A recruiter has forty applications on her desk and a deadline. She drops ten CVs into a free AI tool and asks for a summary of each candidate. The documents contain full names, dates of birth and addresses and, for two of the candidates, a note about an earlier period of long-term sick leave. The summaries arrive in thirty seconds, and the data has left the building.
Across the office, a manager is writing up his notes from a performance review. He asks an AI to turn them into a polished narrative. The text reads well and feels complete. What he does not notice: one paragraph describes an incident that never happened. The model invented it to round out the story. The write-up goes into the personnel file unchecked.
These are not edge cases. They are the two most common ways AI goes wrong in HR: sensitive data entered too easily, and outputs trusted too readily.
#Is AI safe to use in HR?
AI can be used safely in HR, and the time saved is real: summarising, rewriting, structuring, drafting job adverts. But HR is not an ordinary function. HR handles special-category and highly sensitive personal data every day: health, finances, appraisals, identity. A small mistake is easy to make, and the consequences are larger than in almost any other team.
The biggest risk is not the AI technology itself. It lies in two human moments: when someone pastes sensitive data into a tool, and when someone accepts an AI output as fact. Since 2024 there has been a legal dimension too: the EU AI Act classes a large share of HR uses as high-risk, with explicit requirements for human oversight. The seven rules below cover both moments.
Before you set policy, it helps to know which tools your team is probably already using. They fall into two broad groups, and each carries a different kind of risk.
The first group consists of general-purpose AI tools that people paste text into:
- ChatGPT (OpenAI): drafting job adverts and emails, and summarising CVs and review notes. By far the most widely used, and exactly where the paste risk in rule 1 begins.
- Microsoft Copilot: built into Word, Outlook and Teams. Many HR teams already use it without thinking of it as "AI", so sensitive content gets processed unnoticed.
- Google Gemini: the same role inside Google Workspace.
The second group consists of HR-specific systems that help make decisions:
- LinkedIn Recruiter: AI-driven candidate matching and suggested messages; it shapes who you even get to see.
- HireVue: AI video interviews and assessments. Tools like these often fall into the EU AI Act's high-risk category, with duties around human oversight and transparency.
The distinction determines where you step in: in the first group the risk lies in what an employee enters, in the second in what the system itself decides.
Knowing the two groups exist is one thing. Knowing which specific tools your own team has opened (LinkedIn Recruiter, an AI note-taker, a browser extension nobody logged) is another. Most HR leads can name the two or three tools that were formally approved. Almost none can say what else is running alongside them.
Each tool comes with a risk score already attached, drawn from BeeSensible's catalogue of over 1,000 AI tools, so "which tools does HR use" stops being a guess about the two you rolled out and becomes a picture of everything your team has opened.
#Seven privacy rules for AI in HR
#1. Never paste CVs into public AI
Uploading a CV or cover letter to a free tool for a quick summary is a direct data breach. You lose control of personal data a candidate entrusted to you, and you have no clear legal basis for the processing under the GDPR. Many free consumer tiers can also use submitted content to improve their models unless the user switches this off, and that is an individual setting, not an organisation-wide safeguard.
Use only approved, closed AI workspaces covered by a data processing agreement, or strip every identifying detail before you enter anything.
That stripping step is exactly where a check as you type helps. It does not wait for a policy review: it flags the name, the national identifier or the note about sick leave the moment it lands in the prompt box, so the recruiter removes it before pressing Enter rather than after the summary has already gone out.
Here is what that looks like as someone types a CV summary request:
#2. Watch for the bias trap
AI systems learn from historical data. If that data reflects a past preference (for certain schools, ages, names or a gender), the model reproduces it without anyone meaning to. An algorithm that ranks "the strongest candidates" can quietly discriminate on grounds of age, gender or background.
So never let an algorithm decide who the best candidate is on its own. Keep reviewing outputs, monitor for unequal effects across groups and be ready to explain why a candidate did or did not make the shortlist.
#3. Always keep a human in the loop
Under the EU AI Act, many HR uses (automated screening tools, systems that rank candidates, tools that evaluate performance or allocate work) fall into the high-risk category. For those systems, meaningful human oversight is not a recommendation but a legal requirement. The GDPR likewise restricts decisions based solely on automated processing that significantly affect a person.
Treat AI as your assistant, not the final decision-maker. When hiring, appraising, promoting or dismissing people, a person should make the decision and be able to justify it.
#4. Anonymise the data before the prompt
Using AI to turn exit-interview or appraisal notes into a clean write-up? Remove every identifying detail before you enter it. The AI needs the substance of the conversation to rewrite it. It does not need the name, the national identifier or the date of birth.
Work with placeholders and swap them back afterwards:
- "James Carter (Sales)" → [Employee A]
- "Client X" → [Company]
- "DOB 12 March 1987" → [date of birth]
#5. Beware of hallucinations
A major risk is over-reliance on AI. Language models can invent facts ("hallucinate") to make text feel complete and convincing. In a marketing draft that is annoying. In an appraisal or a dismissal file it is dangerous: an invented incident in an official personnel record can damage a career and leave the organisation legally exposed.
Check AI-generated write-ups word for word. AI can take care of the wording; the facts remain your responsibility.
#6. Be fully transparent
Employees and candidates must know when and how AI is used in the HR process. Where algorithms influence recruitment, promotion or appraisals, both the GDPR and the AI Act require you to communicate that openly. Transparency is a duty, and above all a matter of trust: people who discover that AI quietly helped decide their career lose confidence faster than you can rebuild it.
Describe your use of AI specifically in your privacy notice, your recruitment process and your staff handbook, and be clear about who makes the final decision.
#7. Offer safe alternatives
A ban without an alternative does little. Surveys consistently find that a large majority of employees use unapproved AI tools at least sometimes, often because the approved route is missing or too cumbersome. This is shadow AI: without a good route, people find another one.
Instead, make sure your HR team has access to safe enterprise versions, and invest in AI literacy so everyone knows what is and is not allowed. The goal is not less AI, but safer AI on a route people actually use.
#The cheat sheet: what you never paste into public AI
HR handles data that becomes a breach in the wrong context. These are the categories that should never go into an open AI tool unfiltered, each with a safer alternative.
The table above covers what happens in a prompt box. But plenty of this same data reaches AI the long way round: as an attached PDF, not typed text. A CV, a signed appraisal form, an exit report exported from an HR system: these get opened and shared as documents well before anyone thinks about pasting a paragraph into a chat window.
Opening the file first and seeing exactly which fields are sensitive, before it goes anywhere near an AI tool or an external inbox, closes the gap the cheat sheet above cannot: the moment the data leaves as a document rather than a prompt.
#What is really at stake
Under the GDPR, entering identifiable personal data into an unsuitable tool can be a data breach. Organisations must assess every breach: where there is a risk to people's rights and freedoms, it must be reported to the data protection regulator (the ICO in the UK) within 72 hours of becoming aware of it. Health data, national identifiers or financial details usually bring an incident within that duty. GDPR fines can reach €20 million or 4 per cent of global annual turnover. Liability sits with the organisation, not the individual employee.
On top of that comes the EU AI Act. For HR systems classed as high-risk, there are obligations around human oversight, transparency and documentation, with their own enforcement and penalties, separate from the GDPR. A hiring algorithm that decides without human review is a privacy risk, and it may be a breach.
And then there is reputation. A candidate who learns their CV was fed into an AI tool without consent, or an employee who finds an invented incident in their file, will talk about it. In a labour market built on trust, that damage is hard to undo.
#Where HR has already gone wrong with AI
Amazon stopped using an experimental AI system that scored CVs after finding it disadvantaged women. The model had been trained on a decade of applications, mostly from men, and had learned to downgrade CVs containing the word "women's". The system never went live. Source: Reuters, October 2018.
#2023: iTutorGroup and the EEOC
The US Equal Employment Opportunity Commission settled a case against tutoring company iTutorGroup for $365,000. The company's recruiting software automatically rejected applicants by age: women aged 55 and over, and men aged 60 and over. It is an early example of enforcement against age discrimination by an automated HR system. Source: U.S. EEOC, August 2023.
#2023: Samsung restricts generative AI
Samsung restricted the use of generative AI tools after employees entered confidential company information into ChatGPT. The incident captures the core risk of public AI at work: not an outside attack, but data that employees enter themselves. Source: Bloomberg, May 2023.
#2024-2025: Mobley v. Workday
A US federal court allowed a discrimination case against HR software vendor Workday to proceed, on allegations that its AI-based screening tools disadvantaged applicants on grounds of age, race and disability. In 2025 the court permitted an age-discrimination claim to advance on a collective basis. The litigation is ongoing, but it signals that the providers of AI hiring tools, as well as the employers using them, may face scrutiny. Source: reporting on Mobley v. Workday, U.S. District Court, 2024-2025.
The common thread is not a single vulnerability in one tool. It is the combination of human action, trust in the output and the speed at which AI processes data.
#What settings and policy do not solve
Settings decide what an AI tool may do with your data. Policy decides what is allowed. Neither decides what an employee types into the box on a busy Tuesday afternoon.
No setting stops someone copying a CV into a free tool. No policy intercepts a date of birth or a national identifier in a meeting note before it is submitted. The most common privacy mistakes in HR are not configuration errors. They are decisions made in a few seconds, during ordinary work, by people who are not thinking about privacy at that moment. That is the point settings and policy do not reach.
#Before the prompt: what HR sees while typing
BeeSensible checks text fields for personal data, including the input box of AI tools, as you type or paste. The extension automatically recognises 65 types of sensitive data: names, national identifiers, dates of birth, phone numbers, bank account numbers and more. When sensitive content appears, BeeSensible highlights it in the text: yellow for sensitive data such as a name or a date of birth, red for critical data such as a national identifier. For each highlight, the employee can mask it with a numbered label such as [NAME1], or remove it, before the text goes to the AI tool.
Analysis runs in memory on an EU-based detection service, and the text is discarded after detection. Message content is not stored. The employee decides, unless the organisation has set "Mask on paste" or "Mask on send" to "On by default" or "Always on".
In HR, that means a CV summary, a performance write-up or an exit note containing a name, a national identifier or a date of birth is flagged in the input box before the prompt is sent. If a recruiter drags CVs into ChatGPT or Claude, the organisation can switch on "Mask on attach": BeeSensible then holds the file first, and the recruiter chooses to review and mask it or to add the original.
AI is fast, and that speed is exactly what makes it useful. The same speed causes most privacy mistakes in HR. Settings, policy and AI literacy reduce risk, but they do not reach the moment when someone pastes a CV or asks for a write-up. A highlight in the input box does: it shows what is in the prompt before it leaves the organisation, without blocking anything by itself.
From tomorrow, take the name and the identifier out of the CV before you press Enter, and read the AI's write-up before it goes into the file.
ChatGPT
