Algorithmic bias in assessment is a systematic difference in scores or selection outcomes between demographic groups that is not explained by real differences in the skill being measured. In hiring, the working benchmark is the four-fifths rule: if one group's selection rate falls below 80% of the highest group's rate, the tool is presumed to have adverse impact. In 2026, two large bodies of evidence — a Stanford-led audit of four million applications and more than 150 independent bias audits of commercial AI tools — finally let us replace opinion with data on how biased AI assessment actually is. The answer is uncomfortable for both camps: less biased than humans on average, and still capable of failing badly in specific jobs, for specific vendors, in ways aggregate statistics hide.
What the 2026 evidence actually shows
In May 2026, researchers from Stanford, Chapman and Northeastern published the largest audit of a hiring algorithm to date: over 4 million applications from 3 million candidates across 156 employers, all screened by the same commercial game-based assessment platform. The headline findings:
- 10.6% of individual positions showed adverse impact against Black applicants under federal thresholds.
- Roughly a quarter of all applications from Black candidates went to positions producing outcomes that would count as discriminatory under US employment law.
- Around 30% of Black applicants hit at least one position with a discriminatory result.
The most important detail is methodological. The vendor had previously reported no bias — accurately, at the aggregate level. Pool every position and every employer together and the disparities average out. Examine job by job, which is what employment law actually requires, and one in ten positions fails. A fairness claim computed at the wrong level of analysis is not a fairness claim at all.
The study also documented an "algorithmic blackball" effect: because candidate scores were reused across employers for up to 330 days, a candidate who scored poorly once could be silently rejected everywhere the platform operated. About 4% of applicants who applied to ten positions were rejected by all ten — more than chance predicts. One bad test day became a portable, invisible disqualification.
The counter-evidence: AI is still beating the human baseline
Before concluding that algorithmic assessment should be abandoned, look at the comparison group. Warden AI's analysis of 150+ bias audits covering over a million test samples found that 85% of audited AI systems met the four-fifths threshold, and that audited AI tools achieved an average impact ratio of 0.94, against roughly 0.67 for human-only processes drawn from peer-reviewed benchmarks. By that measure, audited AI delivered substantially fairer outcomes for racial minorities and for women than unstructured human judgment. EEOC filings tell the same story from another angle: the overwhelming majority of US discrimination claims over the past five years concern human decisions, not algorithmic ones.
Both findings are true at once. Human screening is the higher-bias baseline — decades of research on identical CVs with different names established that long before AI arrived. And specific AI deployments can still produce legally actionable disparities, because variance between vendors exceeded 40% in Warden's bias scores. "AI" is not one thing. The fairness of your assessment stack depends on which tool you bought, what data it learned from, and whether anyone has audited it at the level that matters.
The regulatory clock: paused, not stopped
The EU AI Act classifies AI used for recruitment, candidate evaluation and employee assessment as high-risk, with obligations covering human oversight, transparency to candidates, logging and bias monitoring. The compliance deadline was 2 August 2026 — this week, as originally enacted. In practice, the EU's Digital Omnibus package, adopted this summer, deferred the high-risk obligations to 2 December 2027, citing delays in standards and enforcement infrastructure.
Do not misread the deferral. The prohibitions have been in force since February 2025: emotion recognition in interviews, biometric categorisation inferring protected characteristics, and social scoring are already banned, with penalties reaching €35 million or 7% of global turnover. The high-risk classification of hiring and assessment tools is settled law; only the date moved. Any Indian assessor, GCC or BPO screening candidates located in the EU is in scope regardless of where the company sits. Sixteen extra months is time to build an audit habit, not permission to skip one.
How to audit your own assessment stack
The practical lesson from 2026's evidence is that fairness must be measured where decisions happen, not asserted in a vendor brochure. A workable audit routine:
1. Compute adverse impact per job, not per platform
Run the four-fifths calculation for each role (or tightly related job family) with enough volume to be meaningful. Aggregated dashboards are how the Stanford study's disparities stayed invisible for years.
2. Interrogate score reuse
If your platform carries scores across requisitions or clients, ask for how long, and whether a candidate can retest. Score portability is efficient right up until it becomes a blackball.
3. Demand vendor evidence at the right level
Ask vendors for job-level or client-level impact ratios, the demographic composition of training data, and third-party audit results — not a single aggregate fairness figure. A 40% spread between vendors means selection is your biggest fairness lever.
4. Keep humans in the loop where stakes are high
Both the EU AI Act and GDPR Article 22 point the same way: consequential rejections need meaningful human review. Structured AI scoring plus human oversight beats either alone. This is the design logic behind AssessAll's AI-graded scenario responses — the AI applies a consistent rubric to every candidate's open-ended answer, while recruiters see the evidence and reasoning behind each grade rather than an unexplainable number, and proctoring outcomes are reported as High/Medium/Low integrity bands instead of a black-box verdict.
5. Re-audit on a cadence
Bias is not a one-time certification; it drifts as applicant pools, roles and models change. Quarterly per-role impact checks are a reasonable floor for volume hiring. Usage-priced platforms make this economically sane — on a pay-as-you-go model like AssessAll's (₹30/US$0.50 per assessment credit), running a monitoring sample costs pocket change compared with defending a disparate-impact claim.
What this means for Indian hiring teams
India's volume-hiring context — campus drives, BPO and BFSI funnels processing thousands of candidates per requisition — is exactly where both the promise and the risk concentrate. High volumes make human screening least reliable and most biased, so structured, AI-assisted assessment has the most fairness upside here. But high volumes also mean a biased tool damages more candidates faster, and job-level disparities reach statistical significance quickly. Teams serving European clients should treat December 2027 as a hard date; everyone else should treat the Stanford findings as the reason not to wait for a law.
The takeaway: the 2026 evidence says AI assessment is already fairer than the human baseline on average — but only audited AI, measured job by job, deserves that claim. Pick vendors who show their numbers at the level where hiring decisions actually happen, and check your own.