The AI Said No. You Pay.

Your AI tool doesn't get sued. You do.

That's the whole message of a 30-second video Michael Blakley shared on LinkedIn this week, made by Screenloop. It's short, and it says something a lot of hiring teams still don't want to hear.

What the Video Says

The video calls itself a "hiring risk briefing". It opens with two lines: "Your AI rejects the candidate. You face the claim."

When AI gets it wrong, the law doesn't blame the software. It blames the company that used it.

Then it walks through how one AI rejection turns into your problem, in four steps:

1. AI scores and filters the CVs.
2. A qualified candidate is rejected.
3. They ask for a human review.
4. You can't explain the model.

At that point, the burden is on you.

It ends with a line I agree with: "let AI take the notes, let humans take the decisions."

I wanted to check the law and the numbers behind it before I shared it with you. Here's what I found. I'm not a lawyer, so treat this as a map and check the details with yours.

The Legal Risk Is Already Here

The UK

Discrimination compensation under the Equality Act 2010 has no upper limit. Under section 136, once a candidate shows facts that point to discrimination, it's on you to prove you didn't discriminate. "The tool decided" is no defence.

The UK also changed its rules on automated decisions. Section 80 of the Data (Use and Access) Act 2025 came into force on 5 February 2026. If a decision about someone is made with no meaningful human involvement, they have the right to challenge it and ask for a person to look at it. The ICO consulted on new guidance for this in March 2026.

The EU

Some AI at work is simply banned. Since February 2025, the EU AI Act has banned AI that reads people's emotions at work, and that includes candidates in a hiring process. Breaking that rule can cost up to €35 million or 7% of global turnover, whichever is higher.

So if your video interview tool claims to "read confidence" or "detect enthusiasm" from a face, ask your vendor about it this week.

The US

In Mobley v. Workday, a candidate over 40 says he was rejected more than 100 times through employers using Workday's AI screening. In May 2025, a court let the case go ahead as a collective action. Workday's own filings said about 1.1 billion applications were rejected through its tools in that period. In June 2026, the court let the core claims continue.

Nobody has won or lost on the merits yet. But 1.1 billion rejections is a lot of people who can ask "why?"

Why AI Gets It Wrong

This is the part I find most interesting, as someone with a master's degree in AI.

Names. Kyra Wilson and Aylin Caliskan at the University of Washington took more than 550 real CVs and changed only the names. The three AI models they tested preferred white-associated names 85% of the time. They preferred female-associated names 11% of the time, against 52% for male-associated names. They never preferred Black male names over white male names.

Order. David Rozado tested 22 AI models on hiring decisions. He gave each model two CVs and asked which candidate fitted the job better. The models picked whoever was listed first 63.5% of the time, and 21 of the 22 models showed this bias.

Think about that for a second. If order didn't matter, it would be 50/50. Your candidate's chances can depend on where their CV sits in a list.

The Honest Trade-Off

These are research setups. They aren't the exact tools you buy, and a vendor may have fixed some of these problems. But they show the kind of mistakes AI makes. If you can't see inside your tool, you can't know which of these mistakes it's making with your candidates.

I'm not against AI in hiring. I use it. AI that helps a person decide is very different from AI that decides.

Three Things to Do This Week

1. Map every AI decision and make the review real. Write down every step where a tool can reject, rank or score a candidate. Most teams I talk to can't do this from memory. Then look at who reviews those decisions. A recruiter clicking "approve" on 300 AI rejections isn't meaningful human involvement. If the person never disagrees with the tool, ask what they're actually doing.
2. Question your vendor, then test the tool yourself. Ask whether the tool has been tested for bias and by whom. Ask whether it reads emotions from video or voice, and whether it can tell you why it ranked someone low. Then run your own test. Take one CV, change only the name and run it again. Swap the order of two candidates and see if the result flips. It takes an afternoon.
3. Be able to explain every rejection. If a candidate asks why they were rejected, you should be able to answer in two sentences. "The score was low" isn't enough. Use clear criteria and a scoring rubric, so you can point to what was missing. Let AI take notes, summarise, schedule and draft. Let people decide who moves forward.

That's how we run interviews at AssessDev. They're human to human and structured, with written notes and scores we can explain.

A Quick Checklist

• We know every step where AI can reject or rank a candidate.
• A real person reviews AI decisions, and sometimes disagrees.
• We can explain any rejection in plain words.
• Our vendor has answered our bias questions in writing.
• No tool in our process reads emotions from faces or voices.
• We've tested the tool ourselves with swapped names and order.
• Candidates know where AI is used and how to ask for a person.

Do you know every place AI makes a decision in your hiring process? And could you explain one of those decisions if a candidate asked tomorrow?

Sources

• Michael Blakley on LinkedIn, sharing Screenloop's "Hiring risk briefing" video (credited to Richard South). TA and Recruitment folks... 12 October 2026.
• UK Equality Act 2010, s.124 (remedies) and s.136 (burden of proof).
• UK Data (Use and Access) Act 2025, s.80 (automated decision-making), in force from 5 February 2026 under SI 2026/82. ICO, draft guidance consultation on automated decision-making. March 2026.
• EU AI Act, Article 5(1)(f) and Article 99. Future of Privacy Forum, Red lines under the EU AI Act: emotion recognition in the workplace.
• Mobley v. Workday, N.D. Cal. Holland & Knight, Federal court allows collective action lawsuit over alleged AI hiring bias. May 2025. Proskauer, Law and the Workplace, AI bias lawsuit against Workday reaches next stage. June 2025. Duane Morris, ruling on Workday's motion to dismiss. 24 June 2026.
• Wilson and Caliskan, University of Washington. Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval. AIES 2024.
• David Rozado. The strange behavior of LLMs in hiring decisions. 19 May 2025. Paper: arXiv:2505.17049.

This post first appeared in Shortlisted by Kat. Photo by Sasun Bughdaryan on Unsplash.

Decisions Made by People

Every AssessDev tech assessment is a live, structured interview run by a senior engineer or CTO. Each one comes with a written report and scores you can explain to the candidate.