AI Still Can’t Reliably Evaluate Employee Feedback, Study Finds

What the Research Actually Found

PYX Labs, a research lab sponsored by Perceptyx, benchmarked seven AI models from OpenAI, Google, Anthropic, and xAI across 84 employee listening tasks. The evaluation criteria were developed by psychologists and organizational behavior specialists—not engineers—which makes the methodology worth paying attention to.

The results split cleanly along one line: structured versus unstructured work.

When tasks had clear, verifiable answers, AI models passed between 76% and 82% of them. Solid performance for routine categorization and pattern recognition.

When tasks required interpreting open-ended employee feedback and producing a coherent, accurate takeaway, that pass rate dropped to as low as 33%.

That’s not a minor gap. That’s a different category of capability entirely.

Where AI Breaks Down: Synthesis

The lowest-scoring capability across every model tested was synthesis—the ability to pull together multiple, sometimes conflicting signals into one clear conclusion.

Synthesis scores ranged from 14% to 57%, the widest performance gap of any task in the study. The report described the failure point precisely: models struggled when they had to weigh incomplete, emotional, or context-dependent signals and resolve them into one clear takeaway.

That description maps almost perfectly to what employee feedback actually looks like in practice. People don’t write in clean categories. They write about frustration, confusion, competing loyalties, and things they can’t quite name. That’s exactly where AI falls apart.

The study also flagged something more serious: rare but meaningful instances where models fabricated statistical outputs or failed to stay within the constraints of the underlying dataset. In a performance management context, that’s not a minor bug. It’s a liability.

The Adoption Problem

Here’s the tension that makes this research matter beyond the lab.

Only 20% of companies report that their managers are effective at giving feedback and coaching, according to a 2025 WTW report. That’s a real, documented gap in human capability. So companies have been turning to AI to fill it—37% of respondents to the same WTW survey said they already use AI tools as part of their performance management process.

AI is being deployed to solve a human problem, before the AI is actually reliable enough to solve it.

Joseph Freed, chief product officer at Perceptyx and head of PYX Labs, framed the core issue clearly: the question isn’t whether AI can produce fluent-sounding answers. It’s whether AI understands what good looks like in a workplace context.

Right now, the evidence says it often doesn’t.

What This Means for HR Teams Using AI Tools

This research doesn’t argue that AI has no place in HR workflows. It argues for a more honest accounting of where AI adds value and where it introduces risk.

A practical way to think about it:

• AI is useful for tagging themes, flagging response volume trends, sorting structured survey data, and surfacing patterns across large datasets.
• AI is unreliable for synthesizing open-ended responses, interpreting emotional or ambiguous signals, and generating recommendations that directly influence decisions about individual employees.

The risk isn’t that AI gives a wrong answer. The risk is that it gives a confident, fluent, wrong answer—and no one catches it before it shapes a performance review, a promotion decision, or a layoff.

The Takeaway

If your HR team is using AI to process employee feedback, the PYX Labs findings are a useful calibration tool. The models available today—even the most capable ones—are not reliable interpreters of complex human experience at work.

That doesn’t mean stop using AI in HR. It means build human review into any workflow where AI is synthesizing open-ended feedback or generating people-related recommendations. The 33% pass rate on interpretation tasks isn’t a reason to panic. It’s a reason to design your process accordingly.

Observe what AI can actually do. Then choose how much to trust it.