Sources#

• Do job seekers value procedure in AI hiring only for error correction? Evidence from a conjoint experiment
• Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews

Choosing the AI Interviewer While Paying for a Human Decider#

The question#

The #oq/now bullet on Procedural Value in AI Decisions asks this. The page already reconciles the sign: the conjoint prices decision authority, and the field experiment never varied it, because humans made every hiring decision in both arms. That leaves the magnitude. A +0.272 premium for a human decider sits beside 78.4% of applicants choosing a machine. Is that residual population (US Prolific vs Philippine BPO applicants) and stakes (a hypothetical forced choice vs a live application)? Or would US job seekers facing a real application also take the AI?

Short answer#

The residual is not population and stakes. Most of it is not a residual at all. Restate the conjoint on the object the field experiment actually varied, and its stated penalty is 0.030, not 0.272. That leaves the field's tilt toward the AI to be explained by AI-side pulls the conjoint never priced: speed, and a choice protocol that puts the AI agent on the line. Population and stakes change how large the majority is. Within the vault's evidence they do not turn it into a minority, and real stakes push toward the AI.

So on the vault's evidence, US job seekers with a real application would most likely also take the AI by a majority, provided a human still decides and the AI is faster. Nobody has measured the US level, and that measurement is the part of the question that needs an outside source.

1. Put both numbers on the same object#

The conjoint's marginal means (Do job seekers value procedure in AI hiring only for error correction? Evidence from a conjoint experiment Table C3, reconciled on Procedural Value in AI Decisions):

The +0.272 is the jump from the bottom row to the top two. The field experiment's applicants never faced that jump. Whichever interviewer they picked, a human recruiter reviewed the audio, transcript and test scores and made the call (Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews §2.4, "Assessment across treatments"). What they chose was how much of the process the machine ran under unchanged human authority. The only conjoint contrast with that shape is between the top two rows: the AI takes over more of the funnel, and the human keeps the decision. Respondents paid 0.030 to avoid that (p < 0.001), about a ninth of the authority premium. In an exploratory test that margin did not move with the error rate either.

The proxy is imperfect. "AI screens first" means the human never sees the applications the AI rejects, which is more consequential than an AI conducting an interview a human then reviews in full. If anything, 0.030 therefore overstates the stated penalty on the field's object. The conjoint has no interview-modality attribute at all, so this is the nearest contrast available, not a measurement of modality.

On that object, the comparison is a 3-point stated penalty against a +28-point revealed tilt (78.4% vs an even split). A tilt that large calls for a positive pull toward the AI, which a smaller human premium cannot supply. The conjoint cannot produce one, because every one of its attributes is costless and none is speed.

2. What pulls toward the AI in the field, and why the conjoint can't see it#

The field paper names the two pulls itself (Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews §5.3, "Predicting choices"):

• Convenience. AI interviews can be taken at the applicant's chosen time, including immediately. Time to interview fell from 0.51 to 0.32 days (Controlled Variance: AI's Edge as Reduced Dispersion). The conjoint holds timing fixed by construction, because every feature it offers is free and instant.
• Attitudes. 47% of survey respondents expected AI to affect them positively at work, against 19% negative, and the attitude predicts choice: 77% / 72% / 65% chose the AI among the positive, balanced and negative groups (Controlled Variance: AI's Edge as Reduced Dispersion, The Automation–Optimism Link).

A third pull is in the raw and not yet on any concept page: the choice protocol leans toward the AI. Applicants who did not choose within a set time frame were called by the AI voice agent. It explained the process, described how it worked, and then asked them to choose. The firm built this deliberately to give applicants a first exposure to the agent and reduce apprehension (raw §2.4, footnote 10). The paper does not report how many choices were made during that AI call. For those applicants, though, the revealed choice was made while already talking to the machine, with "keep talking to it" as the zero-effort option. The non-response rate after that call was small (1.4% walk-in, 2.7% remote), so the AI-call cohort's size is unknown, but the default runs one way.

3. How far population and stakes move the number#

These are the within-vault handles on each factor. None of them reverses the majority.

• The convenience margin, inside the field experiment. Walk-in applicants, already at a recruitment centre where a human could interview them in person, chose the AI 69.3% of the time; remote applicants chose it 81.6% of the time (Controlled Variance: AI's Edge as Reduced Dispersion). Narrowing the convenience gap moves the share by about 12pp and leaves a clear majority.
• The attitude margin, inside the field experiment. Even the 19% who expected AI to hurt them chose it 65% of the time. A US population more sceptical than the Philippine sample would need to sit well below the field's own pessimist group to fall under half. The vault holds no US interview-choice-by-attitude data that would place it there.
• The population margin, inside the conjoint. The authority premium barely varies across the subgroups measured: prior AI evaluation +0.269 vs none +0.282 (interaction -0.013, p = 0.36), no detectable age interaction, women +0.289 vs men +0.253 (Do job seekers value procedure in AI hiring only for error correction? Evidence from a conjoint experiment Table E1, Appendix H). The gender signs cross between instruments. US women price human authority more than men, while Filipino women choose the AI interviewer more than men (80% vs 76%, p < 0.001). The populations differ, so this proves nothing alone. But it is the pattern expected if the instruments measure different objects, and the opposite of what one shared aversion to AI expressed at two strengths would produce.
• The stakes margin. Both studies point the same way. In the conjoint, intention to apply (3.24) exceeds belief in legitimacy (2.85), and 7.8% doubt AI hiring yet intend to apply (Procedural Value in AI Decisions). Within person, the human-authority preference correlates with application intention at |r| ≤ 0.14 (Appendix H). A respondent's stated authority premium barely predicts even their stated intention to apply, let alone a behavioural choice. With a real application at stake, the measure moves toward compliance, not refusal. So stakes predict a higher US AI share than a no-stakes vignette would, not a lower one.

4. The vault's other stated-vs-revealed pairs follow one rule#

The question asked for this check. Each pair the vault holds either agrees or has its disagreement traced to an object or population mismatch:

The dissertation pair is the one counter-weight, and it is worth stating plainly: there, a stated/revealed gap that looked like attitudes failing to govern behaviour turned out to be population. So population mismatch can produce an apparent gap. In the hiring pair, though, the population difference sits on top of an object difference that already explains the sign and most of the magnitude (§1). Of the two, the object mismatch is the one doing the work.

Verdict#

• Settled by synthesis. Population and stakes do not explain the magnitude gap. On a common object, the stated penalty is 0.030, and the field's tilt comes from AI-side pulls the conjoint never priced (immediacy, an AI-led choice default, positive attitudes). Population and attitudes move the revealed share within roughly a 12pp band in the field's own data. Stakes push the revealed share toward the AI. The vault's other stated-vs-revealed pairs agree when object and population match.
• Prediction, not measurement. US job seekers facing a real application, with a human keeping the decision and the AI offering faster scheduling, would choose the AI by a majority. That prediction would fail if any of three conditions held: the AI is offered without a speed advantage, the AI holds decision authority (which the conjoint prices at 0.272), or the choice is elicited by a human rather than by the agent.
• What would close it. A US field interviewer-choice arm under retained human authority, or US platform opt-in data for AI interviews. Alternatively, a conjoint that adds interview modality and time to interview as randomized attributes alongside decision authority. That would price the field's object directly, and the conjoint's own costless-feature limitation invites it. The share of Jabarian & Henkel's choices made on the AI's call is also unreported and would size the default effect.

Sources#

• Procedural Value in AI Decisions and Do job seekers value procedure in AI hiring only for error correction? Evidence from a conjoint experiment: Table C3 marginal means (0.604 / 0.575 / 0.320), the +0.030 within-authority contrast, Table E1 subgroup AMCEs, Appendix H (prior-experience interaction, gender +0.037, |r| ≤ 0.14 convergence check), §4.4 legitimacy vs intention (2.85 vs 3.24, 7.8%).
• Controlled Variance: AI's Edge as Reduced Dispersion and Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews: §2.4 treatment definitions (human evaluation in all arms; the Choice condition's AI-call default and footnote 10); §5.3 choices (69.30% walk-in, 81.63% remote, 78.41% pooled; non-response 49/3,469 and 268/9,927), attitude gradient and gender split; §7.1 time to interview.
• The Automation–Optimism Link: the belief→delegation gradient and its compliance-vs-endorsement caveat.
• AI Adoption in Scientific Work, AI Writing in PhD Dissertations: the population-matched comfort-vs-behaviour pair.
• Telemetry vs. Survey Measurement: the ICONIQ/Ramp convergence and the three-instrument table on the hiring experiment.
• Cost-per-Task Over Cost-per-Token: the a16z vs Ramp object mismatch.
• Arithmetic here (0.030 as about a ninth of 0.272; +28pp as 78.4% minus 50%; the 12pp walk-in/remote spread; the non-response shares) is the vault's own, from the cited figures.