The ChatGPT-Written Performance Review as Exhibit A: Pretext Risk and Calibration Gates

The Capital Math of a 20-Minute Shortcut

Saving twenty minutes per appraisal is a terrible trade when one contradicted termination costs six figures to defend. Unaudited AI review prose turns routine Q1 separations into jury-ready pretext claims.

Run the executive arithmetic on a 400-person company entering November review season. Suppose 50 managers each owe eight appraisals. Drafting an honest, fact-anchored review takes roughly 45 minutes per direct report. Pasting three bullet points into a chatbot cuts that chore to 15 minutes. Across 400 reviews, management saves 200 hours of writing time. At an average manager loaded cost of $85 per hour, the organization saves $17,000 in Q4 administrative effort.

Now price the January collision. Sixty days after ratings lock, a regional sales director comes to HR demanding to fire a 56-year-old account executive—or an analyst recently back from intermittent medical leave—for "chronic missed deadlines and poor executive presence." HR pulls the November appraisal signed eight weeks earlier. Three paragraphs of symmetrical, upbeat chatbot prose praise the employee's "collaborative spirit, adaptable communication style, and consistent commitment to team deliverables." Not one missed deadline appears on the page.

That single document just wiped out ten times your Q4 time savings. Under McDonnell Douglas Corp. v. Green, 411 U.S. 792 (1973), and Reeves v. Sanderson Plumbing Products, Inc., 530 U.S. 133, 147 (2000), a trier of fact may infer unlawful discrimination directly from the falsity or inconsistency of the employer's stated explanation. When a November appraisal praises competencies that management calls terminal defects in January—with zero intervening documented failure—summary judgment evaporates. Defending a single Title VII, ADEA, or ADA claim through federal discovery and summary-judgment briefing routinely consumes $125,000 to $220,000 in outside counsel fees before a dollar of settlement changes hands.

Two Ways AI Appraisals Break Your Defense in Discovery

Generative models create litigation exposure at both ends of the rating distribution. Inflated "soft-no" prose manufactures positive contradictions, while harsh AI-rewritten critiques leave behind discoverable prompt histories.

In calibration audits across AI and changing work, HR business partners encounter two distinct failure patterns. Neither looks dangerous until a separation dispute lands on outside counsel's desk.

Failure Mode 1: Conflict-avoidant inflation ("The polite hallucination"). Most line managers dislike delivering blunt criticism before winter holidays. When a manager feeds a consumer or enterprise assistant a vague prompt such as "write a professional review for Jordan; hit 80% of quota, sometimes disorganized, good teammate, keep it constructive," standard instruction-tuned models default to diplomatic corporate flattery. The output reads like an endorsement. Worse, when managers copy-paste someone else's role template into the prompt, the model often hallucinates responsibilities the employee never held. If an appraisal credits a warehouse lead for "cross-functional client stakeholder alignment," plaintiff's counsel will put the manager on the stand and prove nobody read the evaluation before signing it.

Failure Mode 2: Toxic prompt exhaust. The reverse scenario is even worse. A frustrated supervisor wants HR to approve a performance improvement plan or termination. Before opening the HRIS form, the supervisor opens a personal chatbot account and types: "Write a firm Needs Improvement review for an employee whose migraines and doctor appointments keep slowing down Monday deliverables, without mentioning health." The finished review in Workday looks sanitized. The prompt sitting in the manager's cloud chat history is direct retaliatory evidence under 42 U.S.C. § 12112 and 29 U.S.C. § 2615(a). Pair that risk with our post-leave PIP retaliation protocol before approving any adverse rating near protected leave.

Do not assume chat logs stay private. Under Federal Rules of Civil Procedure 26(b)(1) and 34(a)(1)(A), electronically stored information relevant to a party's claims or defenses—including cloud-hosted drafts and prompt histories under a custodian's practical control—is squarely discoverable. And in United States v. Heppner, No. 25-cr-0503 (S.D.N.Y. Feb. 17, 2026), U.S. District Judge Jed S. Rakoff confirmed that prompts entered into a public generative AI platform enjoy neither attorney-client privilege nor work-product protection.

State Recordkeeping Traps and Consumer Chatbot PII Leaks

Pasting raw peer feedback, compensation figures, or disability notes into a free chatbot creates dual regulatory exposure. California further mandates four-year preservation whenever automated tools facilitate evaluations.

Audit what managers actually paste into browser tabs during review week. A supervisor trying to summarize 360-degree peer comments routinely copies full employee names, base salaries, bonus targets, equity grants, customer account lists, and accommodation schedules into an unvetted web interface. On consumer tiers where model-training defaults remain active, confidential personnel data enters third-party vendor pipelines outside your company's data-processing agreements.

California employers face an explicit statutory retention clock on top of privacy exposure. Under the California Civil Rights Council's Automated-Decision System regulations (2 CCR § 11008.1 and § 11013(c), effective October 1, 2025), an "automated-decision system" covers any computational process using artificial intelligence or machine learning that makes or facilitates human decision-making regarding an employment benefit, explicitly including performance evaluations, compensation, promotion, and termination. Covered California employers (five or more employees) must preserve all personnel records and ADS inputs, criteria, and outputs for four years from creation or from the date of the personnel action, whichever is later. Letting managers generate evaluation text in ephemeral personal accounts you cannot archive puts your organization out of compliance on Day 1.

The Seven-Day Pre-Lock Calibration Protocol

Never open employee visibility or compensation letters until HR runs a structured seven-day calibration gate. Focus human audit time on the exact cohorts where litigation risk concentrates.

You do not need to hand-edit 400 reviews line by line. High-leverage people operations teams run a targeted exception filter seven calendar days before ratings lock and manager delivery:

1. Day T-7 — Pull the Contradiction Cohort: Cross-match every draft rating of "Meets" or "Exceeds" against (a) active informal coaching logs, (b) managers who verbally flagged headcount swaps or Q1 PIPs during budget planning, and (c) employees with bottom-decile objective output metrics. Typically this isolates 6% to 10% of the workforce (24 to 40 files in a 400-person firm).
2. Day T-5 — Run the Three-Artifact Corroboration Rule: Require every "Needs Improvement" or "Unsatisfactory" draft—and every contested borderline file—to cite at least three dated, verifiable work records (missed sprint milestone on August 14, client escalation ticket #4812 on September 9, Q3 quota attainment at 61% against a peer median of 94%). If an LLM summarized those facts inside an approved enterprise workspace, the underlying source records must exist independently in the system of record.
3. Day T-3 — Normalize Protected Leave Denominators: Verify that no employee who took FMLA, PWFA, ADA, or military leave during the review year was penalized against an unadjusted 12-month volume quota. Prorate targets mathematically by active workweeks before locking scores.
4. Day T-1 — Collect Manager Calibration Attestations: Require every people manager to check a mandatory HRIS attestation before employee ratings become visible.

Copy-Ready Manager AI Guardrail & Calibration Attestation

Embed these exact guardrails at the top of your Q4/Q1 manager review instructions and require the four-line attestation checkbox before HRIS submission. Clear bright-line rules beat vague warnings about "responsible AI."

PERFORMANCE REVIEW DRAFTING RULES & PRE-LOCK MANAGER ATTESTATION

1. Approved Tool Boundary (Zero Consumer Chatbots)
   - NEVER paste employee names, self-reviews, peer feedback, compensation figures, medical/leave notes, or disciplinary history into personal or public AI accounts (consumer ChatGPT, Claude, Gemini, or browser extensions).
   - If you use [Approved Enterprise AI Tool] to tighten grammar or structure, restrict inputs to anonymized work bullets from your own notes.

2. Fact-First Drafting Rule (No Unverified Boilerplate)
   - AI tools cannot evaluate employee performance. Every rating—especially "Needs Improvement" or "Exceeds"—must cite specific, dated work examples, metrics, or project outcomes from the review period.
   - Do not submit diplomatic praise for an employee who is missing core expectations. An inflated review blocks HR from approving a PIP or performance separation in Q1/Q2.

3. Protected Leave & Accommodation Neutrality
   - Never reference approved medical leave, FMLA, pregnancy/postpartum accommodations, disability adjustments, or protected complaints in any prompt, draft note, or final review. Prorate annual volume goals for approved leave periods with your HRBP.

MANDATORY HRIS SUBMISSION ATTESTATION (Check Before Locking):
[ ] I personally verified every factual statement, project example, and metric in this evaluation against primary work records.
[ ] I did not enter confidential employee data or protected medical/leave information into any unapproved external AI tool.
[ ] This rating candidly reflects actual performance during [Review Period], and I have disclosed any active performance concerns to HR prior to locking.

Manager Signature: _______________________   Date: _______________

Role Alignment Across HR, Line Managers, and Legal

Pretext defense breaks down when executives demand fast January exits on uncalibrated November paperwork. Assign explicit pre-lock authority across all three seats.

Build this gate directly into your manager support training calendar every October. Make the internal rule unmistakable: HR will reject any Q1 involuntary performance termination request that contradicts a clean appraisal signed within the prior 90 days unless a documented, severe post-review failure intervened.