A candidate who replies to your hiring manager one hour late is about 46% less likely to get hired than one who replies immediately. A five-to-ten minute delay already moves the number. A full day cuts hire likelihood by roughly 90% (Hart, VanEpps, Sezer & Amir, Management Science, 2026).
None of that is about what the candidate wrote.
The finding comes from approximately 11.66 million real marketplace hiring conversations plus nine experiments with more than 8,600 participants, and it survives every control you would reach for: ratings, reputation, prior reviews, and the actual content of the reply (Cornell SC Johnson, 2026). This is hiring response time bias in its purest form โ an unvalidated signal that reaches the evaluator, changes the decision, and leaves no trace in the scorecard.
The uncomfortable part is not that the bias exists. It is that it operates below stated policy. In the experiments, the effect held among managers who explicitly said they did not expect a reply within an hour.
What the Study Actually Measured
The researchers paired transaction data from a large service marketplace with controlled experiments across occupations โ photographers, caterers, doctors, full-time roles (Vanderbilt Owen, 2026). The marketplace data gives scale and realism; the experiments give causal identification. Neither alone would be persuasive. Together they are hard to dismiss.
The mechanism is an inference chain, and it is worth naming precisely because it is where the fix lives. Fast responders are perceived as warmer, as more competent, and โ most consequentially โ as more likely to be responsive in the future (UC San Diego, 2026). That third inference is the one that feels defensible to a hiring manager. It sounds like a prediction about work behaviour rather than a reaction to a timestamp.
It isn't. The study found no occupation that moderated the effect, and it found the effect intact for candidates with poor ratings and for candidates with no ratings at all. If latency were a proxy for observed quality, it would weaken when quality information is present. It doesn't.
Why Hiring Response Time Bias Survives Your Policy
Every mid-market hiring process I have reviewed has some version of a rubric. Most have structured questions. A few have calibration sessions. All of them assume the evaluator is scoring what is on the page.
The rubric governs what the evaluator writes down. It does not govern what the evaluator sees before they write anything down โ and in a threaded inbox or an ATS conversation view, the first thing they see is elapsed time.
This is why the finding is a selection-validity problem and not a training problem. Training corrects beliefs. The managers in this study already held the correct belief; they stated it out loud and then behaved as though they hadn't. A bias that survives explicit disavowal is not going to be dislodged by a slide deck on unconscious bias.
The comparison worth holding in mind: a 2025 meta-analysis of interview-based assessment across 37 studies and 30,646 participants puts criterion-related validity for task performance at ฯ = .30 and contextual performance at ฯ = .28 (Wingate, International Journal of Selection and Assessment, 2025). Those are the good signals โ the ones you spent money and calendar time to collect. Reply latency has no established criterion validity at all, and it is sitting in the same decision, weighted by nobody, audited by nobody.
What Latency Correlates With Instead
Strip the inference away and ask the empirical question: what actually determines whether someone answers a message within the hour?
- Time zone. A candidate three zones east reads your 4pm message the next morning. That is a geography score, not a fit score.
- Caregiving load. School pickup is a fixed 45-minute latency window, applied unevenly across your applicant pool.
- Current employment. The candidate who replies in four minutes may be between roles. The one who replies at 8pm is in a job and respecting it.
- Notification settings. Push notifications on a personal phone versus batched email. This is a preference about attention hygiene โ arguably a positive work trait โ scoring negatively.
- Seniority. More senior candidates have denser calendars and fewer free interstitial moments โ so the signal runs mildly against experience.
- Device and connectivity. A candidate on a laptop-only workflow is structurally slower than one on a phone, regardless of how much they want the role.
None of these predict performance. Two of them correlate with protected or quasi-protected characteristics closely enough that a regulator would find the pattern interesting. And because latency never appears on a scorecard, this contamination is invisible to any audit you currently run.
The Practical Test
Pull your last twenty hires and your last hundred rejections at the same funnel stage. Compute median time-to-first-reply for each group. If the hired cohort is faster by a material margin, you have not found evidence that responsive people perform better. You have found the timestamp gradient reaching your evaluators โ which is exactly what the study predicts, and exactly what your rubric was supposed to prevent.
The 2026 Twist: The Signal Is Both Gameable and Brittle
Here is where this stops being a 2019 bias-training story.
The study's second finding is that the speed advantage disappears when the reply reads as automated or AI-generated (Hart, VanEpps, Sezer & Amir, Management Science, 2026). Speed also does not rescue a badly written reply. Sezer's summary of the boundary condition: the durable advantage is not speed alone, it is being quick and unmistakably human.
Now put that next to the candidate market as it exists this year. Any applicant with a drafting assistant can manufacture the exact latency profile your evaluators unconsciously reward โ sub-five-minute, well-composed, on every message. So the signal is simultaneously corruptible (the candidates who game it are not the candidates you were trying to identify) and brittle (the ones who use AI slightly too visibly are penalised for the tool, not the substance).
A signal that can be faked by the sophisticated and punishes the transparent is worse than no signal. It is a signal with inverted validity, and it is currently running unmonitored inside your funnel.
There is a second-order cost here that ops teams tend to price at zero. Every unvalidated variable you allow into a selection decision consumes decision weight that something validated could have used. The evaluator has finite attention and a finite impression to form. If part of that impression is being set by elapsed minutes, it is not being set by the work sample, the structured answer, or the reference. You did not add noise to a clean signal. You displaced signal with noise.
The Objection: "We Genuinely Want Responsive People"
This is the fair pushback, and it deserves a real answer rather than a dismissal.
Responsiveness is a legitimate job requirement for some roles โ incident response, client service, sales coverage. If it matters, it should be assessed. But there is an enormous difference between assessing responsiveness and absorbing latency.
Assessing it means: naming it as a competency, defining the standard, testing it under conditions you control, and scoring it on the rubric where it can be weighted and audited. Absorbing it means: letting an artefact of the candidate's Tuesday afternoon silently adjust an evaluator's impression of their warmth and competence, with no record that it happened.
The first is selection design. The second is noise wearing selection design's clothes. If responsiveness genuinely matters for the role, the study gives you more reason to formalise it โ not less โ because it demonstrates how powerfully the untested version distorts the decision.
Two Process Controls, Deployable This Quarter
The authors propose procedural fixes rather than attitudinal ones, which is the correct instinct. Both are cheap.
1. Publish the reply window to candidates. State it in the outreach: "Please reply within 48 hours; we review all responses together." This does two things at once. It removes the ambiguity that lets latency carry inferential weight, and it equalises the playing field across time zones and life circumstances. Candidates stop being scored on a rule nobody told them existed.
The failure mode to watch for: teams implement the first control and skip the second, because publishing a reply window feels like it solves the problem. It doesn't. Telling candidates the window changes candidate behaviour; it does not stop the evaluator from seeing that Candidate A answered in three minutes and Candidate B answered in six hours, both comfortably inside the stated window. The inference survives the announcement. Only the second control removes the input.
2. Batch-review responses after a fixed delay. Do not let evaluators read replies as they arrive. Collect them, strip or suppress the timestamps, and review the set together. The evaluator never sees the gradient, so the gradient cannot be inferred. This is the single highest-leverage change, and for most teams it is an ATS view configuration plus a norm โ not a project.
A third control, if you want the measurement: instrument time-to-first-reply as a monitored variable in your funnel data. You are not trying to optimise it. You are trying to detect whether it predicts advancement. If it does, you have a live validity leak and now you can see it.
The Decision to Make Before the Next Requisition Opens
Hiring response time bias is not a story about candidates being judged unfairly, though they are. It is a story about a selection process quietly importing an unvalidated variable with a larger effect on outcomes than several of the signals you deliberately built the process to capture.
You cannot train it out. The managers in the study already knew better and it made no difference.
So the decision this quarter is narrow and concrete: will your evaluators see reply timestamps, or won't they? Everything else about your hiring rubric assumes the answer is no. Right now, in almost every mid-market funnel I have looked at, the answer is yes โ and the 46% is being paid by candidates you never seriously considered, for a reason nobody wrote down.
