Jobseekers now use generative AI to write CVs, prepare for interviews and complete trial assignments. This can lead to more accurate applications and faster preparation, but it can also blur the line between what candidates know themselves and the answers AI generates. Employers therefore need to decide in advance at which stages of the selection process they will allow AI use, when candidates must disclose it, and at which points and by what methods they can assess candidates’ ability to work independently.
When a candidate who does not exist completes the selection process
A synthetic avatar took part in one of Grant Thornton UK’s digital interview processes*. It spoke naturally, gave appropriate answers and ranked among the top performers/in the 99th percentile. It outperformed the vast majority of real applicants. The application only raised suspicions when the phone number provided turned out not to be real.
The example highlighted a fundamental problem with assessment. If an automated tool can achieve an outstanding result in the selection process, an interview or an assignment alone may no longer reveal a candidate’s abilities. Grant Thornton UK therefore worked with Amberjack** to develop an assessment method that examines how applicants use AI. It looks at the questions they ask, whether they check the answer and spot errors, and whether they can improve the result using their professional judgement.
This approach also offers lessons for employers in Hungary. In many roles, AI is now part of everyday work. A blanket ban, however, can prevent employers from assessing a skill candidates will need once they start the job, while unrestricted use can mask gaps in their professional skills. One possible solution is to apply a set of rules that are established in advance and whose application can be checked throughout the selection process.
A ban alone is not enough
At first glance, a complete ban throughout the process may seem straightforward, but it is difficult to enforce in practice. When an assignment is completed at home, it is rarely possible to establish with certainty which tool helped the candidate. AI detection tools are too unreliable to serve as the sole basis for a selection decision, and uncertainty over whether the rules were followed can easily lead to inconsistent assessments.
The process is more transparent when the company tells applicants what it expects from the outset. Candidates need to know whether they can use AI when preparing their application materials, completing a take-home assignment or attending an interview. The rules can also specify when applicants must name the tool they used and briefly explain what they used it for. This allows assessors to base their evaluation on how the candidate actually worked rather than on guesswork.
Three situations, three sets of rules
Each stage of the selection process serves a different purpose, so it makes sense to treat AI use differently at each stage. A well-structured policy can distinguish between three situations:
- AI use is unrestricted: AI can help candidates learn about the role, prepare for an interview, edit their CV or check its language. Candidates remain responsible for the accuracy of the facts they provide.
- AI use is allowed with disclosure: If the selection assignment involves preparing a case study, analysis or presentation at home, it is worth asking candidates to explain briefly which AI tool they used and how they used it. For some assignments, it may also be appropriate to ask them to share their prompts or workflow.
- AI-free stage: External assistance can be explicitly ruled out in an in-person interview designed to assess candidates’ own professional knowledge, numeracy, communication skills or ability to read a situation. Candidates must be told in advance, and the same conditions must apply to every applicant.
Assessment must go beyond the final answer to include how candidates work
An AI-assisted assignment yields meaningful information when employers assess more than the final document or answer. It is at least as important to evaluate how candidates arrived at that result. Do they spot inaccurate claims and verify sources and figures? Can they explain which suggestions they accepted and which they rejected? Can they adapt the answer to the specific client, situation or business objective?
In Grant Thornton UK’s experience, this approach makes critical thinking, decision-making and human skills the focus of the assessment. In professional services, asking the right questions, understanding the client’s situation, taking responsibility and building trust are particularly important. A perfectly worded AI answer does not in itself demonstrate these skills.
A consistent process protects candidates and employers alike
Rules on AI use work when they are incorporated into the entire selection process. It is worth reviewing assignments to see whether they actually assess the abilities that matter for the role. Assessors need a common set of criteria, and candidates must be given the same tools and conditions if using AI is part of the assignment. Particular care is also needed to ensure that differences in access to technology do not put candidates at an undue disadvantage.
The AI tools employers use raise a separate question. The European Union’s AI Act*** classifies certain AI systems used in employment and recruitment as high-risk. Companies therefore need to map where algorithms help screen, rank or assess applicants, what data they use and where humans retain meaningful decision-making authority. Rules governing candidates’ AI use are only credible if employers are also transparent about how they use technology themselves.
Good rules start with the purpose of the selection process
Each organisation must set boundaries that reflect its own roles and selection methods. The aim is simple. The process should show what candidates can do independently, how they work with AI and what decisions they make when the tool gives an inaccurate or incomplete answer. This requires clear expectations, well-designed assignments and consistent assessment.
