How AI-Written Job Descriptions Shrink Your Candidate Pool

If you ask an AI tool to update a job description, it will rarely make it shorter. It works from what you gave it and elaborates. A "nice to have" in the old version comes back as a requirement. A line that said "communication skills" comes back as "demonstrated ability to communicate complex technical concepts to C-suite stakeholders in written and verbal formats." Each change looks reasonable on its own. Put together, they make a longer list of requirements that says less about what the job needs.

The writing isn't bad, and that's part of the problem. It sounds confident and complete, so nobody reads it critically before it goes live. A clumsy draft invites edits. A polished one gets posted.

The people hurt by this are mostly not the experienced candidates, who learned long ago to treat requirement lists as wish lists and apply anyway. The ones who take the list at face value are the ones with less practice at job hunting: people early in their careers, people changing careers, people coming back to work after a gap. They read eighteen requirements, notice they meet twelve, and don't apply. You never see them, so you never know you lost them.

The filtering isn't random, either. One often-quoted finding is that women tend to apply for a job only when they meet about 100% of the listed requirements, while men apply when they meet about 60%. If that's even roughly right, then every requirement you add pushes away some groups of qualified people more than others. You've built a filter you didn't intend, and your ATS will never flag it, because everyone who did apply looks fine.

There's a slower version of the same problem. Suppose a head of engineering once wanted a cloud certification for a role because of a particular technical decision three years ago. That requirement goes into the job description. The next time the role opens, someone gives the old description to an AI tool, and the certification comes through again. The role changes over the years but the requirement stays, because the AI only knows the last document it was shown.

This is why I'm skeptical of a lot of talk about talent shortages. In some specialized fields the shortage is real. But I suspect a fair number of "we can't find anyone" searches are really job description problems. The people exist. They read the posting and decided it wasn't for them.

It helps to ask what a job description is for. A strong candidate reading one is asking two questions: is this worth my time, and can I actually do it? Listing every skill a useful employee might have doesn't help answer either. A description that answers them honestly will usually get fewer applicants, and better ones.

Answering them honestly means separating what's truly required from what would merely be nice. It means writing about the actual work instead of a generic version of the role. And it means being straight about level. A "senior" title with a junior salary attracts people who either didn't read carefully or have no other options, and you don't want to hire either group.

An AI tool won't do this for you. Suppose most experienced people in your market have about three years of Salesforce experience, and the draft asks for five. The tool won't point out that you've just ruled out most of the market. It doesn't know. That judgment has to come from someone who knows what the role really needs.

The damage doesn't stop at the posting. The candidates who do apply are often screened against the same inflated list, which penalizes anyone who misses a box. When the description was written by AI and the screening is automated against its criteria, you get a loop in which one unreviewed document decides who gets considered, and no human judgment enters at any point.

That loop is one of the things we think about most at Asendia. We make voice AI that phones each applicant within hours of applying and interviews them, and this problem is a large part of why it talks to people instead of sending them a form. A form can only ask whether you hold the cloud certification. In a conversation, when someone says no, the next question can be what they've built instead, and that answer usually tells you more than the checkbox would have. The screening criteria are set by your team based on what the role needs, and they don't have to match what the posting says. A hiring manager can mark the certification as preferred rather than required, and the conversation will then ask about equivalent experience instead of rejecting someone who has the skill but not the certificate.

What goes into your ATS afterward is a qualification summary with quotes from the candidate, not a pass or fail on eighteen requirements, so a recruiter can spot the person who misses the formal checklist but clearly has the ability. Agencies running large campaigns use it so that automated screening doesn't repeat a bad job description's logic across hundreds of applicants. If you've wondered whether your ATS was ever designed for this kind of judgment, there's a separate post on what an ATS is actually built for.

Some shortages are real. Others are made at the keyboard, by descriptions that inflate requirements, carry forward old criteria, and substitute template logic for judgment. The fix is a person who knows the role going through the list and deleting everything that's aspirational, which is the one step a better AI writing tool can't do for you. That might take twenty minutes per role, and I'd guess it will do more for the quality of your pipeline than most of what you're doing further down the funnel.

Ready to transform your hiring strategy? Schedule a Demo with our founders today!