AI Pair-Programming Screens Without Creating Audit Risk

Recommendation: treat AI pair-programming as a controlled access session with an evidence pack, or do not use it for hiring decisions. In a typical incident, a team escalates after a bad start. The candidate looked collaborative during a pair session, but no one can reconstruct the session: which parts were authored, what prompts were used, and whether a proxy participated. Your People Analytics team is asked for impact analysis and cannot produce an audit-ready chain of custody. Fraud risk is not theoretical. Checkr reports 31% of hiring managers say they have interviewed a candidate who later turned out to be using a false identity. If your pair-programming mode does not start with identity gating before access, you are increasing the probability that collaboration "signals" were generated by the wrong person. Cost lands in two places: cycle-time waste and replacement risk. SHRM cites replacement cost estimates ranging from 50-200% of annual salary. Pair sessions that are not instrumented also create SLA breakdowns: disputes take days, follow-ups get scheduled late, and time-to-offer drifts at the exact moment where identity is unverified and evidence is missing.

Recommendation: stop using disconnected interview tooling as your collaboration system of truth. Require unified, ATS-anchored audit trails. Legacy ATS workflows capture stage changes and notes. They do not capture the collaboration artifact chain: prompt history, code playback, execution telemetry, and reviewer rubric state at decision time. Coding challenge tools optimize for scoring output, not for reconstructing who did what and when. The market also defaults to sequential checks. Identity is often verified after interviews, which is backwards. Time delays cluster at moments where identity is unverified, and that is where proxy interviews and deepfake attempts are most valuable to an attacker. Without immutable event logs, you cannot build unified evidence packs. Without review-bound SLAs, escalation becomes tribal knowledge. Without standardized rubric storage, you cannot defend consistency across candidates. Shadow workflows then appear to "fix" the gaps, and those shadow workflows are integrity liabilities.

Recommendation: assign owners per control, not per tool. Then bind each control to a timestamp and evidence artifact. Recruiting Ops: owns stage orchestration, candidate routing, and SLA-bound queues. They ensure the identity gate is completed before the pair session link is issued. Security: owns fraud policy, access control, step-up verification triggers, and audit policy for logs and evidence packs. They approve acceptable use boundaries for AI assistance and proxy detection controls. Hiring Manager: owns evidence-based scoring, rubric discipline, and calibration. They must justify decisions with references to logged artifacts, not memory. People Analytics: owns measurement and governance: segmented risk dashboards, reviewer consistency reporting, and audit retrieval drills. They define the minimum viable evidence pack schema required for analysis and defensibility. Systems of truth: ATS for lifecycle state and dispositions. Assessment and verification layers for immutable event logs, code playback, and telemetry. All write back pointers and timestamps into the ATS to keep a single source of truth.

Recommendation: run AI pair-programming in an async-first, risk-tiered funnel with step-up verification and tamper-resistant feedback.

IntegrityLens AI enables AI pair-programming screens as an instrumented workflow so People Analytics can measure, audit, and defend decisions without slowing the funnel: - Identity gating before access using biometric verification (liveness, face match, document authentication) with typical end-to-end verification in 2-3 minutes, time-stamped to the candidate record. - AI coding assessments across 40+ languages with plagiarism detection and execution telemetry, producing artifacts you can use for evidence-based scoring and code playback dispute resolution. - Fraud prevention controls that look for deepfakes, proxy interviews, behavioral signals, and continuous re-authentication triggers during high-risk sessions. - Immutable evidence packs: timestamped logs, rubric snapshots, reviewer notes, and ATS write-backs so the decision chain is reconstructable. - Zero-retention biometrics architecture to support privacy-by-design policies while keeping the audit trail intact.

Do not implement AI pair-programming with these patterns. They increase fraud exposure and make decisions non-defensible: - Allowing access to the pair session link before identity_verified is logged, then trying to " type subSections for block? (But must be exactly 3 bullets; ensure content includes 3 bullets)

External statistics referenced: - Checkr Hiring Hoax (Manager Survey, 2025): https://checkr.com/resources/articles/hiring-hoax-manager-survey-2025 - SHRM replacement cost estimates: https://www.shrm.org/in/topics-tools/news/blogs/why-ignoring-exit-data-is-costing-you-talent

Recommendation: ship a minimum viable, audit-ready pair mode first, then optimize. The goal is reduced time-to-hire with defensible decisions and lower fraud exposure. Implementation checklist: - Lock a driver-navigator rubric to req and level, and store rubric_version_id in every decision. - Enforce identity gate before access and block session links until identity_verified_at is logged. - Require code playback and execution telemetry for every session, not just pass cases. - Set review-bound SLAs: 1 business day for scoring, 4 hours to initiate step-up verification when risk signals fire. - Build a segmented risk dashboard: time-to-event, step-up rate, dispute rate, and reviewer variance by team. - Run a monthly audit retrieval drill: can you produce the evidence pack and approval chain within 30 minutes? Business outcomes to track in People Analytics: time-to-offer delta by role family, dispute rate, step-up verification rate, and post-start performance flags relative to collaboration scores.