Will AI replace HR? What the court and labor record shows (2026)

If you work in HR, you have read the forecast. The screener reads the CVs, the chatbot answers the policy questions, the model drafts the review, and the department shrinks to a dashboard.

There is a record to test that against. It is short, and most of it comes from courts.

Start with the one measurement that names HR professionals. A Danish study linked chatbot surveys to administrative registers and found “precise null effects on earnings and recorded hours,” ruling out effects larger than 2% two years after ChatGPT [source].

What “will AI replace HR” actually asks

Will AI replace HR asks whether software can take over the decisions an HR function is accountable for, not just the paperwork around them.

Those are two different questions. AI clearly does some of the paperwork. The harder question is who answers when the screen rejects the wrong person or the system fires someone.

This post uses four TIN case files: one measurement of AI at work, and three records of software making an employment decision and an authority asking who was responsible.

Will AI replace HR jobs?

No, not on the measured record so far. The only large study that measured HR professionals after ChatGPT found no movement in pay or hours, and the official US projection still has the occupation growing.

The study is NBER Working Paper 33777 by Anders Humlum and Emilie Vestergaard. Its surveys cover 11 exposed occupations, and “HR professionals” is one of them [source].

The latest round “includes responses from 25,000 workers across 7,000 workplaces,” linked to Statistics Denmark records of earnings and hours [source]. The first version concluded that chatbots “have had no significant impact on earnings or recorded hours in any occupation” ([source]).

Read the limits. It is a working paper, not peer-reviewed, and it covers the first two years of adoption. It tells you what has happened, not what will.

What does the job outlook for HR specialists say?

The official US outlook says the occupation is growing. The Bureau of Labor Statistics counts 939,700 HR specialist jobs in 2025 [source].

It projects employment to “grow 6 percent from 2025 to 2035, faster than the average for all occupations,” with “about 73,700 openings” a year [source].

That is the government’s forecast, and forecasts can be wrong. It is cited here as a projection, not as something TIN checked.

What happens when software makes the hiring call?

When software makes the hiring call, the employer still owns the result. That is the consistent line from the US regulator, and it is the opposite of the replacement story.

In EEOC v. iTutorGroup, the agency alleged the company “programmed their tutor application software to automatically reject female applicants age 55 or older and male applicants age 60 or older” [source]. It said more than 200 qualified older applicants were turned away [source].

The company denied the allegations and settled for $365,000 under a five-year consent decree [source]. The Chair’s line is the one HR should keep: “Even when technology automates the discrimination, the employer is still responsible” ([source]).

A caveat the headlines skip: the pleadings describe a programmed age rule, not machine learning. The AI label came from the agency’s initiative, not the complaint. That makes the point sharper, not weaker: the simplest automation already moved liability onto the employer.

Can an HR software vendor be sued for an AI screen?

Yes, a federal court has let that claim proceed. In Mobley v. Workday, Judge Rita Lin granted preliminary certification of a nationwide age-discrimination collective on 16 May 2025 [source].

The collective covers applicants “aged 40 and over who, from September 24, 2020, through the present” were denied employment recommendations through Workday’s platform [source].

Read it for what it is. It is a procedural ruling, not a finding that Workday discriminated, and Workday denies wrongdoing [source]. What it settles is narrower: buying the screen does not hand the question of fairness to the vendor and walk away.

What does “human in the loop” have to mean?

A human in the loop has to be able to change the decision, or the law treats the decision as made by the machine. The Amsterdam Court of Appeal said so about Uber’s driver deactivations in April 2023.

The court held that oversight must be meaningful, “not merely a symbolic act,” carried out by “someone competent and authorised to change the decision” (TIN’s translation of the Dutch ruling) [source].

It ordered Uber to explain the logic behind the deactivations under a penalty of 4,000 euros a day [source]. TechCrunch reported that Uber “could not demonstrate the staff involved had been able to exercise a meaningful check” ([source]).

Drivers are not employees in every jurisdiction, and this was a data-rights case, not an employment tribunal. It still describes the HR job exactly: the person who can overrule the system, and can show they did.

The proof

TIN’s case file on the Danish study quotes both versions of the paper and prints where they differ: 25,000 workers in 11 occupations, no earnings effect.

TIN’s case file on Mobley v. Workday rests on the court order itself, and states that the ruling is procedural: a nationwide age-discrimination collective against an AI hiring tool.

TIN’s case file on iTutorGroup checks the “first AI hiring settlement” label against the pleadings and prints where it overreaches: $365,000 over automated age screening.

TIN’s case file on the Amsterdam ruling quotes the court in Dutch with translations: automated firing and the symbolic human check.

The wider pattern of enforced hiring cases is set out in AI hiring discrimination: what the enforced record shows. None of these files counts HR jobs lost to AI, and this post claims no such number.

What each record shows

The bottom line

The forecast that AI replaces HR treats the function as processing: screening, answering, filing. The record treats it as accountability. A regulator, a federal judge and an appeal court each looked at software making an employment decision and asked the same thing: which person answers for it. Automate the queue, and staff the override with someone who can change the outcome and prove they looked.

That holds for any function where a machine’s decision lands on a person.