The Legitimate Interest Assessment for Keystroke Capture to Train AI
In recent news,* Meta has decided to track employee keystrokes and clicks to train its AI. I have used this example to run an example legitimate interest assessment (LIA) of the same activity. Read why I think why consent fails, why necessity fails, and why the balancing test should too.
People are gathered in the kitchen area, so is the DPO, discussing the weather and current work.
Mark Product: “We have decided that the quickest and most efficient way to get more data for our AI feature is to track all staff’s keystrokes and mouse clicks.”
Sonya Engineering: “I see. How will we explain this to them, though, won’t we need their consent?”
Mark: “We will say it’s to improve the outputs for our customers, so I am pretty sure we can say it’s – what was it called – our legitimate interests (LI) to do this.”
DPO: “Hi, sorry to interrupt. Did you say we were going to track everyone’s keystrokes and mouse clicks?”
Narrator: The DPO did hear correctly and wishes that wasn’t the plan, knowing very well it is.
@sasun1990 on Unsplash
AI Processing Needs a New GDPR Lawful Basis
DPOs discovering current and future plans in this way is common. These kind of team conversations are happening more and more often, from “how can we use more AI at work” to “how do we make sure the AI output is accurate”.
The answer is with more data.
In this data, however, it is very likely that personal data will be included. This is personal data that would have been collected for reason A with a GDPR lawful basis A; but with the new “AI need”, we have a new reason B.
This new reason B covering the AI need needs a new lawful basis B. It is unlikely that a company could re-use the personal data already collected because people wouldn’t expect their information to be used this way – therefore requiring lawful basis B.
This is even more relevant and important to get right in an employment context, where the power balance is severe.
Why Employers Can’t Use Employee Consent for AI Processing
Employers considering this kind of activity can reach for LI as a lawful basis. This is not unreasonable. The recent data changes in the UK make it easier to use automated decision-making / AI on personal data (with safeguards).**
Consent under the UK GDPR must be freely given, specific, informed and unambiguous, and easily withdrawn without detriment (Article 7). The ICO has been consistent that the employer-employee power imbalance means consent is only appropriate where the employee has a genuine choice and control — and almost never do they have it in a monitoring context.***
An employee at a firm (Meta) that has just announced a hiring freeze, layoffs or a reorganisation is not in a position to freely refuse a programme their employer has framed as central to its AI strategy. Pressure does not need to be explicit to be coercive.
So, consent fails. That leaves LI as the only remotely plausible Article 6(1) basis (this is your lawful basis B).
In case you are wondering about the other lawful bases:
Contract does not fit because the monitoring is not necessary to perform the employment contract.
Legal obligation does not fit because no law expressly or otherwise required Meta to do this.
Vital interests and public task are not even in scope.
This is where the LIA does its real work.
Legitimate Interest Assessment of Keystroke Logging
An LIA has three parts: the purpose test, necessity test, and balancing test.
1. Purpose Test: Is There a Legitimate Interest?
Is the interest the employer is pursuing real, lawful, and articulated with enough specificity to assess?
For keystroke capture to train AI, the answer is almost always yes.
Improving the company's AI agents to be commercially competitive is a legitimate business interest. The ICO doesn’t require legitimate interests to be noble and neither does the UK GDPR as long as they are properly assessed (recital 47); commercial interests count, including the controller's own and those of third parties.****
Such an interest includes accelerating productivity to compete in a strategically important technology race.
The Meta case:
Meta's articulated interest — training AI agents that can navigate real software by capturing how people actually use it — is specific, lawful and commercially rational.
The purpose test is passed.
DPO note:
Passing the purpose test in the case of Meta was not going to be a hurdle, considering they are one of the leaders in AI training and development globally.
2. Necessity Test: Is the Processing Necessary?
This is the make-or-break moment and where the DPO's job can be the hardest.
The ICO is clear LI cannot be relied on for employee monitoring "if you can reasonably achieve the same result in a less intrusive way".+ For keystroke capture to train AI, there are less intrusive alternatives on the table:
A paid, opt-in panel of contractors or internal volunteers performing specific work tasks in a controlled environment.
Targeted, time-boxed studies on volunteers from specific teams, with genuine opt-in and meaningful incentives — not "help your daily work train the model" framed as part of the job.
The Meta case:
The infrastructure for either a paid or volunteer panel is possible.
DPO comments:
For most employers, the necessity test fails. The scale and persistence of workforce-wide keystroke capture is not necessary to obtain examples of how humans use software.
It should be demonstrated in writing by a senior manager why the alternatives are inadequate. Vague claims will not suffice.
Function creep is real as the scope expands, monitoring deepens, and the original justification is not revisited.
3. Balancing Test: Do the Rights of Employees Override the Interest?
This is the test that, in my assessment, almost no keystroke-capture-for-AI programme can pass. It is also the test that controllers convince themselves they have passed because the prize is large and the technology race is fast.
Do employees reasonably expect this processing?
Employees expect their employer to log security-relevant events, monitor for misconduct, and collect for product stats.
They do not reasonably expect every keystroke and mouse movement on their work device to be harvested as training data for the systems that may ultimately replace them.
When the people themselves describe the processing as a surprise — as Meta's own staff did, anonymously, to the BBC — the reasonable expectations limb is in trouble.
How intrusive is the processing and what kind of data is used?
Keystroke and screen capture is among the most intrusive forms of workplace monitoring on the spectrum.
Even with a defined application list and sensitive-content filters, the residual risk of capturing special category data (in Slack/Teams channels) is high. The ICO is explicit that incidental capture of special category data triggers Article 9 conditions, and none of those conditions sit comfortably with this purpose.
There are material adverse effects.
First, the chilling effect on employees who know every action is feeding a training pipeline will rationally change their behaviour. They may start writing less candidly, documenting less or raising concerns less often. That is a measurable harm to the workforce.
Second, the same data is being used to train agents that the same workforce reasonably believes are intended to reduce headcount. Employees are being asked, without genuine consent, to accelerate their own replacement. In Meta's case, this is a real possibility.
The power imbalance is the factor that, more than any other, tips the balancing test against the controller. The employment relationship is a context where people have very limited ability to meaningfully assert some of their rights, especially Article 21 right to object to the processing. It exists but an employee exercising it is exercising it against a programme their employer has publicly aligned with its strategic future.
DPO comments:
For workforce-wide keystroke capture to train AI, the balancing test fails.
The interests, rights and freedoms of employees to a reasonable expectation of privacy at work, to not be surveilled for the training of their own replacement, to be open in their professional communication outweigh the employer's commercial interest in faster, cheaper training data.
The programme should be redesigned around a paid volunteer panel with genuine opt-in, or it should not proceed.
LI is not a viable lawful basis.
Final LIA Sign-Off
The employer / controller may not agree with any of this. In fact, a different DPO may also say that the balancing test is passed.
The employer may say that balancing test is passed because the application list is narrow, sensitive content is filtered, and the data is not used for performance management. They may insist that employees retain Article 21 objection rights and can exercise them.
They may even argue, in Meta’s case, that staff should expect this kind of processing because AI is central to the corporate strategy and they chose to work there.
None of these anticipated counter-arguments, individually or together, address the structural problem: that the workforce did not and could not meaningfully agree to this. The harms will be real, systemic and daily, not just a one-off.
A DPO can advise, not decide. The advice may be acted on or it can be just recorded and everyone moves on as if no advice was given. Which of those happens depends on the organisation, their privacy culture and risk appetite.
* https://www.bbc.co.uk/news/articles/cvglyklz49jo accessed 02.05.2026.
** The Data (Use and Access) Act 2026 amends Article 22 of the UK GDPR.
*** https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/employment/monitoring-workers/data-protection-and-monitoring-workers/ accessed 04.05.2026
**** https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/lawful-basis/a-guide-to-lawful-basis/legitimate-interests/?q=appropriate+policy accessed 04.05.2026.
+ https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/employment/monitoring-workers/data-protection-and-monitoring-workers/#dp2 accessed 04.05.2026