Lesson 1: What AI Can and Cannot Do in HR

Before you start: general information, not legal advice

This course gives general information about using AI in HR and recruitment in the UK. It is not legal advice. Employment law and data protection law depend on the facts of each case, and both change. For example, the Data (Use and Access) Act 2025 received Royal Assent on 19 June 2025. It replaced the UK GDPR rules on automated decisions, and its provisions come into force in stages.

Ask an employment lawyer or your data protection officer (DPO) when:

• a decision about a real person (rejection, dismissal, disciplinary action, refusing an adjustment) is based partly on AI output
• your organisation plans a new tool that will process candidate or employee data
• a case involves disability, pregnancy or another protected characteristic, or there is a risk of a tribunal claim
• you are unsure what the law requires today

What a general AI assistant actually is

Tools such as ChatGPT, Gemini, Claude and Microsoft 365 Copilot predict likely text. They are fast and fluent, but being fluent does not make them right. The vendors say this themselves. OpenAI's help pages warn that ChatGPT can produce incorrect or misleading answers that sound confident, including made-up references. Treat an assistant like a quick new junior colleague: helpful for first drafts, but not someone whose work you sign off unchecked.

The three buckets

Every HR task you might give to AI fits into one of three buckets.

1. AI can help draft or summarise (you check everything):

• a first draft of a job advert or job description
• a bank of structured interview questions linked to the role's criteria
• rewriting a letter or policy section in plain English
• an outline for an onboarding checklist or training session
• summarising your own non-confidential notes or a long public guidance document

2. AI does these badly (verify every word or don't use it):

• quoting current law, case names or official guidance. It may invent citations or describe out-of-date rules
• statistics and salary benchmarks. Numbers may be made up or undated
• facts about your organisation, such as your policies, pay scales or history, unless you supplied them
• judging candidates consistently. The same CV can be rated differently each time

3. These must stay with a human:

• final decisions to shortlist, reject, hire, promote or dismiss
• disciplinary outcomes and findings in investigations
• decisions on reasonable adjustments
• difficult conversations: bad news, grievances, wellbeing concerns
• signing off anything sent to a candidate or employee

Worked example: you ask an assistant for a rejection letter. Drafting the wording belongs in bucket 1. If the draft says "as required by the Employment Act 2023, section 12", that is bucket 2: check it, because the reference may be invented. Deciding who gets the letter is bucket 3.

Task 1: Sort your own week

Do this now, using task names only and no personal data:

1. List ten tasks from your last working week.
2. Ask: does this task decide something about a person? If yes, it goes in bucket 3 and stays human.
3. Ask: does it depend on facts, law or figures the tool could get wrong? If yes, it goes in bucket 2, and you verify against a primary source or leave AI out.
4. Otherwise it goes in bucket 1. AI may draft it and you review it before use.
5. Note one bucket-1 task you will try during this course.

What the evidence says about the profession

Always quote a statistic with its source, date and limits.

• CIPD Resourcing and talent planning report 2024 (UK, more than 1,000 HR and people professionals): 31% of organisations used some form of AI or machine learning in recruitment, up from 16% in 2022. Of those using it, 66% said it improved hiring efficiency.
  • Shows: AI in recruitment roughly doubled in two years.
  • Does not show: whether hiring became fairer or more accurate. The efficiency figures are self-reported, and the survey does not separate chat assistants from specialist screening software.
• SHRM State of AI in HR 2026 (published April 2026, 1,908 HR professionals): 39% of organisations had AI in their HR function. Recruiting was the most common area at 27%.
  • Shows: adoption is real but far from universal, and recruitment leads.
  • Does not show: UK practice (SHRM is a US body) or how well the tools work.
• ICO audit outcomes report, 6 November 2024: after auditing AI recruitment tool providers, the ICO made nearly 300 recommendations. It found that some tools inferred gender and ethnicity from candidates' names, and some let recruiters filter out applicants by protected characteristics.
  • Shows: real tools on the market have had serious fairness and data protection flaws.
  • Does not show: how common these problems are across all tools.

DO: write one sentence for each statistic in the form "[Figure], [source], [date]: shows…, does not show…". Use this habit every time you brief a manager.

Task 2: Spot the made-up fact

1. Ask an approved assistant: "Summarise UK law on using AI to screen job applicants, with sources."
2. Highlight every case name, section number, date and statistic in the answer.
3. Check each one against the primary source (legislation.gov.uk, ico.org.uk, cipd.org).
4. Count how many you could not confirm. That count tells you why bucket 2 exists.

Key takeaways

• This course is general information, not legal advice. Use a specialist for decisions about real people, for new tools and for anything with tribunal risk.
• Bucket 1: AI drafts and summarises, and you check. Bucket 2: law, facts and numbers need verifying. Bucket 3: decisions about people stay human.
• Assistants can sound confident while being wrong, including invented citations.
• Quote statistics with source, date and limits: CIPD 2024 (31%, up from 16% in 2022), SHRM 2026 (39%, recruiting 27%), ICO November 2024 (tools inferring protected characteristics).
• The goal is to keep your professional judgement and use AI to save time on the work around it.