AI screening is the use of software to assess job applicants before a human looks at them: reading CVs, scoring answers, ranking candidates or filtering some out. For a company hiring at speed, it is tempting for an obvious reason. A popular role can attract more applications in a week than a recruiter can read in a month.
AI screening is also the part of hiring that regulators watch most closely. This guide explains how AI candidate screening works, the three main types, where they go wrong, the rules that apply in 2026 and 2027, and how to use AI screening at volume without losing the ability to explain your decisions. One disclosure: Cohesyve, the company publishing this, sells a screening product based on job simulations. It appears in the section on types.
What is AI screening?
AI screening is any automated step that scores, ranks, recommends or filters candidates as part of deciding who moves forward. It ranges from a résumé parser that ranks applicants against a job description to a chatbot that asks qualifying questions, to a system that scores a candidate's work on a task.
The legal definitions are broad. New York City's Local Law 144 covers automated employment decision tools. The EU AI Act lists "AI systems intended to be used for the recruitment or selection of natural persons" as high-risk. If a tool influences who advances, assume it is in scope.
How does AI screening work?
Most tools follow the same basic flow.
- Input. The tool takes the job description and the candidate's material: a CV, answers to questions, a recorded response or a completed task.
- Extraction. It turns that material into structured data: skills, job titles, years of experience, answers, or features of the work.
- Comparison. It compares the data with a target: the job requirements, a rubric, or patterns from past hires.
- Output. It produces a score, a rank, a recommendation or a pass or fail.
- Action. A recruiter reviews the output, or the system moves candidates automatically.
The risk sits mainly in steps 3 and 5. If the target is "people like our past hires", the tool learns whatever patterns were in past hiring. If step 5 has no human in it, the tool is making the decision.
Screen high volumes on the work itself, then decide yourself
Give Cohesyve a job description. It sources candidates, tests them on AI simulations of the real work and returns a ranked shortlist your team reviews before any interview.
The three types of AI screening
| Type | What it does | Strength | Main risk |
|---|---|---|---|
| CV parsing and matching (AI resume screening) | Reads CVs and ranks them against the job description or past hires | Fast, cheap, no candidate effort | Rewards keywords and history; can copy past bias; hard to explain |
| Question and chatbot screening | Asks knock-out and qualifying questions, sometimes scores free-text or video answers | Good for eligibility and scheduling | Scoring open answers can be opaque; video analysis attracts extra scrutiny |
| Work-sample and simulation screening | Scores the candidate's work on a task that resembles the job | Direct evidence of ability; easier to explain | Asks candidates for time, so tasks must be relevant and short |
Fixed-library test platforms and newer tools that generate role-specific tasks both sit in the third group. Cohesyve is one of the latter: it sources candidates, tests them on AI job simulations of the real work and returns a ranked shortlist for a person to review.
What is AI resume screening, and why is it risky?
AI resume screening ranks or filters CVs automatically. It is the most common type and the one with the weakest link to job performance, because a CV shows how someone describes their past, not what they can do now.
Three problems recur:
- Inherited bias. A model trained on who was hired before learns those patterns, including ones related to gender, age, background or education, even when those fields are hidden.
- Keyword gaming. Candidates learn to mirror the job description. Applicants who write plain CVs fall down the list.
- No explanation. When a candidate or regulator asks why someone was rejected, "the model scored them lower" is not an answer you can defend.
If you use CV ranking, use it to sort the queue for human review, never to reject.
What are the laws on AI screening?
This is a summary, not legal advice. Rules depend on where you hire and where candidates live, and they change. Check with counsel before you deploy a tool that influences hiring decisions.
| Jurisdiction | Rule | Applies from | Key duties |
|---|---|---|---|
| New York City | Local Law 144 of 2021 | Enforced since 5 July 2023 | Bias audit within one year of use; public information about the audit; notice to candidates (DCWP material says 10 business days before use) |
| Colorado | SB 26-189 (repealed and replaced the 2024 AI Act) | 1 January 2027 | Notice at the point of interaction; a plain-language description of the tool's role within 30 days of an adverse outcome; data access and correction; a right to request human review and reconsideration; records kept at least three years |
| European Union | AI Act, Annex III point 4(a): recruitment or selection | 2 December 2027 (moved from August 2026 by Regulation (EU) 2026/1744) | High-risk obligations, including risk management, data governance, human oversight and transparency |
In Colorado, the Attorney General enforces the law and the act creates no new private right of action. Our guide to AI hiring laws in 2026 covers these and other jurisdictions in more depth.
The common thread: tell candidates when AI is used, be able to explain the basis of an outcome, check for disparate impact, and keep a person able to review decisions.
How to use AI screening at volume, safely
A practical playbook for a company screening thousands of applicants.
1. Name the bottleneck
Pull stage-to-stage times from your applicant tracking system. If the slow step is scheduling, buy scheduling. If it is reading applications, screening is the right category.
2. Prefer evidence of the work
Screen on something the candidate does that resembles the job: a short task, a work sample or a job simulation. It predicts performance more directly than CV history and it is easier to explain. See what a work sample test is.
3. Keep a human on every rejection
Let the tool sort and score; let a person decide who is rejected. It is better practice and increasingly a legal expectation.
4. Tell candidates
Say in the job post or application flow that AI is used, what it does and how to ask for a human review or an accommodation. In New York City and Colorado, notice is a legal requirement for covered tools.
5. Audit before and after
Get the vendor's bias audit before you sign. After launch, compare pass rates across groups at each stage on your own data, and investigate gaps.
6. Write down the rubric
Whatever the tool scores, you should be able to say in plain words what it rewards. If you cannot, you will struggle to meet a 30-day explanation duty.
A worked example (illustrative)
The figures are invented to show the flow.
A company receives 3,000 applications for 40 customer support roles. Instead of ranking CVs, it asks eligibility questions (location, hours, right to work), then invites every eligible applicant to a short role task: answering three realistic customer tickets. Responses are scored against a published rubric. A recruiter reviews the top group and a random sample of lower scores each week to check the scoring. Interviews go only to candidates who have shown they can handle the tickets. Every rejection is reviewed by a person, and candidates are told AI is used and how to request a review.
Questions to ask an AI screening vendor
- What exactly does the tool score, and on what basis?
- Was it trained on past hiring data? Whose, and how was it checked for bias?
- Can you share an independent bias audit dated within the last year?
- What notice wording do you provide for candidates?
- How does a person review or overturn an outcome?
- What do you do when the tool is not confident?
- How will you meet the EU high-risk obligations by December 2027 and Colorado's rules from January 2027?
- How do we export our candidate data and audit trail?
Our AI recruitment tools guide and resume screening software guide cover vendor selection more broadly.
Frequently asked questions
Is AI screening legal?
Yes, in most places, but it is regulated. New York City requires a bias audit and notice for covered tools now; Colorado's rules apply from 1 January 2027; EU high-risk obligations for recruitment AI apply from 2 December 2027. Take legal advice for your locations.
Does AI screening reduce bias?
It can if it screens on evidence of the work against a clear rubric and is audited. It can make bias worse if it ranks CVs against patterns from past hiring.
Can AI reject candidates automatically?
Technically yes, but it is risky. Keep a person on rejections, and in Colorado be ready to describe the tool's role and offer human review after an adverse outcome.
What is the difference between AI screening and an ATS?
An applicant tracking system stores and moves candidates through stages. AI screening is a step that scores or ranks them. Many ATSs now include AI screening features, which bring the same legal duties.
How should candidates be told about AI screening?
Clearly and before the tool is used: what it assesses, how it is used in the decision, and how to request a human review or adjustments. New York City's rules set specific notice requirements for covered tools.