It starts innocently. A manager has six performance reviews due Friday, opens a free AI chatbot, and pastes in notes about an underperforming employee. Name, role, specific incidents, maybe a mention of a medical accommodation or a complaint the employee filed. Thirty seconds later, a polished narrative comes back.
That manager just created legal, regulatory, and reputational exposure the organization will carry for years. Here’s why performance review narratives and similar HR content should stay out of consumer generative AI tools, full stop.
01 You just disclosed confidential employee data to a third party
Performance reviews are among the most sensitive records an organization holds. They contain names, compensation context, health accommodations, disciplinary history, and candid assessments of people’s weaknesses.
When that content goes into a consumer AI tool, it leaves your control. Depending on the tool’s terms, it may be:
- Stored indefinitely on the vendor’s servers
- Reviewed by human trainers for quality purposes
- Used to train future versions of the model
- Retained in logs even after you delete the chat
Most consumer-tier tools reserve some or all of these rights. Your employee never consented to any of it, and your privacy notice almost certainly doesn’t cover it.
02 It is discoverable. Yes, it can be subpoenaed.
This is the risk most managers never consider. Litigation happens in HR constantly: wrongful termination, discrimination, retaliation, wage disputes. When it does, discovery follows.
Chat logs with AI tools are records. If they exist, they can be subpoenaed, whether they sit on your systems or the vendor’s. Courts have already ordered AI providers to preserve user conversations in active litigation, including chats users believed were deleted. Assume anything typed into a chatbot can end up as Exhibit A.
THE DEPOSITION TEST
Picture opposing counsel holding a transcript showing the manager asked an AI to “make this sound less harsh” or “help me justify a low rating.” The prompt itself becomes evidence of intent, bias, or pretext. And unlike your internal documents, you may not even control the retention schedule. The vendor does.
03 You may have just destroyed privilege
If HR is working with counsel on a termination or an investigation, those communications may be protected by attorney-client privilege. Privilege depends on confidentiality. Routing that content through a third-party AI tool, one whose terms allow the vendor to access and use the data, hands opposing counsel an argument that privilege was waived. That’s a catastrophic own goal in litigation.
04 Bias in, bias out, liability everywhere
AI models carry the biases of their training data. Research has repeatedly shown language models describing men and women, or different demographic groups, in systematically different terms. If AI-drafted reviews show a pattern, say, women consistently described as “supportive” while men are “decisive,” you’ve manufactured statistical evidence for a class action.
Regulators are watching this exact issue. The EEOC has made clear that employers own the outcomes of the tools they use. New York City’s Local Law 144 requires bias audits for automated employment decision tools, and several states have followed with their own AI employment laws. “The chatbot wrote it” is not a defense; it’s an admission.
05 The narrative might just be wrong
Generative AI fabricates. It fills gaps with plausible-sounding specifics. An AI-polished review can attribute incidents that never happened, inflate or soften facts, and smooth over the documented specifics you actually need if the review is later challenged. A performance review is a legal document defending an employment decision. Fiction, even flattering fiction, undermines it.
06 Breach and regulatory exposure
Every third party holding your data is attack surface. AI vendors have already suffered incidents exposing user chat histories. If employee personal data leaks through a tool your organization never sanctioned, you may still own the breach notification obligations under state privacy laws, and under GLBA if you’re a financial institution.
Examiners will ask why sensitive personnel data was flowing to an unvetted vendor with no contract, no due diligence, and no data processing agreement. There’s no good answer to that question.
07 Shadow AI is probably already happening
Surveys consistently show a large share of employees using AI tools at work without approval, and most don’t tell anyone. If your organization hasn’t given managers a sanctioned, governed way to work, they’re improvising with free tools on personal accounts. The risk isn’t hypothetical. It’s happening on someone’s phone right now.
EXAMINER LENS
FFIEC guidance treats any third party handling institution or customer data as a vendor relationship requiring due diligence. An unsanctioned consumer chatbot fails that standard on day one: no contract, no SOC report, no data flow mapping, no exit plan.
What to do instead
- 1Write the policy
Explicitly prohibit entering employee personal data, performance content, investigation materials, and legal matters into unapproved AI tools. Name examples.
- 2Offer a sanctioned path
Enterprise-grade AI deployments with zero data retention terms, no training on your data, and a signed DPA are a different risk conversation than a free consumer chatbot. Vet the tool like any other critical vendor.
- 3Train managers on the “front page” test
If you wouldn’t want the prompt read aloud in a deposition, don’t type it.
- 4Keep humans accountable
AI can help with structure and grammar on de-identified drafts. Judgment, specifics, and the final word stay with the manager.
- 5Audit
Monitor for unsanctioned AI use the same way you monitor for any other data leakage channel.
The Bottom Line
A performance review is not busywork. It’s a confidential record, a potential piece of evidence, and a reflection of your organization’s judgment. Pasting it into a consumer chatbot trades a Friday deadline for years of discoverable, unretractable risk.
Write the hard sentence yourself. It’s cheaper than the subpoena.
This advisory is provided for general informational purposes and does not constitute legal advice. Institutions should consult counsel regarding privilege, discovery, and employment law questions specific to their circumstances. © 2026 FinCyberTech. All rights reserved.
