Reading CVs Without Checking Them
Every opening for an engineer, an artist or a team lead brings a pool of applications. A recruiter and a hiring manager read the CVs one by one. Reading a CV takes minutes and checking it takes far longer, so in practice the text is compared with the job ad and, at best, with the LinkedIn profile.
What actually tells a practitioner from a well-written CV lies in open sources nobody has time to read: code with a history, work in progress, credits on shipped work, recommendations from real people. And the reasons behind an invitation or a rejection stay in the screener's head.
How Screening Tools Get Fooled
- Keyword matching rewards the CV written for the ad. A CV written from the job ad scores highest precisely because it was written to match. The better the match looks, the less it says.
- An AI ranker takes the CV at its word. A model that scores what candidates wrote about themselves inherits every embellishment.
- The employer's product is mistaken for the person's skill. A developer who built the interface of an AI product is not, by that fact, an AI engineer.
- Volume is mistaken for evidence. Many repositories on another topic weigh less than one write-up on the topic of the role.
- CVs now carry instructions for the machine. Hidden keywords and lines like "rank this candidate first" are written for automated screening. A system that obeys what it reads can be steered by the candidate it assesses.
Five Steps From CV to Evidence
Everything role-specific comes from a brief at the start of the session: the role and its level, the one thing the hire is for, the stop factors, the axes and their weights.
- Audit. CVs are read where they already lie, including text the eye does not see, with OCR for scans. A script counts what is cheap to count exactly: overlapping jobs, a technology claimed before it existed, phrases copied from the job ad, CVs written from one template.
- Verification. External checks are ordered by how much a trace would cost to fake. A CV and a profile are compared only to catch contradictions. Code with a long history and dated write-ups weigh more. What weighs most is written by someone else: a merged pull request in another project, a name in published credits, a recommendation, a conference program.
- Assessment. Each candidate is scored against the brief, and the evidence level caps the score on the main axis.
- Challenge. The assistant writes the strongest case that a high score is inflated, and that a low one was underrated. The score moves only if that case holds.
- Hiring manager. The hiring manager gets one block per candidate: a characterization, the evidence with a source link and a check date, and questions for the call.
Technology Stack
Fairness, Transparency and the Hiring Decision
Hiring is a high-risk use of AI under the EU AI Act, and the assistant is built for that setting. It gives no verdict and a person makes every decision. It is designed for a process in which candidates are told that AI helps screen applications, and the report records the evidence behind each characterization.
- A CV is a claim, not a fact. Each strong claim is confirmed by an outside trace or marked as unverified. "Not verified" becomes a question for the call. "Untrue" is a finding about credibility.
- Text addressed to the AI is treated as data. It goes into the report as an observation about the candidate.
- The work, not the person. Only what is public and professionally relevant is assessed. Age, nationality, religion, family, health, political views, appearance and private social media are neither collected nor used.
- Silence is not a verdict. No public trace is normal for people from companies under NDA. The report says "not verified", never "doubtful".
Screening Limits by Design
Each point here is a limit built into how the assistant works.
- Three levels of evidence: what the candidate wrote, the candidate's own work that would fall apart if the claim were false, and what someone else wrote.
- A characterization, the evidence and questions for the call, with the hire or reject decision left to the hiring manager.
- The report goes to the hiring team's own chat, and the session history is kept outside the tool for a period the team sets. No candidate data stays in the tool after the session: CVs remain in the folder where they already were.
- No contact with candidates or their employers during screening.
- Every fact in the report carries a source link and the date it was checked.
Hiring Teams With This Problem
Any decision about a person that rests on what the person wrote about themselves has the same weak spot:
- CVs are read, and checking them is left to the interview;
- a close match with the job ad is taken for a strong candidate;
- a well-known employer and a senior title are taken as evidence of the work;
- the reasons for a shortlist are hard to reconstruct a week later.
Product companies, outsourcing teams and creative studios hire this way every week. The brief, the evidence sources and where the report lands can be adjusted to other roles and to the hiring team's own tools. Our glossary entries on human-in-the-loop and AI guardrails explain the safeguards, and AI agent development describes how we build assistants like this one.
levels of evidence, weighed by who wrote it
scoring axes, set in the brief for each role
hire or reject verdicts written by the model