Candidates use AI to produce more applications, better worded. Recruiters, overwhelmed, use AI to filter more. Each side equips itself in response to the other's equipment. The result is a heavier system for everyone, where the useful signal weakens as volume rises.
Recruitment is one of the few areas where both sides of a single interaction have equipped themselves simultaneously. The result is worth looking at, because it foreshadows what will happen elsewhere.
On the candidate side
What AI brings is real and legitimate. Tailoring a cover letter to each job ad, reframing experience in the sector's vocabulary, fixing an awkward phrase: this reduces a disadvantage that had nothing to do with skill. Someone who writes poorly, who is not a native speaker, or who does not know the codes of a given field was penalised on form rather than substance.
From that standpoint, it is a rather welcome levelling, in line with what we described for accessibility.
The problem appears with volume. When applying costs nothing anymore, you apply everywhere. The response rate collapses, which pushes people to apply more, which collapses the response rate further. The serious candidate ends up drowned among those who sent three hundred applications in one evening.
On the recruiter side
Faced with this inflation, automated screening becomes a practical necessity rather than a choice. And it raises three difficulties.
Filtering on formal criteria. A system that sorts on keywords rewards those who know how to write for the machine, not those who would do the best work. Since candidates now use tools that optimise precisely for that, the filter loses its discriminating power.
Inherited biases. A system trained on a company's past hiring reproduces its habits, including its biases. This is the mechanism we described in our article on impossible neutrality: the model learns what it has been shown, and if it has been shown twenty years of homogeneous decisions, it learns homogeneity.
The lack of explanation. A rejected candidate does not know why, and often the recruiter does not either, for the reasons we developed on model opacity.
In Europe, a decision producing significant legal effects cannot rest on fully automated processing without human intervention, and the person concerned has rights to information and to contest the decision. The European AI regulation also classifies recruitment among high-risk uses, with reinforced obligations. Many tools deployed today sit in a zone where compliance deserves serious scrutiny.
What seems to work better
A few practices stand out among those who have taken the issue seriously.
Moving assessment towards situational exercises. If a CV can be optimised by machine, a practical task or an in-depth conversation discriminates more. This is exactly the logic we described for school homework: when production can be delegated, you have to assess differently.
Using AI to broaden rather than to filter. Some teams use it to spot atypical profiles that a standard screen would have rejected, which is the opposite of the problematic use.
Ensuring real human involvement. Not a rubber-stamp validation, but an actual reading of the shortlisted applications.
What to take away
This field illustrates a phenomenon that will appear everywhere: when both sides of an interaction equip themselves with the same tool, the net gain can be zero, or even negative. Everyone runs faster to stay in the same place.
What becomes rare again, in this context, is the costly signal: a recommendation from someone who stakes their credibility, work actually shown, a meeting. These are precisely the things AI cannot produce, and their relative value rises as everything else becomes easy to generate.