The Ugly Side of AI in HR (And Where It Should Stop)
HR stands for Human Resources. It's supposed to be the most human department in the building: the one that hires you, pays you, listens when something's wrong, and handles things with some dignity when it's time for you to go.
So it's a little strange that it's also the department racing hardest to hand things over to machines.
Here's the thing though: AI in HR isn't good or bad. It's a tool. Tools work great in some hands and turn into a problem in others. The real question was never whether HR should use AI. It's where the line sits between "this makes work better" and "this makes people feel like they're being watched."
Let's start with the part that's actually working.
The good: AI that shows up when it's needed
Unilever uses AI-powered games to assess candidate skills instead of just filtering resumes by keyword. It's a small shift with a big implication: hire for what someone can actually do, not for whether their resume survived a scanner. Other companies have built AI assistants that answer benefits questions at 3am, when the HR team is asleep and the employee's open enrollment deadline is very much not.
The results back it up. Companies using AI for personalized learning paths and early attrition warnings are seeing engagement scores climb from 3.2 to 4.1, and retention jump from 75% to 85%. Those aren't rounding errors. That's the difference between a team that sticks around and one that quietly starts updating LinkedIn.
This is AI doing what it's supposed to do: catching things humans would've missed, and being available at 3am so a person doesn't have to be.
The bad: when the machine gets it wrong
Then there's the part that isn't working.
Tools like HireVue, which analyze facial expressions and speech patterns during video interviews, penalize candidates who communicate differently than whatever pattern the model was trained to reward. Hiring bias lawsuits over exactly this are already in federal court. California regulated the practice in 2026, which should tell you how far past "early concern" this has already gotten.
This isn't a rounding error either. It's a system deciding who gets a second interview based on how closely someone's face and voice match a statistical average, then calling that objectivity.
The ugly: the layer nobody agreed to
Here's where it gets genuinely uncomfortable. Some companies are passively scanning employee Slack and email to gauge sentiment. Not with a survey. Not with a conversation. Quietly, in the background, and employees don't always know it's happening.
Think about what that actually replaces. It used to be that if a manager suspected someone was disengaged, checking out, or struggling with something, the move was to walk over and ask. Awkward, sure. Sometimes the answer was "I hate my manager." Sometimes it was "my kid's been in the hospital and I haven't told anyone." Either way, you found out by talking to a person.
Now a sentiment model can flag the same person as a risk based on word choice in a Slack thread, and nobody has to have the conversation at all.
Accountability is not the same as control
This is the line, and it's worth being precise about it.
Accountability needs to be direct, not indirect. When someone isn't engaged, the answer is to go talk to them, look them in the eye, and sit through the awkward moment long enough to find out what's actually going on. That's an uncomfortable skill. It's also the entire job of managing people, and it's the exact skill set that gets weaker every time a company outsources "checking in" to a dashboard.
The tools designed to create efficiency shouldn't be eroding the thing that makes an organization actually function: the ability to have hard conversations, be curious about disagreement, and show up for someone as a whole person. Used well, AI extends that capacity. A benefits assistant that answers a 3am question is giving someone their evening back. A hiring tool that screens for actual skill instead of resume keywords is giving a qualified candidate a shot they wouldn't have otherwise gotten.
Used badly, AI does the opposite. Sentiment analysis on Slack without a real conversation isn't management. It's surveillance wearing management's clothes. And you can't get to the bottom of an interpersonal problem without being interpersonal about it. There's no dashboard workaround for that.
HR should absolutely use the efficiency. Just spend it on the conversations, not instead of them. The best HR technology, AI included, should hand more trust and choice back to people. Not use it to monitor them more closely.
- Watch the full episode: The ClaudeCast breaks this down in detail, unscripted. Watch on YouTube
