Why AI Won't Replace Human Judgment.

"Everyone changes. Humans can be ten different people in a day. The un-algorithmic alacrity with which they wobble, my goodness."  - Ultron (VisionQuest via Marvel/Disney)

Short answer: AI won't replace human judgment, because good judgment depends on the one thing AI is built not to do, which is change with the moment.

Ultron's line comes from VisionQuest, the Disney+ series where Vision, an android, consults the AI personas living in his own programming to figure out who he is. Leave it to a synthetic being to spot what we keep missing about AI and ourselves.

You may have heard this one. In a 2011 study published in PNAS, researchers tracked more than 1,000 parole decisions by Israeli judges. At the start of a session, judges granted about 65% of requests. Right before their food break, that number slid toward zero.

After the break? Right back up to about 65%. (Critics have pointed out that case scheduling explains part of the effect, so don't tattoo the exact numbers on your arm. But the broader point has held up.)

Same judges. Same law. Same kinds of cases. Different people, depending on when they last ate.

We usually tell that story as a cautionary tale about human bias. And it is one. But it's also a pretty good description of what it means to be human. We change. Constantly.

A human can be ten different people in a day, and an AI is, more or less, the same thing every time you ask it. The pitch for AI is that it doesn't wobble. The real question is whether that's its biggest advantage or its biggest limit. I'd argue it's both, and understanding the difference is the whole game. 

Why do humans wobble?

Daniel Kahneman and his co-authors gave part of this a name in their book Noise: occasion noise. The same expert, looking at the same facts, reaches a different conclusion depending on mood, fatigue, the weather, or whatever happened in the meeting before. Doctors do it. Underwriters do it. You do it every time you answer an email differently at 8am than you would at 6pm.

That part is real, and it's worth fixing where consistency matters. But lumping all human variability into "noise" misses something big. A lot of the wobble isn't random. It's adaptation.

Humans are built to change state in response to what's in front of them. We get sharper when the stakes go up, softer when someone's struggling, funnier when the room needs it, and quieter when it doesn't. The focused 9am version of you, the frazzled one after four back-to-back calls, and the generous one after lunch aren't malfunctions. They're one person running different settings for different situations, usually without even noticing the switch. 

What does AI do instead?

AI works the other way around. Give a model the same input and you get roughly the same output. It doesn't get hangry, it doesn't have a bad Tuesday, and it doesn't carry yesterday's argument into today's conversation. For a lot of work, that's a superpower.

But notice how AI actually changes. It adapts to the prompt, not to the room. Its deeper shifts happen in releases: a new model, new training data, a version number that ticks up every few months. Between those updates, it's essentially the same entity, answering the millionth question with the same disposition it brought to the first.

Humans, meanwhile, shift states with the kind of un-algorithmic alacrity VisionQuest was talking about, and no actual algorithm can match it. That's the irony. We built machines to be predictable and then started worrying they'd replace the least predictable thing on the planet. The wobble is exactly what makes that replacement so hard. 

What do the ten people actually do?

Think about a recruiter on a screening call. The candidate stumbles through the first two questions, then lights up when the conversation turns to a project they led. A good recruiter clocks the difference between nerves and a skills gap in about ninety seconds, then shifts gears: slows down, warms up, asks a better question. That's one person becoming a different interviewer mid-call because the moment demanded it.

Or take the manager who walks into a team meeting planning to push hard on a missed deadline, sees three exhausted faces, and decides the push can wait until Thursday. Was that inconsistent? Absolutely. Was it the right call? Probably. (Ask anyone who's ever worked for a manager who pushed anyway.)

None of that comes from a rulebook. It comes from a person changing who they are, on the fly, based on what's happening right now. This is the stuff that never shows up in a job description, and it's exactly what people remember about the best colleagues they ever had.

Nobody tells a story about the boss who responded identically to every situation. They tell stories about the one who knew when to ease off, when to lean in, and when to close the laptop and ask how you were really doing. 

Won't AI just learn to read the room?

That's the obvious pushback, and vendors are certainly trying. Emotion AI tools promise to read sentiment from your voice, your typing patterns, even your face on a video call. And sure, a model can learn that a shaky voice often signals stress.

But spotting a signal isn't the same as understanding a person. The manager who eases off isn't running a sentiment score. She's drawing on six months of history with her team, a hunch about what's going on at home for one of them, and a feel for what this group needs this week.

More importantly, she owns the call. If she gets it wrong, she owns it, repairs it, and earns back trust. A model can't do the repair part, because nobody trusts a system the way they trust a person who showed up for them.

Regulators seem to agree that something here is off-limits. The EU AI Act flat-out bans emotion recognition systems in the workplace, with narrow exceptions for medical and safety reasons. When lawmakers in Brussels and HR leaders agree, it's worth paying attention.

That's the real ceiling. AI will get faster, cheaper, and more consistent every year. What it can't do is have a stake in the outcome, and in work that involves people, having a stake in the outcome is most of the job. 

So where does each one fit?

Once you see the distinction, the division of labor becomes much clearer. Some work rewards sameness. Answering the same benefits question for the four hundredth time, checking every application against the same criteria, applying a policy the same way in Ohio as in Oregon. That's consistency work, and AI does it better than we do, mostly because it never has a 4:45 pm on a Friday.

Other work rewards the wobble. Negotiating an offer with a candidate who's clearly on the fence. Telling someone their role is changing. Coaching a new manager through their first hard conversation. That's adaptation work, and it depends on a person who can become whoever the moment needs. 

The smartest use of AI sits right at that border. It can even help keep the human side honest by flagging when decisions drift with the clock or the calendar. Think less robot replacement and more air traffic control. The machine watches the patterns. The humans still land the plane. 

How should HR leaders split work between AI and human judgment?

Start by sorting the work. For any task you're thinking of handing to AI, ask one simple question: does this need to come out the same every time, or does it need to change with the situation? If it's the first, automate it without guilt. If it's the second, keep a human in the seat, and give that human the time and context to actually wobble well.

Then protect the human parts of the job on purpose. If AI takes the admin load off your managers (and it should), don't backfill that time with more meetings. Spend it on the conversations a model can't have: the career chat, the hard feedback delivered well, the check-in that isn't really about the project.

And please stop trying to make people more machine-like. Rigid scripts, one-size-fits-all playbooks, and dashboards that punish any deviation all chip away at the one capability AI can't copy. If you want a workforce that's hard to replace, build one that can adapt.

Walt Whitman said he contained multitudes. So does everyone on your team, and that's the one thing no model can replicate. AI can be consistent. It can't be ten people. The future isn't about which one wins. It's about building organizations smart enough to know which one each moment needs.

Frequently asked questions

Will AI replace human judgment in HR?

No. AI is built for work that needs the same answer every time. Judgment calls that depend on context, history, and owning the outcome still need a person.

What is occasion noise?

It's a term from Kahneman, Sibony, and Sunstein's book Noise. It describes the same expert reaching different conclusions on the same facts depending on mood, fatigue, or timing. 

Is emotion recognition AI legal in the workplace?

Not in the EU. The EU AI Act bans AI systems that infer people's emotions at work, with narrow exceptions for medical or safety reasons.

How do I decide which HR tasks to hand to AI?

Ask whether the task needs to come out the same every time or needs to change with the situation. Automate the first, and keep a human in the seat for the second.