What Meta’s AI Layoff Lawsuit Reveals About Performance Data

A lawsuit against Meta has reignited a familiar debate: Should artificial intelligence play a role in performance and workforce decisions that affect people’s careers?

It is an important question. But it may not be the most useful one.

The more revealing question is: What information was the AI asked to evaluate?

According to a lawsuit filed by 26 Meta employees, the company used internal AI systems, activity-monitoring data, AI usage dashboards, and algorithmically assisted performance rankings as part of a layoff-selection process. The employees allege that the process failed to account adequately for medical, parental, and family leave, effectively penalizing people whose absence reduced the activity the systems could measure.

Meta disputes those allegations and says organizational decisions were made by people, not AI. The case is still unfolding, and many of the details about how the technology influenced the process remain unclear. Associated Press reporting on the lawsuit provides an overview of the allegations and Meta’s response.

But the uncertainty surrounding the case is exactly why it matters.

Before concluding that AI made a good or bad decision, organizations need to understand the performance system surrounding it: the data the technology received, the outcome it was instructed to optimize, the context it could not see, and the people responsible for validating its output.

The Meta case is not simply a referendum on AI.

It is an early test of whether the performance systems organizations have today are ready for the decisions AI will increasingly help them make tomorrow.

Start with the inputs, not the algorithm

When an AI-supported talent decision becomes controversial, attention naturally turns to the technology.

What model was used? Did an algorithm make the decision? Was the output biased?

Those questions matter. But even the most sophisticated AI system cannot produce a reliable assessment from incomplete, poorly structured, or irrelevant information.

As Betterworks CEO Doug Dennerline put it during a recent leadership conversation:

“I need to know more information about how they were making those decisions—how AI was involved, how much data they looked at, and what their program design was.”

Those are the questions every organization should be prepared to answer before AI influences a consequential talent decision:

• What outcome was the system asked to optimize?

• Which performance signals were included?

• How were those signals weighted?

• Did the information represent the employee’s complete body of work?

• What important context was missing?

• Could a manager challenge the recommendation?

• Who remained accountable for the final decision?

These are not minor technical details. They determine what the system is actually evaluating.

An algorithm asked to identify the lowest performers will produce very different results depending on whether “performance” means business outcomes, goal progress, manager ratings, hours online, AI adoption, keystrokes, or some combination of those signals.

The question is not only whether AI was used.

It is how the organization defined performance in the first place.

Traditional reviews were not designed for AI-era decisions

Long before generative AI entered the workplace, performance reviews had a data problem.

Many organizations still evaluate performance through one or two formal review cycles each year. Managers reconstruct months of work from memory, employees summarize their own accomplishments, and the process concludes with a rating intended to represent an entire body of work.

The limitations are familiar:

• Recent events receive more attention than earlier contributions.

• Work performed across teams may be invisible to the direct manager.

• Changing priorities are not always reflected in the original goals.

• Feedback delivered throughout the year is scattered across systems.

• The final rating compresses a complicated performance story into a single number.

An employee can receive a strong review and then be selected for a layoff months later based on a different set of criteria. Even when there is a legitimate business reason for the decision, the disconnect creates confusion, distrust, and potential legal risk.

AI did not create that problem.

But AI can amplify it.

If a model is working from two historical ratings, incomplete documentation, and a manager’s recollection of recent events, it does not have a reliable view of performance. It has a faster and more scalable way to analyze the same fragmented record.

That is why AI-native performance management requires better data, not simply faster reviews.

A capable model pointed at thin, episodic performance records can still produce a confident but unreliable conclusion.

More employee data is not necessarily better performance data

The allegations against Meta also highlight an important distinction between employee activity and employee performance.

Modern workplaces generate enormous amounts of data:

• Messages sent

• Documents created

• Meetings attended

• Time spent in applications

• System logins

• Tasks completed

• AI tools used

• Tokens consumed

• Keystrokes or mouse activity

These signals are easy to count.

That does not mean they accurately represent contribution.

An employee can produce a large volume of visible activity without advancing an important company priority. Another employee might solve one critical customer problem, improve a high-value process, make a strategic decision, or enable an entire team while producing relatively few trackable interactions.

The employee’s output may be affected by a change in responsibilities, a disability accommodation, an under-resourced project, a new manager, a delayed dependency, or a strategic priority that shifted midyear. A dataset may record the visible result without explaining the conditions surrounding it.

This is one of the central risks of using workplace data for high-stakes decisions: What is easiest to measure is not always what matters most.

The goal should not be to give AI the greatest possible volume of employee data.

It should be to give AI relevant, representative, and appropriately governed evidence of performance.

AI needs a body of work, not a snapshot

The opportunity for AI in performance management is not simply to summarize an annual review or calculate a rating more efficiently.

It is to help organizations understand performance across time.

That requires a richer foundation of information, including:

• Goals connected to current business priorities

• Progress toward measurable outcomes

• Regular manager and employee check-ins

• Timely feedback from colleagues and stakeholders

• Coaching and recognition

• Changes in responsibilities

• Project and milestone results

• Relevant context from the systems where work happens

Together, these signals create an evolving body of evidence rather than a periodic snapshot.

AI that understands performance can surface patterns across this information: where an employee consistently delivered, where priorities changed, which obstacles affected progress, and which conclusions are supported by evidence.

This represents a fundamentally different use of AI than asking a model to rank employees from a handful of historical ratings.

It also changes what performance management needs to become.

Instead of waiting for employees and managers to manually reconstruct the past at review time, real-time performance management creates a continuous record of goals, feedback, conversations, and outcomes as work unfolds.

Continuous employee feedback is an important part of that record. Feedback captured near the work itself can preserve examples and perspectives that would otherwise be forgotten months later.

The objective is not to monitor every action an employee takes.

It is to create enough relevant context that neither AI nor a manager has to guess what happened.

Human validation has to be meaningful

Meta says its organizational decisions were made by people rather than AI. The plaintiffs allege that AI-driven systems and rankings materially shaped the process.

That disagreement points to another question organizations will increasingly need to address: What does it actually mean for a human to remain “in the loop”?

A person approving a recommendation does not necessarily make the process human-led.

Managers can be influenced by automation bias—the tendency to accept the output of a system because it appears objective or analytically sophisticated. Human review becomes little more than a formality when the reviewer cannot see the underlying evidence, understand the criteria, or meaningfully challenge the result.

Meaningful human validation requires that reviewers can:

• See which evidence informed the recommendation

• Understand the criteria the system applied

• Identify information the system may have missed

• Add relevant context

• Correct inaccurate data

• Challenge or override the output

• Document why the final decision was made

The NIST AI Risk Management Framework similarly emphasizes the need to define human roles, responsibilities, and oversight when organizations deploy AI systems.

AI should be a source of insight.

It should not become an unchallengeable source of authority.

Calibration becomes more important—not less

AI could eventually make some talent decisions more consistent than those produced by managers working from memory and subjective impressions alone.

That is a meaningful opportunity.

“If all that data was there, it would probably make them better at doing that than a human,” Dennerline said. “Because it would be impartial.”

The key phrase is if all that data was there.

AI cannot create impartial decisions from inconsistent definitions of performance. It cannot supply missing context on its own. It cannot know whether one team’s manager scores generously while another applies a much higher standard unless the system has the evidence and design required to identify that discrepancy.

This is where modern performance calibration remains essential.

Calibration gives leaders an opportunity to examine whether:

• The same standards are being applied across teams

• Ratings and recommendations are supported by evidence

• Important context has been considered

• Managers are interpreting performance consistently

• Certain employees or groups are being affected differently

Traditional calibration often falls short because leaders enter the conversation with static review summaries, spreadsheets, and manager recollections.

Modern performance review calibration can bring goals, feedback, performance history, and other relevant evidence directly into the decision workflow.

AI can help leaders assemble and analyze that information.

Calibration provides the forum for testing the conclusion.

What HR leaders should take from the Meta case

The allegations against Meta have not yet been proven, and many details about the company’s process remain contested or unknown.

HR leaders do not need to decide who is right to recognize the broader warning.

AI will increasingly influence decisions about promotions, performance ratings, succession, internal mobility, workforce planning, and restructuring. As that happens, the quality of the organization’s performance foundation will become more consequential.

Before using AI to inform a high-stakes talent decision, organizations should be able to answer several questions.

What are we defining as performance?

The organization should distinguish meaningful outcomes and contributions from convenient proxies such as activity volume.

Does the data represent performance across time?

A recent rating, isolated project, or short-term activity pattern may not represent an employee’s full contribution.

What context could the system be missing?

Leave, accommodations, changing priorities, role transitions, resource constraints, and team dependencies can all affect the signals AI receives.

Can the output be explained?

Managers and HR leaders should be able to understand which evidence contributed to a recommendation and how the conclusion was reached.

Can the output be challenged?

There should be a defined mechanism for correcting information, adding context, and overriding the recommendation.

Who owns the final decision?

Accountability must remain with clearly identified human decision-makers—not with a model, vendor, dashboard, or abstract “system.”

These questions should not be introduced after an employee challenges a decision.

They should shape the performance program from the beginning.

Responsible AI starts before the AI

The lesson from the Meta lawsuit is not that AI should be removed from performance management.

Used responsibly, AI has the potential to help organizations identify patterns, reduce reliance on manager memory, apply standards more consistently, and make better-informed talent decisions.

But the model itself is only one part of the system.

AI will amplify whatever performance foundation it is given.

A process built around isolated reviews, activity monitoring, incomplete context, and superficial human approval will produce faster versions of the same flawed decisions organizations already struggle with.

A process built around continuous performance evidence, clear criteria, responsible governance, and meaningful human judgment can produce something better.

That is why preparing for AI-powered talent decisions begins before an organization selects a model or launches a new feature.

It begins by connecting the goals, conversations, feedback, and outcomes that show how work is actually happening.

It begins by replacing recollection with evidence.

And it begins by recognizing that responsible AI requires better performance management underneath it.

Betterworks Performance Management creates a continuous view of goals, feedback, conversations, and performance evidence, helping organizations establish the context that both managers and AI need to make better decisions.