Candidate Voice Report 2026: Drop-Off Rates, Fairness Perceptions, and What 2,500+ Recent Applicants Want from AI Hiring

The Headline Picture

The most important finding in this year's data only appears when you broaden the comparison set. Most public discussion of AI in hiring frames the question as 'AI screen versus a thoughtful human interview.' For applicants outside the resume-shortlist tier — the majority of applicants for the majority of roles — that's not the comparison their employer is actually making. The realistic alternative to an AI touchpoint is no touchpoint at all: a resume goes into a queue, gets cursory or no review, and the applicant either receives a templated rejection days later or never hears back. The 614-respondent no-AI baseline cohort in this study exists to put numbers on that comparison. Measured against it, candidates who experienced any AI touchpoint rated their overall experience 0.6 points higher and were more than twice as likely to know what actually happened to their application. A well-deployed AI screen gives applicants a chance to tell their story that the most common alternative simply doesn't.

Within AI, modality matters — particularly on completion. Voice AI shows the lowest mid-process abandonment (9%), followed by async video (16%), skills assessment (22%), and chat AI highest at 28%. The 19-point gap between top and bottom is meaningful. But the gap between vendors within a single modality is larger still. In our editorial testing of 14 voice AI platforms, overall-experience scores spanned 1.8 points from best to worst (4.2 to 2.4). For the 11 async video platforms tested, the spread was 1.6 points (3.7 to 2.1). Both within-modality vendor spreads are wider than the 0.8-point cross-modality range across the four category averages. Which vendor an employer chose, and how thoughtfully that vendor handled edge cases, mattered more to the candidates we surveyed than which modality the employer chose.

Communication is the deployment-side lever that turns any quality vendor into a strong candidate experience — disclosure clarity, time-to-response, and a clearly available human alternative move overall-experience ratings more than the modality choice does within the same disclosure pattern. The candidate side of AI recruiting is a vendor-quality and deployment problem more than it is a category problem.

The Baseline: AI Screening vs. No Screening

The 614-respondent baseline cohort applied to roles in the same 90-day window through a process that included no AI touchpoint of any kind. The cohort was matched to the primary sample on role type, headcount band, and industry category. The numbers below compare the average AI-screened applicant across modalities (n=2,587) to that no-AI baseline (n=614).

The gap is largest on the dimensions most associated with feeling considered: knowing the application was actually reviewed, having a chance to demonstrate something beyond what's on the resume, and getting a definitive outcome at all. The no-AI baseline isn't the no-friction ideal — it's the silent-rejection reality. AI screening, when it comes from a quality vendor and includes basic disclosure and outcome communication, clears that bar by a wide margin.

Mid-Process Abandonment by Modality

We asked respondents to anchor on a single specific recent AI touchpoint and report one of three states: they did not start the touchpoint despite being invited, they started it but did not complete it, or they completed it. The middle column — started but did not complete — is the operational mid-process abandonment metric and the most actionable signal for buyers, because it represents candidates the employer has already lost effort and brand exposure on. Values are percentages of all respondents in the modality cell, and rows sum to 100%.

Procedural Fairness Ratings by Modality

Respondents rated each AI touchpoint on a 5-point scale across four dimensions: procedural fairness ('the process was fair'), informational clarity ('I understood what was being evaluated'), accommodation ('it was easy to request an alternative if I needed one'), and overall experience. Process-quality questions were asked before respondents revealed or were asked about outcome, to separate experience from result.

The Disclosure Effect

We asked respondents two binary questions about how the process was communicated: was AI use disclosed before the touchpoint started, and was a human-alternative option clearly offered. We then compared overall experience ratings across the four resulting cells. The pattern is the largest single deployment-side effect in the dataset — and it holds inside every modality.

Vendor Quality Variance Within Modality

Modality-level averages obscure a much larger story: the spread of candidate experience between vendors inside a single modality is wider than the spread between modalities. We see it in two places.

In the survey itself, in the minority of follow-up interviews where the vendor behind the employer's process was identifiable, within-modality experience ratings spanned more than a full point on the 5-point scale for every modality with sufficient identification. Most candidates can't reliably name the platform an employer used, which is why the vendor-by-vendor picture leans on editorial testing rather than the survey.

In Recruiting Tech Reviews' editorial testing program — 60 AI recruiting platforms reviewed in the 2026 cycle, with hands-on candidate-side testing where the modality permits — the variance is more pronounced. For voice AI specifically, overall experience scores across 14 platforms ranged from 4.2 at the top of the category to 2.4 at the bottom. For async video, the spread across 11 platforms was 3.7 to 2.1. In both modalities, the vendor spread is wider than the cross-modality average gap.

Four vendor-side factors track most closely with the high end of the range:

1. Quality of the underlying AI and naturalness of dialogue. Robotic-sounding text-to-speech, awkward pauses, broken turn-taking, and noticeable response latency concentrated in lower-rated platforms. The platforms at the top of voice AI testing handled interruptions, clarifying questions, and pace adjustments without breaking the conversational frame.

2. How edge cases are handled. Lower-rated platforms struggled with predictable edge cases — heavy accents, unstable connectivity, candidates asking off-script questions, requests to repeat or rephrase. Higher-rated platforms had thought through these paths in advance and degraded gracefully when something went wrong, including offering a clean fallback to a human when the AI couldn't proceed reliably.

3. Instructional design and onboarding clarity. Platforms that opened with a clear setup — what to expect, how long it would take, what was being evaluated, what to do if something went wrong — rated meaningfully higher than platforms that dropped candidates into the experience without context.

4. Scoring methodology. Platforms using documented rubric-based scoring — explicit, defined criteria the candidate could be told about in advance — rated higher on perceived fairness than platforms using opaque algorithmic-summary scoring, by 0.5 points within voice AI and 0.7 points within async video.

What Candidates Praise, and Where Frustration Concentrates

Two reviewers worked through the open-ended responses independently and identified the same themes coming up over and over.

Three praise themes dominated the data. Candidates valued the convenience of being able to self-schedule and self-complete on their own time, a benefit that came through strongest in chat AI and skills assessment. They valued getting a clear, prompt outcome regardless of which way the decision went, which was raised in every modality. And, inside voice AI specifically, candidates frequently described conversations as feeling natural and as giving them room to ask clarifying questions — a theme that was heavily concentrated among candidates who happened to be interviewed by the higher-rated voice AI vendors in our editorial testing.

The themes candidates raised most often as frustration:

Silence after completion. 'I never heard back' is by a wide margin the most common complaint in the dataset. 44% of candidates rejected at an AI stage said this, along with 22% of candidates who advanced past it but lost visibility into the process afterward. This is a configuration choice on every major platform, not a product limitation. Among rejected applicants in the no-AI baseline cohort, the same complaint reaches 71%. AI candidates are markedly less likely to be left in the dark, but neither share is acceptable.

Discomfort with the AI itself. Robotic-sounding voices, awkward dialogue, latency or technical glitches mid-interview, and unclear evaluation criteria appear in roughly 38% of the 1,142 negative open-ended responses we coded. The complaints are unevenly distributed across vendors: in editorial testing they clustered heavily in the lower-rated platforms in each modality and were sparse in the higher-rated platforms. Vendor-quality differences explain more of this frustration than the modality choice does.

Surprise about AI use after the fact, with no advance disclosure: 28% of the 2,587 candidates we surveyed.

Inability to ask a real person a question when something went wrong: 33% of the candidates we surveyed, concentrated in deployments without a visible human-alternative path.

The pattern is hard to miss. The most frequently-cited candidate frustrations aren't about the existence of AI in the process. They're about communication, vendor quality, and accessible escalation paths — process and product features that vary widely between vendors, and even between deployments of the same vendor.

Where Demographic Variance Shows Up

We tested for meaningful demographic differences in process-experience ratings, segmenting on age, role type, prior experience with AI hiring, accommodation needs, and self-reported demographic categories aligned with U.S. EEOC and UK Equality Act conventions. Findings are reported as candidate perceptions of process — they are not adverse-impact analyses of the underlying AI systems.

The meaningful differences:

Age. Candidates 50+ rated async video lower on procedural fairness (2.6 vs. 3.2 for under-50) and informational clarity (2.4 vs. 3.0). Voice AI ratings were closer across age bands. Age effects on chat AI and skills assessment were small and within noise.

Accommodation needs. Respondents who reported needing any form of accommodation rated accommodation availability 1.0 point lower across every modality. The accommodation gap is the most consistent disparity in the dataset and points directly to a deployment-side fix: a visible, low-friction, no-questions-asked path to a human alternative.

First-time vs. repeat AI candidates. First-time AI candidates rated overall experience 0.4 points lower than repeat candidates with comparable outcomes. First-encounter context appears to amplify negative reactions, and advance disclosure substantially closes the gap.

The segments where differences were small enough to be within noise: gender, race/ethnicity at the segment sizes we report, geography within the U.S./UK frame, and household income.

Accessibility, Disability, and Neurodivergence

The accommodation-availability gap is the most consistent disparity anywhere in this dataset, and it shows up in a way that is more diagnostic than the topline numbers suggest. The 1.0-point ratings drop for respondents who reported needing any form of accommodation holds across every modality, every age band, every role type, and every employer-size segment we could segment cleanly. No other demographic cut produces a gap that consistent. That tells us the issue is structural to how AI hiring is currently deployed rather than specific to any modality or vendor.

The specific accommodation patterns that came up most often in the open-ended responses, ranked by frequency:

No visible path to a human. The single most-cited frustration in the accommodation sub-cohort was that there was no clearly visible way to request a human alternative without feeling like the request itself would penalize the application. Candidates described searching the apply flow for an accommodation link, finding only a generic ADA email address, sending a request, and either receiving no response or being asked to justify the request before any alternative was offered. Where the human-alternative option was offered upfront and described as a free choice (not gated on documentation), accommodation-needs respondents rated the experience comparably to the broader sample.

Time pressure on synchronous formats. Candidates self-identifying as neurodivergent — most often ADHD, autism spectrum, or learning disability — described synchronous voice AI and chat AI screens as harder when the format imposed a hard time limit per response or made interruption easy. The same candidates rated voice AI more favorably when they could pause to think, ask the AI to repeat a question, or take a few seconds before responding without being cut off. Voice AI platforms that handled clarifying questions and processing pauses gracefully rated above-average even with this sub-cohort. Voice AI platforms that did not, rated well below.

Async video as the lowest-rated modality for accommodation needs. Async video drew the largest accommodation-related ratings drop of any modality (1.4 points lower than the modality average). The complaints concentrated on the visual-performance dimension — being recorded, the time pressure of pre-set response windows, the absence of conversational repair when something went wrong — rather than on the AI scoring itself. Several respondents described re-recording attempts costing them more time than a live conversation would have.

Technical-accessibility breakdowns. Live captions missing or out of sync on async video, screen-reader incompatibility on chat AI flows, keyboard-only navigation breaking inside embedded interview widgets, and color-contrast issues on skills assessment platforms each came up in the dataset. None of these is a deep AI problem. They are platform-engineering choices that vendors do or do not invest in.

For the legal and compliance frame, the relevant regimes — the Americans with Disabilities Act in the U.S., the Equality Act 2010 in the UK, and the EU AI Act's high-risk-system requirements for employment use — converge on the same operational requirement: a candidate must have a meaningful, non-penalizing path to an alternative process when the AI touchpoint is not appropriate for them. The data is consistent that the platforms and deployments where that path is visible, low-friction, and not gated on disclosure outperform on every accommodation-related metric we measured.

Where Buyers Should Focus to Fix the Candidate Side

Four operational levers, in rough order of effect size:

1. Choose a quality vendor inside whatever modality fits the role. The largest source of candidate-experience variance in the dataset is between vendors within a modality, not between modalities. Treat candidate-experience criteria — voice naturalness, edge-case handling, rubric-based scoring with disclosable criteria, fallback paths when something goes wrong — as first-class evaluation criteria in vendor selection, not as nice-to-haves layered on after capability and integration. The 1.8-point spread we observed across voice AI vendors in editorial testing is wider than any cross-modality comparison the survey can make.

2. Default to disclosure and a real human alternative. The disclosure-plus-alternative cell outperforms every other cell in the dataset by a meaningful margin, holds across every modality, and costs essentially nothing to deploy. Most platforms already support both. The blocker is policy, not product.

3. Close the silence loop. The single most common complaint in the open-ended data is 'I never heard back.' Candidates rejected at an AI stage with a prompt, clear notification rate the experience 0.7 points higher than those rejected with no follow-up — and are over 3x more likely to say they would re-apply. AI-stage rejections handled well beat the silent-resume baseline by a wide margin. Handled badly, much of that advantage evaporates. Automated rejection notifications are a basic platform feature. Turning them on for AI-stage rejections is a configuration choice.

4. Re-evaluate any modality (or vendor) that consistently underperforms. Async video at the category-average level rates lowest on perceived fairness and informational clarity, but the spread inside the category is wide — tighter deployments of higher-quality async video platforms perform comparably to other modalities. The data argues for either upgrading the vendor (within or across modalities) or tightening deployment, not for abandoning the modality wholesale.

How This Survey Was Conducted

The survey was fielded online January 14 through February 22, 2026. Eligible respondents for the primary cohort were adult U.S. or UK job seekers who applied to a role in the prior 90 days that included at least one AI touchpoint of an identified type (voice AI screen, chat AI screen, async video interview, or skills assessment). Verification questions confirmed the AI modality before substantive questions began.

Final usable primary sample: 2,587 respondents, stratified across modality (target n>=500 per modality, achieved n=512–701), role type (frontline/hourly, professional/knowledge, technical, executive), outcome (hired, progressed past AI stage, rejected at AI stage, withdrew), and demographic segments sufficient to support segment-level analysis without small-n disclosure risk.

Baseline cohort: 614 respondents who applied in the same 90-day window through a process that included no AI touchpoint of any kind, used as a comparison set on outcome, communication, and overall-experience measures. The baseline cohort was matched to the primary sample on role type, headcount band, and industry category.

Recruitment used independent survey panels with verified-applicant screening, supplemented by candidate community partnerships. No employer-administered channels were used — a deliberate methodological choice to capture the full applicant population, not the population of finalists. Process-experience questions were asked before outcome to structurally separate experience from result. The 60 follow-up interviews were conducted with a stratified random sub-sample.

Vendor-quality observations cited in the report draw on Recruiting Tech Reviews' editorial testing program, which conducted hands-on review of 60 AI recruiting platforms in the 2026 review cycle, including 14 voice AI and 11 async video platforms with candidate-side testing. The full survey instrument and editorial testing methodology are available on request to credentialed researchers and journalists.

Limitations and How to Use the Findings

All process-experience data is self-reported. Most candidates can't reliably identify which vendor's platform delivered a given AI touchpoint (employer branding typically obscures the underlying vendor), so vendor-by-vendor survey findings are limited to the subset of follow-up interviews where the platform was identifiable. The broader vendor-quality observations in this report are anchored on Recruiting Tech Reviews' editorial testing rather than on representative candidate-side vendor data.

The no-AI baseline cohort provides a comparison set against the realistic alternative most applicants face, but it is not a controlled experiment — process design, employer mix, and role characteristics differ between cohorts even after matching, and the comparison should be read as descriptive rather than causal.

Survey panels reach a substantial cross-section of recent applicants but under-represent applicants without sustained internet access, applicants for whom English is not the dominant language (this edition is U.S./UK English-only), and applicants who fully exited the labor market after the experience. Findings on perceived fairness by demographic segment are candidate perceptions, not adverse-impact analyses of underlying systems, which require employer-side outcome data the candidate-side instrument cannot provide.

For buyers: use the modality-level fairness, drop-off, vendor-variance, and disclosure-effect findings as a calibration layer for vendor and program decisions. The two most actionable findings are (1) treating candidate-experience criteria as first-class in vendor selection — given how much variance there is between vendors inside a single category — and (2) the disclosure-plus-alternative deployment policy, which has measurable candidate-experience impact, is available within every major platform, and most employers have not turned on.