Research into AI-written CVs and AI-assisted recruitment raises a question for employers and candidates: how much should writing style influence who gets shortlisted?
A candidate uses AI to improve their CV. An employer uses AI to assess it.
Both are trying to make recruitment more efficient. But when AI operates on both sides of the application, an unexpected advantage may emerge: the system reviewing the CV could favour the kind of language it produces itself.
Recent coverage in the Financial Times has brought renewed attention to this issue. For international candidates and hiring teams, it raises a practical question about how professional ability is recognised.
AI-written CVs: what the research found
In AI Self-preferencing in Algorithmic Hiring: Empirical Evidence and Insights, researchers Jiannan Xu, Gujie Li and Jane Yi Jiang examined whether large language models favour CVs generated by the same model.
Their controlled experiments found that the tested models preferred their own generated CVs over human-written alternatives and those produced by other models, even when content quality was controlled.
In simulated hiring pipelines across 24 occupations, candidates using the same model as the evaluator were 23% to 60% more likely to be shortlisted than equally qualified candidates submitting human-written CVs. These are relative differences in simulated shortlisting, not percentage-point increases or measured improvements in real-world job offers.
The University of Maryland’s explanation of the research highlights a possible influence from model-specific writing styles, independent of applicants’ qualifications. This behaviour is described as self-preference bias.
The findings do not establish that every recruitment system behaves this way. They identify a risk in the models and evaluation settings studied.
When polished writing becomes a hiring signal
Clear communication matters. A well-organised CV helps an employer understand what someone has done and how their experience relates to a role.
The difficulty begins when presentation carries more weight than the experience behind it.
Consider two candidates who have done comparable work. One describes their contribution in straightforward language. The other uses a polished, AI-assisted summary. If the second receives a stronger assessment because its phrasing appeals to the evaluator, the ranking may reflect something beyond professional suitability.
For international hiring, this deserves particular attention. Candidates bring different writing conventions and levels of confidence in the language of the application. Employers need to decide which communication skills are essential for the job and how to assess them fairly.
An elegantly written paragraph can help explain a capability. It cannot, on its own, demonstrate that the capability exists.
What candidates can take from this
The research is not a reliable shortcut for choosing which AI tool should write your CV. It does not show that matching an employer’s model will improve your chances in every hiring process.
For AI-written CVs, a more useful approach is to use AI as an editor while keeping control of the substance.
Start with the facts: your responsibilities, the tools you used, the problems you addressed and the outcomes you can explain. Then ask AI to improve clarity, shorten repetition or help with translation.
Review the result carefully. A sentence that sounds more impressive may also imply more responsibility than you actually held.
For example, the following illustrative statement offers little detail:
“Delivered innovative solutions to optimise business performance.”
A more informative version might be:
“Built a weekly inventory report in Excel that helped the operations team identify stock discrepancies before ordering.”
The second gives a recruiter something concrete to explore. What data did you use? How did you build the report? What changed after the team started using it?
You do not need a numerical achievement in every sentence. You do need an accurate account that you can discuss confidently.
What employers should ask of AI screening
For hiring teams, the research offers a reason to examine how recommendations are produced.
Before relying on an AI-generated ranking, define the evidence that matters for the role. Relevant experience, specific skills and practical requirements should have an explicit place in the assessment.
A useful internal test is to present the same qualifications in different writing styles and check whether the ranking changes substantially. If it does, investigate what the system is rewarding.
Reviewers should also be able to connect a recommendation to information in the application. An assessment such as “strong candidate” is less useful than an explanation identifying relevant experience and showing where the role requirements remain unmet.
For later stages, focused interviews and proportionate work samples can help teams explore the capabilities described in a CV. Human review should involve questioning the recommendation and examining evidence.
These are practical steps we recommend when considering the implications of the research, rather than a guarantee that any one approach removes bias.
The relevance for Ekspats and Spark
At Ekspats, this discussion reinforces the importance of making skills, experience and work preferences explicit.
Spark is Ekspats’ international talent matching platform, where candidate profiles and company positions are compared through these factors. For candidates, maintaining accurate, detailed and current profile information is an important part of that approach.
Structured information can give hiring teams a clearer basis for comparison. It still needs thoughtful evaluation; using AI or structured profiles does not automatically make a system free from bias.
As AI becomes part of both writing and reviewing applications, recruitment technology needs to be judged by how well it helps people understand professional suitability.
The question worth asking of any recommendation is simple:
What evidence shows that this person could do this job well?
Sources: