AI liability: ask your AI what you'd ask a new hire

You can’t move for apocalyptic nightmare scenarios about AI at the moment, so here's one for you. A customer, the ICO or even a judge asks why you made a decision, and you hold up a piece of paper with four words on it. "AI told me to."

I ain't no fancy talkin' big city lawyer. Taking legal advice from me is like taking fashion advice from our Engineering team. But Wordnerds turns what customers say into what organisations do, and we use AI in a range of forms to do it, so this is something we think about a lot.

This guide covers who is liable, what happens when AI goes wrong and what to ask a supplier. Then it sets out the Hiring Test, four steps for checking an AI the way you'd check someone new to a job: write the job description, test before you trust, supervise where it matters, and keep the file.

Who is liable when AI analyses your customer feedback?

I wish I had a cheerier answer. It’s you. Whichever AI or supplier your organisation uses. Under UK GDPR, the organisation deciding why and how feedback is analysed must be able to demonstrate its processing complies. Using a vendor doesn't move that duty. The ICO's guidance on AI (2023, now under review after the Data (Use and Access) Act) says "overall accountability for data protection compliance lies with the controller, ie your organisation", and "You cannot delegate these issues to data scientists or engineering teams."

In the Law Commission's words (July 2025), autonomous and adaptive AI systems "do not currently have separate legal personality".

Wordnerds CEO and AI legality bore Pete Daykin went into more detail on the issue in his frankly magnificent CX Corner Newsletter (sign up now why not?):Who is liable when AI gets it wrong?

What happens when AI gets it wrong?

We’re still figuring out all the delightful and imaginative ways that AI could find to bite you in the bum, and real consequences are starting to raise their ugly heads.

Suspected AI mistakes are already costing organisations money, reputation, even people's careers. In a June 2025 High Court claim against the London Borough of Haringey, the claimant's lawyers had cited five fictitious cases; the barrister and law centre were each ordered to pay £2,000. The court couldn't establish whether AI was used, but it warned heads of chambers and managing partners they will be asked what they did to make sure their lawyers check what AI produces.

In October 2025 Birmingham's safety advisory group banned away fans from Aston Villa's match against Maccabi Tel Aviv on West Midlands Police's advice. The police report cited a Maccabi match against West Ham that never happened, which one officer told the inspectorate came from a Microsoft Copilot search, one of eight inaccuracies it listed. The Chief Constable's retirement was announced in January 2026.

None of this is new. Remember the COVID exam fiasco? On 26 August 2020 Boris Johnson told pupils their grades were "almost derailed by a mutant algorithm", as if it had trapped the politicians and the locked pod bay doors. With AI, the reasoning is even harder to see.

Should you avoid AI for customer feedback altogether?

Honestly, plenty of organisations haven't started. The government's 2025 AI adoption survey (published January 2026) found around 1 in 6 UK businesses use AI. As I see it, people fall into either fear and inertia, missing out on the immense power of AI, or they go at it gung ho and store up major problems for the future.

But "should we use AI?" just isn't the right question. Imagine asking a colleague how the numbers on a project were checked. They lean back in their chair and say, "Oh, don't worry about that! We got a human to do it."

A human. Call me a mistrustful lanky Scottish pillock if you must, but I'd have a BUNCH of follow-up questions. An infant is a human. A Premiership footballer is a human. Charlie Sheen is a human. Being a human doesn't mean you're qualified to do the job.

Which human? How did you choose them? How were they trained? How do you know they're doing a good job?

Congratulations. These are exactly the questions you should be asking of your AI.

What should you ask an AI supplier?

The first question to ask any supplier that offers AI is “Can I do it again?” We could name a boatload of AI providers who fail on it. It’s also, sad to say, not enough.

Run the same data through the same AI twice: do you get the same answer? Researchers have seen accuracy vary by up to 15% across repeated runs on the same input (Atil et al., 2025).

So why isn’t it enough? Well, if what your AI produces is bad, getting the same bad thing every time isn't exactly a ringing endorsement.

So the other question you need to ask is “Can you show me what it’s doing?”

Channel your inner maths teacher and ask to see the working. Don't settle for the model's own explanation: research presented at NeurIPS in 2023 found a model's step-by-step explanations "can systematically misrepresent the true reason for a model's prediction".

UK guidance suggests a similar set of due diligence:

• Test the supplier's claims before you buy, and agree regular accuracy reviews (ICO, 2023).
• Define acceptable performance and ask for process logs (UK Guidelines for AI procurement, 2020).
• Ask whether your data is used to train future models (AI Playbook for the UK Government, 2025).

How do you make AI accountable?

The Hiring Test makes AI accountable by asking of it what you'd ask of a person doing the job, and keeping a record you can hand over when someone asks.

Your team defines the categories, sets the pass mark, checks the work and signs off the record; the AI does the reading at scale.

Things get more specific from here, as we zoom in on Wordnerdery, but the same steps work with any tool that can show its working.

At Wordnerds, The Hiring Test is normally a 2-4 week period at the start of the onboarding process, and then it's ongoing throughout our work with customers.

How do you define what your AI should look for?

Step 1: write the job description. What exactly is it the AI should do? Give every category a written definition: what's in and what's out, with an owner and a date. One published research method records each code with a "label, definition, description, qualifications or exclusions and examples" (Roberts, Dowell and Nie, 2019).

Here at Wordnerds, every category has a specific definition showing what's in and out of the theme. That definition is what the AI is held to, and what you show when someone asks why a comment ended up where it did.

One thing to watch here: sometimes when trying to categorise data, people can send out definitions so broad that PT Barnum would take you aside and ask you to narrow things down.

The Department for Transport's evaluation of its AI consultation tool (2025) found that deciding "how a theme should be defined and at what level of aggregation" took judgement, and noted past human analyses with themes as general as "Benefits (general)".

If a definition can't tell a reviewer what doesn't belong, it isn't finished.

How do you test AI analysis before you trust it?

Step 2: test before you trust. Your AI's probation period. A trained and qualified person (Charlie Sheen can stop sending us CVs) checks a sample of the data in every theme against its definition. The goal here isn't 100%; no system gets there, people included.

Set the pass mark before you look at the results. Methods researchers call agreement thresholds "ultimately arbitrary" (O'Connor and Joffe, 2020), so the number is your call.

Compare the AI with your own people, and test blind. In a government test of its Consult tool (published May 2025), the AI matched a senior adviser 62% of the time, two policy advisers matched that adviser 70 to 73% of the time, and the two advisers agreed with each other only 62% of the time.

How do you check AI's work without re-reading everything?

Step 3: supervise where it matters. So you've trained your model, tested it until you trust it, and you're ready to sit back and enjoy all the efficiencies and time savings.

But what if your AI is specifically trained to look at a set of issues, and then the data changes? Customers start talking about different things, their language evolves, the already-subjective barriers separating themes merge. If you're not careful, your carefully curated themes can look more outdated than my super-relevant Charlie Sheen references. But if you spend all your time trying to find flaws in your model, are you actually saving time?

At Wordnerds, we use unsupervised topic modelling alongside the trained themes, so that if there's a material change in what you're looking for, or something comes up that you didn't expect, you can be made aware without having to be in the data all the time.

You should also test again if the model changes. The Department for Transport's 2025 evaluation reports that it re-engineered and re-evaluated its prompts whenever it changed or upgraded models. This step leaves you with a check log.

If you want more on this step, four of the best and brightest Nerds discussed intentional AI adoption in more depth here.

What records should you keep of how AI reached an answer?

Step 4: keep the file, so you can answer two questions: "How did you get to that answer?" and "How did this mistake happen?" The government's transparency template for public-sector algorithms makes a ready-made checklist. It asks for:

• a senior responsible owner
• how human review works
• the model version
• how the model's performance was evaluated
• a maintenance and review schedule

Different sectors are moving at different speeds. In finance, the FCA found good practice where senior managers were accountable from the design stage and firms kept "auditable records of decisions, including rationale, challenge, and escalation" (October 2026).

At Wordnerds, the platform collects this automatically: who checked a theme, the changes they made, and a Health Score showing how well the theme is performing.

For councils and other public bodies, the file can be requested under Freedom of Information. On who owns the definitions once everyone can question the data, see Who owns the answer?

What are the common mistakes with AI feedback analysis?

These are the pitfalls we are forever seeing with AI feedback:

• Checking with the AI's answers in view. Seeing the AI's answer first can anchor the checker, as the Department for Transport's 2025 evaluation warns. Check blind.
• Trusting one overall score. An average hides the weak categories. Report each category.
• Assuming the model stays the same. The same model's behaviour "can vary greatly over time" (Chen, Zaharia and Zou, 2024, testing 2023 models). Re-test after every change.
• Testing on someone else's data. GPT-4's performance varies "across tasks, even within datasets" (Pangakis and Wolken, 2025). Test on your own comments.
• Looking for errors but not gaps. In the Department for Transport's 2025 evaluation, without human review, its tool found 124 of the 165 themes analysts had identified, and a 2025 clinical study in npj Digital Medicine found AI notes left things out more often than they made things up. Ask what's missing.

What happens when someone asks how your AI got its answer?

So. A few months after all that, we're back in the room with the customer. Or the judge. (Or both, judges buy things too).

You're asked, "how did you get this answer?", and you open the file. The definition of the theme is there, with its owner and the date it last changed, next to the agreement check and health score, topics showing any gaps in the framework, and the name of the person who signed it off.

These steps aren't designed for a magical world where nobody ever makes a mistake. They mean that when a mistake is made, you have a clear, transparent process to understand what happened, why it happened and how to fix it.

If you'd like to talk through what transparent AI needs to look like for your organisation, talk to our team.

---

Subscribe to CX Corner →

---

This article was last reviewed on 2026-11-02. Next review: 2027-01-31.