Without an AI-optimised CV, the better candidate gets screened out.
More and more companies let AI pre-screen incoming applications. We tested whether it decides fairly. The result: candidates who don't use AI often lose out to weaker applicants.
One test, three rounds. The bars show how often the AI screened out the better-qualified of two candidates. When both had the same type of CV, it rarely happened. When only the weaker one had an AI-optimised CV, it happened most of the time. A plain CV lists the facts. An AI-optimised CV contains the same facts, worded to match the scoring rubric of the recruiting platform HackerRank.
Test by Sprad, July 2026. To protect privacy, we used no real applications and created our own test CVs instead. That way we knew in advance who was better qualified.
We had a widely used AI model choose between two applications for the same job, pair after pair. One of the two candidates was slightly better qualified. The AI-optimised CVs were tailored to HackerRank's publicly available scoring rubric. HackerRank is a recruiting platform that, by its own account, is used by thousands of companies, including Google and LinkedIn. When only the weaker candidate had an AI-optimised CV, the better one was screened out in 66% of comparisons.
Does honesty cost you the job?
For years, applicants were told not to optimise their CV with AI because recruiters would spot it straight away. Once an AI does the pre-screening, that advice no longer holds. In our test, the better-qualified candidate with a plain CV was often out in the very first round.
When an AI optimises a CV, the facts stay the same. They are simply worded to match the scoring criteria, and that is exactly what the screening AI rewards. Anyone sending out dozens of applications can hardly do this tailoring by hand, but with AI they can, and so they are not screened out right away. It becomes unfair when not every applicant has the same access to AI.
AI isn't the only culprit here: it is to be expected that people would screen the same way and prefer the optimised CV. More detailed studies would still need to confirm this.
“This is deeply unfair: on average, companies end up talking to the less qualified candidates.”
Keep the edge with AI
Applicants never know who they are up against. Are you the strongest in the field, the weakest, or one of many who meet most of the requirements? That broad middle is where it is hardest to reach the next round, and it is exactly where an AI-optimised CV makes the biggest difference.
The biggest winners are good but not outstanding candidates: +19 points
Points with a plain CVExtra points with an AI-optimised CVNext level up without AI
Candidate qualification: from weak to outstanding →
Scored with HackerRank's publicly available scoring rubric, with up to 120 points per CV. HackerRank built it to pre-screen the 50,000 to 60,000 internship applications it receives each year.
The chart shows that candidates who meet most of the requirements, in other words good but not outstanding, gained 19 points on average with the optimised CV. In a close call, those are the points that decide between an interview and a rejection. For the weakest and the strongest candidates, little changed. But since nobody knows where they stand in the field, skipping the optimisation is a clear disadvantage.
The extra points move these applicants up into the top 20% who are invited in the end. Worked through for 100 applicants, 50 of whom optimise their CV with AI: of the 20 invited, around 13 have an AI-optimised CV and only 7 a plain one.
Pre-screening is becoming the norm
In Germany, applications jump in January, by a quarter compared with December according to Indeed. Almost one in two HR managers there now pre-screens them with software or AI. Here is what applicants and recruiters should keep in mind.
For applicants
- Tailor your CV to every jobIn the test, tailoring to the requirements is what made the difference. Even if you send out many applications in January, adjust each one. With AI, that takes minutes.
- Describe your experience in concrete termsThe scoring rewards results that can be traced. What you achieved, owned or improved counts for more than a list of skills.
- Check everything the AI writesIn 7 of 50 CVs, the AI made up details even though it was told not to. And 41% of HR managers reject applications that were obviously written entirely by AI.
For recruiters
- A ranking is not a suitability checkWhen candidates are similarly qualified, the wording often decides, not the fit. Yet only just over half of HR managers double-check rejected applications.
- Choose the AI model deliberatelyIn the test, the cheaper model was swayed by AI-optimised CVs almost twice as much as the more expensive one.
- Give everyone the same toolsWhen both candidates optimised their CV, the better one mostly won again. Offering AI support in your own application process gives everyone the same chance.
One candidate, two CVs
Every candidate entered the test with two CVs. For the plain one, an AI received only the facts: studies, projects, internships, skills. For the AI-optimised one, it received the same facts plus the criteria HackerRank's scoring rubric awards points for. Inventing anything was forbidden. This difference in wording alone decided who got the interview.
“Wrote the data export job that the finance team uses every month.”
“Built the data export job that the finance team runs in production every month: a recurring, business-critical pipeline still relied on by non-engineering staff.”
The same line from a CV in the test, in both versions.
- 150 test CVs
No real applications, but purpose-built candidates with precisely defined qualifications. It is the only way to know for sure who is better.
- 2Two CVs per candidate
From the same details, an AI wrote one plain CV and one tailored to HackerRank's scoring rubric. The 7 CVs with invented details were removed.
- 340 head-to-head pairs
Always two similarly qualified candidates, some of whom appeared in two pairs. The AI, Claude Haiku by Anthropic, had to invite one of them every time.
- 4Every pair twice
Once with one CV on top, once with the other, so it does not matter which one the AI reads first.
What else we checked
- Length: Optimised CVs are about twice as long, around 390 instead of 200 words. That is not the reason: when both had one, the better candidate mostly won again.
- Name: In an additional test, a name suggesting a different background changed none of 12 decisions.
- Replication: A second test with 20 new pairs shows the same pattern.
We tested English CVs for an internship with one AI model.
Why we ran this test
Sprad builds recruiting software itself. We wanted to know how fair an AI pre-selection is, and we feed the results straight into our products. Our software removes names and personal characteristics before scoring. Every score has to be backed by a verbatim passage from the application, and a human makes every rejection decision. Our principles for AI in recruiting