AI Advice in Hiring: Wages, Surplus, and the Gender Gap

For experimental details, see the Experimental Design (Public) section.

2026-10-10

2026-11-30

H1. We test the effect of AI assistance on the wages employers offer.

H2. We test whether the effect of AI assistance on the wages employers offer differs by worker gender.

H3. We test the effect of AI assistance on the expected surplus of employers.

H1. We first test whether the level of the AI’s recommendations differs from the level of employers’ offers. For each worker, we compute the mean no-AI offer of employers and the mean AI recommendation shown to employers. Pooling these worker-level means, we regress wages on an AI-source indicator (i.e., a dummy equal to one for the AI’s recommendation and zero for employers’ offers), with worker fixed effects and standard errors clustered at the worker level. We then test the AI effect within employers: we regress the wage offer on the worker attributes (a female indicator, age, education, preferred subject, and three warm-up correctness indicators) and the AI indicator (i.e., a dummy equal to one when the employer has access to the AI’s recommendation while setting the wage), with employer fixed effects and standard errors two-way clustered at the worker and employer levels.

H2. We first test whether the gender gap in the AI’s recommendations differs from the gap in employers’ offers. Pooling the same worker-level means as in H1, we regress wages on the worker attributes, the AI-source indicator, and the AI-source indicator interacted with each attribute, with standard errors clustered at the worker level. We then test the AI effect within employers: we regress the wage offer on the worker attributes, the AI indicator, and the AI indicator interacted with each attribute, with employer fixed effects and standard errors two-way clustered at the worker and employer levels.

H3. We first test whether the AI’s recommendations are more accurate than employers’ offers. For each worker, we compute the mean expected bonus lost (see its definition in the Experimental Design section) of employers’ no-AI offers and, separately, of the AI’s recommendations shown to employers. Pooling these worker-level means, we regress the loss on the AI-source indicator, with worker fixed effects and standard errors clustered at the worker level. We then test the AI effect within employers: we regress the employer’s expected surplus (see its definition in the Experimental Design section) on the AI indicator, with employer fixed effects and standard errors two-way clustered at the worker and employer levels.

Each regression uses the offers from Sessions 1 and 2 (see the session structure in the Experimental Design section), pools the real and simulated setups (in the real setup the wage offer actually affects the hired worker’s bonus, whereas in the simulated setup the offer is hypothetical; employers are told their setup and domain at the start of the experiment) and is run separately by domain: math and literature are two separate pre-registered tests of each hypothesis, and we report both.

H4. We test the effect of AI assistance on the expected surplus of workers.

H5. We test whether the effect of AI assistance on wage offers, the gender gap, and employer and worker surplus persists after the AI is withdrawn.

H6. We test whether the accuracy gain in employers’ offers under AI assistance, if any, comes from less disagreement across employers (variance) or from a better typical offer (bias).

We also consider several exploratory analyses:

E1. We explore the heterogeneity of the AI effects in H1–H4 across employer characteristics (gender, age, education, AI usage).

E2. We explore whether AI assistance narrows the gender gap in employers’ offers when one exists, and whether this differs across the two domains.

E3. We explore whether AI assistance changes how wage offers depend on the worker attributes other than gender (age, education, preferred subject, and warm-up performance), using the AI indicator interactions already included in the within-employer H2 regression.

E4. We explore how much of the AI’s advice employers adopt, and how adoption varies with whether the recommendation carries an explanation, whether it confirms or challenges the employer’s prior ranking of the worker’s gender (from their earlier offers or predictions), and employer characteristics (gender, age, education, AI usage).

E5. We explore how the real versus simulated setup shapes behavior: first the offer level, gender gap, and surplus in employers’ initial no-AI decisions (before any AI exposure), then the AI effect on each.

E6. We regenerate the recommendations using additional frontier LLMs and compare each model’s offer accuracy and gender gap.

H4. We regress the worker’s expected surplus (see its definition in the Experimental Design section) on the AI indicator, with employer fixed effects and standard errors two-way clustered at the worker and employer levels, separately by domain. We omit data from the simulated setup.

H5. To each of the within-employer regressions in H1–H4 we add the Session 3 offers and a post indicator (i.e., a dummy equal to one for these offers), entering each regression exactly as the AI indicator does.

H6. For each worker, the mean expected bonus lost decomposes into a variance part (from the dispersion of offers across employers) and a bias part (from the gap between the worker's mean offer and the worker's true value); we compare the Session 1 (no-AI) and Session 2 (with-AI) offers on each part. Inference is based on a nonparametric bootstrap that resamples workers with replacement and recomputes the full decomposition on each draw; we report percentile intervals for the AI-induced change in the variance part, in the bias part, and in their difference. Run separately by domain.

This experiment runs in two waves. In the first wave, a pool of job “workers” complete test questions and a background questionnaire; their performance and attributes generate the profiles shown to “employers.” This wave tests no hypothesis. In the second wave, employers decide how much to pay to hire these workers.

Worker task. Workers answer questions in two domains, math and literature (5 warm-up and 20 official questions each), and report gender, age, highest education, and preferred school subject; country is recorded by Prolific. A profile displays a worker’s gender, country, age, education, preferred subject, and answers to three fixed warm-up questions in the relevant domain (chosen from the five warm-ups using the realized worker pool, before the employer wave).

Employer task. Employers are told their randomized setup (real or simulated) and domain (math or literature) in the hiring-task instructions. For each worker, an employer sets the maximum wage (0–100 cents) they would pay to hire that worker. The worker is hired if this wage is at least a price drawn at random between 0 and 100 cents; if hired, the employer earns 5 cents per correct answer the worker gave on the 20 official questions in the employer’s domain, minus the price. Each employer evaluates 18 distinct workers across three sessions of six (drawn at random from the pool, gender-balanced with three male and three female per session): Session 1 without AI; Session 2 with AI; Session 3 again without AI. In Session 2, the employer additionally sees, for each worker, a recommended wage from a commercial AI chatbot (GPT-6.1 Sol, default parameters) that observes the same profile as the employer.

We vary three treatments, two between employers and one within. The between-employer treatments are:
- Setup: in the “real” setup, employers are told “the hiring is real, so the worker you hire actually receives the money you pay”; in the “simulated” setup, “the hiring is simulated, so the worker you hire never actually receives the money you pay.”
- Domain: each employer’s task and payment concern either math or literature; the profile shows only that domain’s warm-up performance.

The within-employer treatment is explanation: whether the AI’s recommendation shows a brief explanation of its reasoning or the wage number alone. For every employer, two of the three workers per gender show the explanation, and one shows the number alone (four of the six AI recommendations carry an explanation); which workers is randomized within gender.

The AI recommendations are generated offline in advance with default parameters. For every profile, we additionally generate offline the recommendation under the opposite gender label, holding all else fixed, and we regenerate recommendations under additional frontier LLM models; these support exploratory analyses.

After the 18 decisions, employers complete a prediction task: they predict the average number of correct answers (out of 20) for each of nine worker groups (male; female; education above and below a fixed threshold; younger and older than a fixed age threshold; and favorite subject math, science, or English; both thresholds are set from the realized worker pool before the employer wave). Finally, employers answer a background questionnaire (gender, age, education; hiring experience; AI-usage frequency; trust in AI generally and in the task recommendations; and whether the AI’s recommendations reflected worker performance). We give employers two attention checks and drop from analysis any employer who fails either.

Payments. Employers receive a show-up payment; in addition, one of each employer's 27 decisions (18 hiring, 9 prediction) is selected at random for that employer's bonus. If a hiring decision is selected, a price p is drawn at random on [0, 100]: if w >= p, the worker is hired at price p and the employer earns 5 cents per correct answer the worker gave on the 20 official questions, minus p; otherwise, no hiring. A 100-cent buffer is added whenever a hiring decision is selected (so the realized bonus for employers is never negative; it does not affect the decision-relevant surplus). If a prediction is selected, we use a binarized quadratic scoring rule: the employer receives a 200-cent bonus with probability max(0, 1 - (p - t)^2 / 100), where p is the prediction and t is the true group average (both out of 20). Workers receive a show-up payment for the first-wave task and, in the real setup, an additional bonus equal to the sum of the prices p from every selected-for-payment hiring decision in which a real-setup employer hires them (w >= p); in the simulated setup, they receive no such bonus.

Outcome measures. Write w in [0, 100] for the wage offer (cents) and V = 5s for the worker's true value (cents), where s is the number of correct answers (out of 20) in the employer's domain. Per decision and in expectation over the BDM price (buffer excluded), the worker surplus is w^2/200 cents, the employer surplus is Vw/100 - w^2/200 cents, and the expected bonus lost (inverse accuracy; lower is more accurate) is (V - w)^2/200 cents.

Not available

All randomization is done by computer.

We randomize the domain-by-setup cell (math/real, math/simulated, literature/real, literature/simulated) between employers using blocked randomization with blocks of 4 (one employer per cell). Within each employer, we randomize the draw of the 18 workers, their assignment to the three sessions of six, and their order of presentation, subject to three male and three female per session. Within each employer, we randomly choose, for each gender, which two of the three workers show the AI’s explanation. Within each employer, we randomize the order of the nine prediction groups.

The individual (employer).

No

Workers: 400 participants. Employers: 800 participants.

14,400 wage offers (800 participants × 6 worker profiles × 3 sessions) and 7,200 predictions (800 participants × 9 prediction questions). At most 13,440 AI recommendations per model (800 participants × 6 worker profiles × 1.4 attrition buffer × 2 gender labels, actual and counterfactual). The attrition buffer ensures enough recommendations are prepared offline, since employers who start but drop out still consume recommendations.

Of the 800 employers, approximately 200 are assigned to each of the four domain-by-setup cells. AI exposure varies within employers: each employer makes six no-AI offers in Session 1, six AI-assisted offers in Session 2, and six no-AI offers in Session 3, with three male and three female workers per session. Of the 800 × 6 = 4,800 Session 2 offers, two-thirds (3,200) display the AI’s explanation and one-third (1,600) display the number alone, balanced by worker gender.