Can We Use AI to Screen Job Applicants? The Law and the Evidence

Can we use AI to screen job applicants?

You can use AI to screen job applicants with a line drawn in the middle of the process. Everything that moves an application through your pipeline can be automated. Deciding who advances should stay with a person, and in several states that is now closer to a requirement than a preference.

The distinction that matters is between handling and judging. Reading a CV into structured fields is handling. Confirming receipt, asking three qualifying questions, checking a licence number, offering interview slots: handling. Scoring candidates against each other, ranking them, filtering anybody out automatically: judging.

Automating the handling half is where the time is anyway. For a business hiring a handful of people a year, most of the hours go into acknowledgements, scheduling, chasing and the awkward silence that makes good candidates accept somewhere else. None of that involves a decision about a person's suitability, and none of it carries the legal exposure.

The judging half is where both the law and the evidence have arrived. Several states and one city now regulate it directly, the regulators treat a vendor's tool as the employer's responsibility, and the research says the tools carry bias the employer cannot see.

There is one more reason to hold the line. A hiring decision you cannot explain is a hiring decision you cannot defend, to a rejected candidate, to a regulator, or to yourself in six months when the hire is not working. A ranked list from a system nobody can interrogate is not an explanation.

What does the research say about bias in AI resume screening?

The research on bias in AI resume screening found severe disparities, in a setup designed so that nothing but the name differed.

Kyra Wilson and Aylin Caliskan of the University of Washington audited three open-source retrieval models, the kind that match resumes to job descriptions, using 554 resumes and 571 job descriptions and generating more than three million comparisons. Their paper, Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval, was presented at the AAAI and ACM Conference on AI, Ethics and Society in October 2024. Resumes with white-associated names were favoured in 85.1% of cases. Female-associated names were favoured in 11.1%. Black male candidates fared worst of all, with the models preferring other candidates close to 100% of the time across the combinations tested.

Read the limits with the result. The researchers tested open-source models rather than commercial hiring products, because commercial products are closed and cannot be audited from outside. That is a caveat about this study and it is also the problem: the tools being sold to employers cannot be independently checked, and the ones that can be checked behave like this.

The mechanism is not mysterious. A model trained on text produced by a society that hires unequally will reproduce the pattern, and a name is enough of a signal. Nothing in a screener's output flags that this has happened, which is what separates this failure from an obvious one.

The practical consequence for an employer is specific. If a tool ranks candidates and you follow the ranking, you have adopted whatever pattern sits inside it, and under United States employment law the outcome is your responsibility rather than the vendor's. California's regulations make that point explicitly by reaching vendors and tool developers who screen applicants on an employer's behalf.

Which hiring laws apply when we use AI to screen candidates?

The hiring laws that apply when you use AI to screen candidates come down to four United States rules in 2026, and the trigger for all of them is using a tool to assist a decision about a person rather than the size of your business.

Rule	Who it covers	What it requires	In force
New York City Local Law 144	Employers using an automated employment decision tool for jobs in NYC	Independent bias audit within the past year, a public summary of the results, and notice to candidates at least 10 business days before use	Enforced since 5 July 2023
Illinois HB 3773, amending the Human Rights Act (Public Act 103-0804)	Employers with employees in Illinois	No use of AI with a discriminatory effect in recruitment, hiring, promotion, discipline or discharge; no zip code as a proxy for a protected class; applicants and employees must be notified when AI is used	1 January 2026
California Civil Rights Council regulations under FEHA	California employers with five or more employees in the state	Automated decision system data kept four years; anti-bias testing is relevant to a discrimination defence; vendors and developers that screen on an employer's behalf are reached as agents	1 October 2025
Colorado	Colorado employers	The original AI Act was stayed and then replaced by a narrower act built on notice and human review	Replacement effective 1 January 2027

Each of those rows has a primary source worth reading before you rely on a vendor's summary of it: the Department of Consumer and Worker Protection's own FAQ for New York City, the Illinois amendment as HB 3773, Public Act 103-0804, and the Civil Rights Council's rulemaking record for California.

One wrinkle on Illinois: the statutory duty is live, and the detail is not. The Illinois Department of Human Rights proposed the rules defining what a compliant notice looks like on 15 May 2026 and withdrew them on 2 June 2026. Give the notice anyway and keep a record of what you sent.

Two things sit underneath the table. Existing discrimination law applies to an automated decision exactly as it applies to a human one, which is the part that covers you everywhere else in the country. And a vendor's assurance is not a defence: the employer makes the decision, so the employer answers for it.

If you hire inside the European Union, employment screening sits in the high-risk category of the EU AI Act, and the timetable for those obligations has moved more than once. Get the current date from counsel rather than from a vendor's compliance page.

The wider map, including Colorado's replacement act and the federal position, is in does a small business need an AI policy. Staffing firms screening on behalf of clients carry a different set of obligations, set out on AI for staffing agencies.

Which parts of hiring are safe to automate?

Six parts of hiring are safe to automate, and together they cover most of the clock time between an application arriving and an offer going out.

Acknowledgement and status updates. Every applicant hears back within minutes, and again when something changes. This is the cheapest reputational win in hiring and most businesses still fail it.

Structured capture. Pulling a CV into consistent fields so a human reviewer compares like with like. Note what this is not: scoring. Presenting the same six fields for every candidate is the opposite of a black box.

Knockout questions you would ask anyway, as long as they are objective and job-related. A licence, a certification, legal authorisation to work, availability for the shift pattern. Write them down, apply them identically, and keep a record. Anything that is a judgment is not a knockout question.

Interview scheduling, including rescheduling and reminders. Pure calendar work, and it removes days of waiting from a pipeline.

Interview preparation. Generating a structured interview guide from the job description, so every candidate gets the same questions, which is also the strongest defence against inconsistent hiring that exists.

Reference and document chasing. Sequences, not decisions.

What is left is the part a person does: reading the applications that pass the objective filters, interviewing, and choosing. For a business hiring ten people a year, automating the six above gives the hiring manager their week back without putting a model anywhere near a judgment. The same shape applies across the trades, where speed of response decides who you even get to interview, and those capabilities sit on the home and trade services board.

What should we never let AI decide in hiring?

Never let AI decide five things in hiring, and each one is a decision rather than a task.

Who gets rejected. An automatic rejection based on a model's ranking is the single highest risk action in this category, and in New York City it triggers the audit and notice obligations outright.

Anything scored from a video or a voice. Facial expression, tone and speech-pattern scoring have the weakest evidence base and the strongest regulatory attention, and they disadvantage candidates with disabilities in ways that are hard to detect and hard to defend.

Personality or culture inference. A model guessing at traits from a document is producing an opinion with a number attached, and the number makes it look like a measurement.

Anything that uses a proxy for a protected class. Illinois names zip code explicitly. Others are just as effective and less obvious: the name of a school, a gap in employment, membership of an association.

Anything where you cannot say why. If the answer to "why was this candidate ranked below that one" is that the tool said so, you have no answer at all.

One operational rule sits underneath all five: a human reviews every rejection, or at minimum a sample of rejections, and that review is recorded. California requires automated decision system records to be kept for four years, and the records are what turn a good process into a defensible one. The general pattern for designing a system so its mistakes get caught is in what happens when AI gets it wrong.

How do we use AI in hiring without creating a legal problem?

Use AI in hiring without creating a legal problem by working through six steps in this order, most of which are documentation rather than technology.

Write down what the tool does. Which stage it touches, what it produces, and whether its output influences a decision. A tool that writes job descriptions is not a tool that screens applicants, and conflating them in your own head is how notice obligations get missed.

Tell candidates. Illinois requires notification where AI is used in these decisions, and New York City requires notice at least ten business days before a covered tool is used, with information about the characteristics it assesses. Notice costs nothing and it is the first thing a regulator asks for.

Keep the human decision real. Not a rubber stamp on a ranked list. A reviewer who sees the application rather than only the score, and who can and sometimes does disagree with the tool.

Keep records. Applications, outputs, decisions, reviewers, dates. California sets four years as the floor for automated decision system data, and keeping them everywhere is simpler than remembering where you do not have to.

Test for disparity yourself. Compare pass rates at each stage by the categories you are allowed to analyse. New York City requires an independent audit for covered tools, and California treats anti-bias testing as relevant to your defence. Running the numbers twice a year is cheap, and the absence of them is conspicuous.

Put it in the policy. Who can use which tool at which stage, who reviews, how long records are kept. If you do not have one yet, does a small business need an AI policy covers what it should say.

Where BLACKSIG builds hiring workflow, the acknowledgements, the structured capture, the scheduling and the record keeping run on our infrastructure and we operate them from there, with the screening decisions, the interviews and the sign-off staying with your people. The systems we build do not rank candidates.

What should we ask an AI hiring vendor before switching it on?

Ask an AI hiring vendor seven questions before switching it on, and a vendor who cannot answer the first three in writing has answered all seven.

Does your tool score, rank or filter candidates, or does it only organise information? The answer determines which laws apply, and most product pages are deliberately vague about it.

Has it had an independent bias audit, who performed it, when, and can we see the summary? For a tool used in New York City this is a legal requirement rather than a courtesy.

What happens to candidate data: where is it stored, how long for, is it used to improve your models, and can we delete it on request?

Will you support us if a candidate or a regulator challenges a decision, and is that in the contract?

Can we turn off the scoring and keep the rest? Often the answer is yes and nobody asks.

What does the tool do when it has low confidence: flag for review, or guess silently? The honest answer separates two categories of product.

Can we export our data if we leave? Hiring records you are legally required to keep should not be hostage to a subscription. One final check that is not a question: before trusting a ranking, take twenty applications you have already decided on, put them through the tool, and compare. If it disagrees with your judgment in ways you cannot explain, that is the finding, and it arrived cheaply. The vendor evaluation checklist for any system touching your data is in is AI automation safe with my business data.

Frequently asked questions

How much does it cost to automate our hiring process?

What moves the number is how many roles you hire for, whether applications arrive through one job board or six, whether you run an applicant tracking system or a shared inbox, how much of the scheduling involves other people's calendars, and what records you need to keep. A business hiring six people a year from one job board is a different job from a multi-site operator hiring continuously across three states with different notice rules. BLACKSIG does not publish a rate card; scope is set on the call once the strategy work has established what is worth doing. The drivers are in how much AI automation costs.

How does AI screening actually work?

Most tools convert a CV into structured fields, compare those fields against the job description, and produce a score or a ranking. Some add questionnaire responses or assessment results. The scoring step is a model's judgment about similarity, not a measurement of capability, which is why two tools given the same candidates produce different orders.

Is AI screening better than using a recruiter?

Different costs and different failure modes. A recruiter brings market knowledge, candidate sourcing and judgment, and charges a fee per hire. A tool processes volume instantly and brings whatever bias it learned. For roles with hundreds of applicants, the strongest setup is automated handling plus objective knockouts plus human review, and a recruiter for the roles where sourcing rather than filtering is the hard part.

Can our hiring be automated if we only hire a few people a year?

The administration can, and the screening is not worth touching. At low volume the time goes into acknowledgements, scheduling and the follow-ups that fall through, and automating those improves your candidate experience immediately. Scoring a handful of applicants with a model adds risk and saves minutes.

Do we have to tell candidates we are using AI?

In Illinois and New York City, yes, with New York City requiring at least ten business days' notice before a covered tool is used, and other states moving in the same direction. Tell them everywhere regardless: it costs nothing, it is the first thing a regulator asks for, and candidates who find out later tell other candidates.

Can We Use AI to Screen Job Applicants? What the Law and the Evidence Say

The short version