Generation 1: Boolean search (1995–2010)
The original. Recruiter types "Java AND (microservices OR Kubernetes) NOT junior" into the ATS and gets a list of CVs containing those tokens. Brittle, slow, and biased toward candidates who used the exact phrasing the recruiter searched for.
Still ships in most legacy ATS systems as the default search. Most "search for candidates with X skill" features are this, dressed up with a nicer UI.
Generation 2: TF-IDF + keyword scoring (2010–2020)
Statistical relevance scoring. The system tokenizes the JD and each CV, computes term-frequency / inverse-document-frequency vectors, and ranks candidates by cosine similarity. Better than boolean — it surfaces candidates with related phrasing — but still keyword-anchored. A CV that says "led the move from EC2 to ECS" outranks one that says "migrated cloud infrastructure" even if the second is more relevant.
Most "AI screening" features in 2026-era ATS systems are still this generation, with a thin LLM "explanation" layer added for show.
Generation 3: Embedding similarity (2020–2024)
Modern semantic search. The JD and each CV get embedded into a high-dimensional vector via a model like OpenAI ada-002 or sentence-transformers, and ranked by vector similarity. This catches the "led the move from EC2 to ECS" vs "migrated cloud infrastructure" case — they embed near each other.
Big improvement, but still optimizes for similarity, not fitness. A senior architect's CV embeds near a junior dev's if they both list the same stack — the model has no concept of "this candidate has 8 years of evidence and this one has 18 months."
Generation 4: Evidence-based LLM evaluation (2024–today)
The current frontier. A large language model reads the structured intake (must-haves, nice-to-haves, soft signals) and reads each CV in full, then scores the candidate against each criterion with a citation back to the specific CV text that justified the score. CVPRO's STEP0 formula is one example: Skills 40% + Experience 25% + Domain 15% + Location 10% + Recency 10%, with 42 evidence points underpinning each score.
Two characteristics distinguish a real Generation 4 system from a Generation 2 system with an LLM bolted on: (1) it shows you the evidence span — the exact sentence on the CV that triggered the score — for every criterion, and (2) the score is reproducible: re-running the same CV against the same JD produces the same score within a tight margin.
How to test a vendor demo
Run these three checks during any AI screening demo:
- Test 1: Phrase variance. Submit two CVs that describe the same experience in different words ("led the EC2-to-ECS migration" vs "owned cloud infrastructure modernization at scale"). A real Gen 4 system scores them within 5 points of each other. A Gen 2 system scores them 20+ points apart.
- Test 2: Evidence citation. Click a candidate's score. The system must show you the exact CV sentences that justified each criterion. If it shows you a generic "the candidate has React experience" with no citation back to the CV, it is Gen 2 with an LLM wrapper.
- Test 3: Negative signals. Add a CV that has the right keywords but missing experience depth (e.g., a junior dev who listed every framework they touched in a bootcamp). A Gen 4 system flags this and ranks the candidate low. A Gen 2 system ranks them high because the keywords match.
What to do with the time AI screening saves
AI screening compresses 5 hours of CV review into 15 minutes. The time you save should not go into more recruiters or more requirements per recruiter — it should go into deeper conversations with the top 10 candidates per role. The recruiters who win in 2026 are the ones whose candidates feel "this person actually understood what I wanted to build" — not the ones who can shovel 50 candidates into the funnel.
Bias and audit considerations
AI screening is not automatically less biased. The screening model inherits whatever bias is in the historical hiring data and the JD itself. Two safeguards: (1) require evidence citations for every score so a human can audit the reasoning, and (2) test the screening output for adverse impact across protected groups (gender, age, regional language signals on Indian CVs) at least quarterly. A Gen 4 system that exposes evidence makes both checks possible. A Gen 2 system that returns black-box scores makes both impossible.