Could AI decide who gets laid off? What the Meta lawsuit means for UK employers

A serious allegation against Meta has put a very uncomfortable question in front of employers: could an AI system help decide who loses their job?

According to the discussion source, 26 Meta employees have sued the company, claiming it used artificial intelligence systems to select people for layoffs and that this disproportionately targeted workers on medical, parental or family leave. The employees are among 8,000 people Meta said it would lay off, described as about 10% of its workforce.

The claim is not that a chatbot casually fired people. The allegation is more specific and, frankly, more relevant to real businesses: internal AI systems, keystroke and activity-monitoring data, AI token-usage dashboards and algorithmically assisted performance rankings were allegedly among the methods used to determine who would be laid off.

Meta has not had its liability established in the material provided. These are legal allegations, and the company response is not disclosed in the source. But as a case study in AI governance, it is exactly the sort of thing UK organisations should be paying attention to.

Why AI layoff decisions are so risky

Redundancy selection is already sensitive. Add AI, productivity scoring and employee monitoring data, and the risk profile changes quickly.

The core problem is that many workplace metrics are not neutral just because they are numerical. Keystrokes, activity levels, dashboard usage and output volume may look objective, but they often measure availability as much as performance. Someone on medical leave, parental leave or family leave may naturally have fewer recent signals in the system.

The lawsuit reportedly argues that some scores and ratings, by design, could not be accumulated by an employee who was on protected medical or family leave, or whose output was reduced by a disability. That is the crucial point. If a system rewards visible activity without properly accounting for authorised absence or reasonable adjustments, it can turn a protected circumstance into a negative performance signal.

That is not a mysterious AI failure. It is a design failure, a governance failure and possibly a human review failure.

The UK lesson: do not confuse productivity data with fairness

For UK employers, the practical warning is simple: productivity data is not the same thing as fair evidence.

Many businesses now have more digital exhaust than they know what to do with. Collaboration tools show message counts. Code platforms show commits. CRM systems show calls and tasks. AI tools may show token usage, prompt volume or automation activity. Monitoring tools may track login patterns, mouse movement, keystrokes or application use.

These signals can be useful in limited contexts. They can help identify bottlenecks, support overloaded teams or spot underused systems. But they are dangerous when treated as a direct proxy for employee value.

A parent returning from leave, a disabled employee working with adjustments, or someone dealing with a health issue may produce a very different data trail from a colleague working full-time without interruption. That difference may be entirely legitimate. A model that ignores the context can make the wrong thing look efficient.

What an algorithm can miss in redundancy selection

AI systems are good at finding patterns in data. They are not automatically good at understanding whether those patterns are fair, lawful, relevant or humane.

In this case, the allegation is that Meta did not account for protected leave when taking employee scores into account and did not pause the system for an individualised, leave-neutral and accommodation-neutral review. Whether that is proven is not disclosed. But the governance principle is worth spelling out.

Any AI-assisted HR process should ask questions such as:

• Does the data penalise people for periods of approved absence?
• Are disability adjustments properly represented in the scoring logic?
• Can managers explain why a person was selected without hiding behind the system?
• Has the employer tested whether certain groups are disproportionately affected?
• Is there a meaningful human review, or just a rubber stamp?
• Can employees challenge the decision and see the basis for it?

If those questions cannot be answered clearly, the organisation is not ready to use AI in decisions as consequential as redundancy.

Human review must mean more than clicking approve

A common defence of workplace AI is that there is still a human in the loop. That phrase sounds reassuring, but it can be misleading.

If a manager receives a ranked list, assumes the model has already done the hard work, and approves it under time pressure, that is not meaningful oversight. It is administrative cover. Human review only matters if the reviewer has enough information, authority and time to challenge the system.

In redundancy decisions, that means looking beyond the score. Why is the person ranked where they are? What data is missing? Did leave affect the result? Were adjustments considered? Are there comparable employees with different circumstances? Has the selection pool itself been defined fairly?

The more consequential the decision, the stronger the review process needs to be. Losing a job is not the same as receiving a product recommendation.

Employee monitoring creates a second layer of risk

The discussion also mentions keystroke and activity-monitoring data. That matters because AI does not just introduce risk at the decision stage. It can also amplify risk at the data collection stage.

Monitoring employees can create privacy, trust and data protection issues, particularly if staff do not understand what is being collected or how it may later be used. A dataset gathered for productivity analysis can become far more sensitive if it is later used in performance management or redundancy selection.

For UK organisations, this is where data protection governance becomes more than paperwork. Employers should be clear about purpose, necessity, proportionality, retention and access. They should also think carefully before combining multiple data sources into a single performance score, because aggregation can make a system feel scientific while hiding weak assumptions.

There is also a cultural issue. A workplace that measures people by activity trails can quietly encourage performative busyness. That is bad management with or without AI.

How UK businesses should use AI in HR without sleepwalking into trouble

AI can still have a useful role in HR. It can help summarise policy documents, identify skills gaps, draft role profiles, support workforce planning and improve internal reporting. The issue is not AI in HR full stop. The issue is using AI to make or heavily influence high-impact decisions without proper controls.

A sensible UK employer should treat AI-assisted redundancy, performance scoring and monitoring as high-risk internal systems. That does not necessarily mean banning them. It means slowing down enough to design them properly.

1. Map the decision. Be clear where AI is used, what data it uses and who makes the final call.
2. Test for unfair impact. Look for patterns affecting people on leave, disabled employees, carers, part-time workers and other groups that may be disadvantaged by activity-based scoring.
3. Separate absence from performance. Make sure approved leave and reasonable adjustments do not become hidden penalties.
4. Keep evidence explainable. If you cannot explain the basis for a selection decision in plain English, the process is probably too opaque.
5. Give humans real authority. Reviewers must be able to override the system and document why.
6. Document the governance. Keep records of testing, assumptions, limitations, sign-offs and challenge routes.
7. Consult before deployment. Where monitoring or automated decision support affects staff, involve HR, legal, data protection leads and employee representatives where appropriate.

This is also where business leaders should resist vendor theatre. A dashboard can look polished and still encode bad assumptions. For a wider example of how AI systems can reflect odd defaults and hidden bias, see my piece on AI, bias, randomness and cultural defaults.

The real question is accountability

The Meta allegation is a useful reminder that AI does not remove accountability from employers. If anything, it increases the need for accountable design.

When an organisation uses an algorithmic ranking, it is making choices about what counts. It chooses the data, the weighting, the exclusions, the thresholds and the review process. If those choices disadvantage people who are absent for protected or legitimate reasons, pointing to the software will not be a satisfying answer.

For employees, the concern is obvious: nobody wants their career reduced to a dashboard score that ignores the reality of their life. For employers, the risk is equally clear: badly governed AI can turn a difficult restructuring into a reputational, legal and cultural mess.

The lesson for UK businesses is not that AI should never touch HR. It is that AI should not be allowed to quietly redefine fairness. If a system cannot handle protected leave, disability adjustments and human context, it should not be anywhere near layoff decisions.