When Two AIs Are Talking to Each Other, Nobody Is Actually Hiring.
Recruiting can be tough, and that was before AI. Small teams with no recruiter are turning on whatever screening tool came with their ATS because it saves time. Candidates have since figured out that AI is reading their resumes before any human does so they're using AI to write resumes that match job descriptions as closely as possible. What you end up with is one AI generating the resume and another one screening it, and the humans don't enter the conversation until the field has already been narrowed by an algorithm nobody really set up intentionally.
This is a problem for a few reasons, and it's a bigger legal problem than most small employers realize.
The legal piece
Mobley v. Workday is the case to know. Derek Mobley, a Black man over 40 with a disability, applied to more than 100 jobs through employers using Workday's AI screening tools and was rejected from every single one, often within minutes. The case is now a class action covering every applicant over 40 screened out by Workday's tools since September 2020, with race and disability claims included.
The courts have confirmed that the employer using the tool carries liability, not just the vendor that built it. Workday's argument that age discrimination law doesn't cover job applicants was rejected. Employers using Workday's HiredScore AI screening features have been pulled into the litigation for using that exact tool in their own recruiting, even though they didn’t build it and have no say in its algorithm.
What AI actually optimizes for
AI screening tools match what's in your job description to what's on a resume. That sounds useful, and it can be, but it means the tool is optimizing for familiarity. Candidates whose careers look exactly like the role get surfaced (or candidates who revise their resumes to look exactly like the job description). Candidates with non-linear paths, career gaps, unconventional backgrounds, or experiences that don't map to your specific keywords get filtered out.
At a small company, those are often your most interesting candidates. Someone who's worn a lot of hats, pivoted industries, or built something scrappy tends to thrive in a 20-person company in a way that a perfect resume match sometimes doesn't.
What to do about it
Don't let the tool reject anyone automatically. With the exception of hard legal requirements, like needing a work permit or a specific license or credential, no candidate should get an automated rejection before a human has looked at them. Sorting and ranking candidates is fine. But a tool making a final no on your behalf, without any human review, is where the legal exposure lives and where you lose candidates worth a second look. Someone should be putting eyes on every application, even briefly.
Add screening questions. This is the most practical thing a small team can do without a recruiter. Screening questions are short, specific questions every applicant answers upfront, before any AI tool touches their application. They work because even if a candidate uses AI to help draft an answer, they have to actually engage with the question. The quality of the answer tells you how someone thinks, not just whether their resume has the right words. And if someone doesn't answer, or gives a generic non-answer, that's useful information too.
Good screening questions are specific to the role and necessitate reflection. A few examples:
For an operations role: "Describe a process you inherited that wasn't working. What did you change and why?"
For a sales role: "Tell us about a deal you lost. What would you do differently?"
For a generalist or startup hire: "What's something you taught yourself in the last year that changed how you work?"
For any role: "How do you do your best work?" or “What is it about this specific role that has you excited?”
Know what your tools are actually doing. A lot of small teams have AI screening active and don't know it, because it came default with their ATS or was turned on during setup. Log into your ATS, look at how candidates are being ranked or filtered, and check whether any automated recommendations are running before a human sees an application. If you can't find the answer in the settings, email your vendor and ask directly. You are entitled to know what criteria your tool is using and what data it was trained on.
Check who you're filtering out, not just who you're surfacing. Pull your rejection data occasionally and look at it. Are you consistently screening out candidates with certain types of backgrounds? Career changers? People with gaps? Liability lies in these kinds of details.
Write job descriptions for the person, not the algorithm. If your job description is a list of keywords designed to feed a screening tool, you're narrowing your own pool before you've seen a single application. Write for the human you actually want to hire. Be honest about what's required versus what's a nice-to-have. That alone will change the quality of who applies.
The bottom line
Employers need to know what their tools are doing, add a layer that requires actual human thought, and make sure they are the one making the decision, not the algorithm. The Mobley case is still moving through the courts, but the direction is clear: AI in hiring is not a neutral administrative function, and the employer is responsible for who it rejects.
