prediction of Section 5.4 by comparing pass rates of reapplicants with first-time applicants on the same listings.
8. Limitations and conclusion
This paper has three main limitations. First, its evidence is documentary: published terms, one company-authored research paper, press reports of litigation, and regulator statistics. It does not observe platform-level data on applicants, seats or hours. Second, the litigation we cite consists of unproven allegations, and Alignerr does not publish referral terms that could be compared with those of Mercor and micro1. Third, the model is deliberately simple. It treats seats as fixed within a period and abstracts from the platform’s reputation, which in practice limits over-issuance because disappointed applicants become disappointed reviewers.
Subject to these limits, the argument is as follows. When a platform can draw applicants into an open-ended pool at no cost to itself, earns something from each of them, and pays the people who recruit them on volume, it will issue more claims on work than there is work. The applicant buys each claim with time and data, holds the default risk, and is encouraged after each failure to buy another claim whose payoff is correlated with the last. The corresponding remedies are pool-level disclosure, risk retention by referrers, portable scores and payment for long assessments.
References
Akerlof, G. A. (1970). The market for “lemons”: Quality uncertainty and the market mechanism. Quarterly Journal of Economics, 84(3), 488–500.
Arrieta-Ibarra, I., Goff, L., Jiménez-Hernández, D., Lanier, J., & Weyl, E. G. (2018). Should we treat data as labor? Moving beyond “free”. AEA Papers and Proceedings, 108, 38–42.
Ashcraft, A. B., & Schuermann, T. (2008). Understanding the securitization of subprime mortgage credit. Foundations and Trends in Finance, 2(3), 191–309.
Autor, D. H. (2001). Wiring the labor market. Journal of Economic Perspectives, 15(1), 25–40.
Bolton, P., Freixas, X., & Shapiro, J. (2012). The credit ratings game. Journal of Finance, 67(1), 85–111.
Burks, S. V., Cowgill, B., Hoffman, M., & Housman, M. (2015). The value of hiring through employee referrals. Quarterly Journal of Economics, 130(2), 805–839.
Coval, J., Jurek, J., & Stafford, E. (2009). The economics of structured finance. Journal of Economic Perspectives, 23(1), 3–25.
Gray, M. L., & Suri, S. (2019). Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass. Houghton Mifflin Harcourt.
Harris, J. R., & Todaro, M. P. (1970). Migration, unemployment and development: A two-sector analysis. American Economic Review, 60(1), 126–142.
Horton, J. J. (2010). Online labor markets. In Internet and Network Economics (WINE 2010), Lecture Notes in Computer Science 6484, 515–522. Springer.
Kässi, O., & Lehdonvirta, V. (2018). Online labour index: Measuring the online gig economy for policy and research. Technological Forecasting and Social Change, 137, 241–248.
Keys, B. J., Mukherjee, T., Seru, A., & Vig, V. (2010). Did securitization lead to lax screening? Evidence from subprime loans. Quarterly Journal of Economics, 125(1), 307–362.
Montgomery, J. D. (1991). Social networks and labor-market outcomes: Toward an economic analysis. American Economic Review, 81(5), 1408–1418.
Pallais, A. (2014). Inefficient hiring in entry-level labor markets. American Economic Review, 104(11), 3565–3599.
Rochet, J.-C., & Tirole, J. (2003). Platform competition in two-sided markets. Journal of the European Economic Association, 1(4), 990–1029.
Spence, M. (1973). Job market signaling. Quarterly Journal of Economics, 87(3), 355–374.
Tullock, G. (1980). Efficient rent seeking. In J. M. Buchanan, R. D. Tollison, & G. Tullock (Eds.), Toward a Theory of the Rent-Seeking Society (pp. 97–112). Texas A&M University Press.
Yazdani, N., Mahajan, A., & Ansari, A. (2025). Zara: An LLM-based candidate interview feedback system. arXiv:2507.02869 [cs.HC]. https://arxiv.org/abs/2507.02869
AIGigJobs. (2026). Alignerr review & guide 2026: Pay, interview, how to apply. https://www.aigigjobs.com/platforms/alignerr
Ashby. (2025). Ghost jobs: Talent trends report. https://www.ashbyhq.com/talent-trends-report/reports/ghost-jobs
Federal Trade Commission. (2024, December). Paying to get paid: Gamified job scams drive record losses (Data Spotlight). https://www.ftc.gov/news-events/data-visualizations/data-spotlight/2024/12/paying-get-paid-gamified-job-scams-drive-record-losses
Mercor. (2026a). Referral policy and FAQs. https://talent.docs.mercor.com/policies/referrals
Mercor. (2026b). Referrals. https://www.mercor.com/referrals/
Mercor. (2026c). Instant work offers. https://talent.docs.mercor.com/new-releases/instant-offer
Hausfeld LLP. (2026). Mercor data breach (case page for Ananthula et al. v. Mercor.io Corporation et al., №3:26-cv-03362, N.D. Cal.). https://www.hausfeld.com/what-we-do/current-claims/mercor-data-breach
micro1. (2026a). Referral program. https://micro1.ai/referral-program
micro1. (2026b). Sub-affiliate program. https://www.micro1.ai/sub-affiliate
micro1. (2026c). Referral terms (last updated 8 October 2026). https://www.micro1.ai/legal/referral-terms
micro1. (2026d). Zara: AI recruiter. https://micro1.ai/zara
Teiss. (2026, April 24). Mercor faces class-action lawsuits after data breach exposes contractor information. https://www.teiss.co.uk/news/mercor-faces-class-action-lawsuits-after-data-breach-exposes-contractor-information-17404
TechCrunch. (2025a, January 9). Scale AI hit by its second employee wage lawsuit in less than a month. https://techcrunch.com/2025/01/09/scale-ai-hit-by-its-second-employee-wage-lawsuit-in-less-than-a-month
TechCrunch. (2025b, October 27). Mercor quintuples valuation to $10B with $350M Series C. https://techcrunch.com/2025/10/27/mercor-quintuples-valuation-to-10b-with-350m-series-c/
