The Loop Problem: When AI Screens AI
A candidate uses AI to reverse-engineer your job description and build a resume optimized for how your system scores.
Your applicant tracking system then uses AI to scan that resume for signals of qualification.
The score the candidate receives reflects how well their AI understood your AI’s criteria. That may or may not have anything to do with whether they can do the job. You are no longer evaluating candidates. You are running a contest to see whose AI is better.
The data confirms it is already happening
According to Greenhouse’s 2026 AI Hiring Report, 91% of recruiters and hiring managers have spotted or suspected candidate deception, and 74% say they are more worried about fake credentials than a year ago. AI-generated resume exaggeration was the most common form observed, reported by 63% of respondents.
That data raises a harder question. If your screening process cannot distinguish a genuinely qualified candidate from an AI-optimized application by someone who is not qualified, what is it actually measuring? In most cases, vocabulary alignment. Whether the document contains the same words as your job description. That is keyword matching with extra steps.
From a keyword problem to skills in action
Tigran Sloyan, CEO of the skills platform CodeSignal, put it plainly: the existing process winnows candidates by searching for skills keywords in resumes as a filter. That was always a weak proxy for competency. AI has made it a broken one. When candidates know your system weights certain terms, they include them. When AI writing tools know the same thing, they add them automatically.
The organizations moving past this are not abandoning AI. They are changing what they ask it to do, building skills-in-action assessments that replicate real work. Sloyan’s framework separates three tiers worth screening for:
- Essential skills: core capabilities tied directly to job outputs.
- Foundational skills: the underlying capabilities that make essential skills possible.
- Tool skills: platform-specific proficiency that changes as technology changes.
The governance gap
There is a quieter dimension: legal exposure. In May 2025, a federal judge conditionally certified a class action against Workday covering applicants over 40 rejected by its AI screening since 2020, filings alleging roughly 1.1 billion applications during the period. The EEOC settled its first AI age-discrimination case in 2023, involving a tool that rejected older applicants. The tell was an applicant who submitted two identical resumes with different birth dates. Only the younger one got an interview.
The Loop is not an argument against AI. It is an argument for understanding what your AI is doing, which most organizations cannot fully explain.
