Hiring in the AI era: keeping assessment fair and trustworthy.
Candidates can now use AI to write applications, rehearse interview answers and, sometimes, complete online tests. That changes what a polished application or a strong online score can tell you.
On this page
- Decide and publish your policy on candidate AI use
- Treat unsupervised results as provisional
- Put more weight on evidence formed live
- Choose formats that help, without relying on them
- Do not rely on AI detectors
- Expect polished applications, and sift accordingly
- Keep it fair to everyone
- Hold your own use of AI to the same standard
- A process that holds up
Eight steps to make sure the evidence you rely on is the candidate’s own, plus the UK rules if you use AI yourself.
Most candidates who use AI are doing what they have always done with a friend who writes well: asking for help. A few will use it to answer questions that were meant to be theirs alone. A fluent personal statement or a high score on an unsupervised test may now reflect the candidate, the tool, or a mix of the two.
Two responses are tempting and both are mistakes. One is to carry on as before. The other is to treat every good application as suspicious. The better course is to decide which evidence must be the candidate’s own, and to design the process so that it is.
Decide and publish your policy on candidate AI use
Be specific. Using AI to check spelling, structure a CV or prepare for an interview is something many employers will accept. Using it to answer test questions, write answers to assessed written questions, or feed answers during a live interview is something most will not.
Write down what is allowed at each stage, tell candidates before they start, and ask them to confirm they have followed it. Clear rules protect honest candidates and give you a fair basis for acting if they are broken.
Treat unsupervised results as provisional
Anything completed at home, unsupervised and online – an ability test, a situational judgement test, a written exercise – can be useful as an early sift, but should not settle a decision on its own. We suggest building in a supervised step for shortlisted candidates: a short re-test, a live version of the exercise, or a conversation in which the candidate talks through their written answers and the reasoning behind them.
Put more weight on evidence formed live
Some evidence is hard to outsource. A structured interview with real follow-up questions asks the candidate to go beyond any prepared answer: “What exactly did you say?”, “What would you do differently?”. A work sample done in the room or on camera, a presentation followed by questions, a role-play with a realistic customer or colleague: all of these show what the candidate can do, not what they can generate. See how to run a structured interview.
Choose formats that help, without relying on them
Some formats are harder to game than others. In one study, when AI chatbots were asked to complete personality questionnaires as an ideal applicant would, forced-choice formats – where candidates choose between equally attractive statements – reduced their ability to inflate scores, but did not remove it; the most capable model still scored highly (Phillips and Robie, 2024, using models available in mid-2023).
In another, five chatbots all performed at least as well as the applicants on a text-based situational judgement test, and three scored significantly higher (Mittelstädt and colleagues, 2024). Format helps. It is not a defence on its own.
Do not rely on AI detectors
Tools that claim to detect AI-written text are unreliable. OpenAI withdrew its own classifier in July 2023, citing its low rate of accuracy. A detector score is not evidence, and rejecting or accusing a candidate on the strength of one risks being both wrong and unfair. If you have a concern about an application, test it at interview: ask the candidate to expand on what they wrote.
Expect polished applications, and sift accordingly
When everyone can produce a fluent covering letter, fluency stops telling you much. Score applications against two or three specific criteria agreed in advance, and ask for concrete evidence: what the candidate did, with whom, with what result. Consider replacing long open statements with a few short, specific questions about real experience, which you can then follow up at interview.
Keep it fair to everyone
Apply the same rules to every candidate. Remember that some disabled candidates use assistive technology, including AI-based tools, and that a request to use one may be a request for a reasonable adjustment. Make it easy to ask, and decide case by case. Check whether candidates from different groups are progressing through each stage at different rates, especially after you change the process.
Hold your own use of AI to the same standard
If you use AI to screen, rank or assess candidates, the rules are clear and tightening. See below.
If you use AI in your own hiring process (UK)
- Data protection. The ICO audited AI recruitment tools between August 2023 and May 2024 and, in November 2024, published recommendations for recruiters and providers: monitor tools for accuracy and bias; tell candidates how AI is used and how to challenge decisions; collect only the data you need and set retention limits; complete a data protection impact assessment early; be clear who is controller and who is processor; and document your lawful basis.
- Automated decisions. Since 5 February 2026, new Articles 22A to 22D of the UK GDPR, inserted by the Data (Use and Access) Act 2025, govern significant decisions based solely on automated processing. Such decisions are permitted subject to safeguards, including telling people about them, letting them contest the decision and obtain human intervention; tighter restrictions apply where special category data is involved. A rejection made by a tool with no meaningful human involvement is likely to be a significant decision.
- Equality. The Equality Act 2010 applies to the outcome however the decision was reached. If a tool disadvantages a group, the employer answers for it.
- Buying well. The Department for Science, Innovation and Technology’s Responsible AI in recruitment guide (March 2024) sets out questions to ask suppliers before and after you buy.
This is a summary, not legal advice. Check the current position with your data protection officer or adviser before you introduce a tool.
A process that holds up
Put together, a process for the AI era looks familiar: an application scored against agreed criteria; an online sift where volumes justify it, treated as provisional; then supervised, live evidence – a structured interview with real probing, and a work sample or exercise that shows the candidate doing the job – with references at the end to verify the history. None of it requires new technology. It requires deciding which evidence counts, and making sure you have seen it first-hand. The selection method chooser suggests a mix for your role; the SJT explainer covers one of the formats most affected.
Drawn from great{with}talent’s work on assessment design for AI-era recruitment (2025), rewritten as general method. Research: Phillips, J. & Robie, C. (2024), “Can a computer outfake a human?”, Personality and Individual Differences 217, 112434; Mittelstädt, J. M. et al. (2024), Scientific Reports 14, 27449. Regulation: ICO, AI tools in recruitment audit outcomes (November 2024); UK GDPR Articles 22A–22D (Data (Use and Access) Act 2025, s.80; in force 5 February 2026); DSIT, Responsible AI in recruitment (March 2024). Assessment for every stage of hiring is part of FindingPotential.
The clearest thirty minutes in your week.
Book a walk-through with a real person, or start with a pilot or a 30-minute conversation.