Are AI Evaluation principles Really New?

For decades, product teams have relied on established human-centred design (HCD) principles and evaluation methods to understand whether products and services work for the people who use them.

Now AI is rapidly becoming part of the products and services we interact with, changing the nature of the experiences we need to evaluate. And, as with any major technological disruption, there is a temptation to assume that the old rules no longer apply, that a fundamentally different technology requires an entirely new approach to evaluating the experiences built around it.

But how much of the UX rulebook really needs rewriting? Does a technology that behaves differently require an entirely new set of evaluation principles – or do the principles we already have still hold up?

Largely not. The heuristics we already use still apply to AI products. What changes is how much weight some of them carry, particularly when it comes to trust.

Why AI seems to need new rules

It is worth asking first why generative AI feels like an exception at all. The instinct has some basis.

Traditional software is deterministic: the same request always gives the same result – a calculator returns 4 for 2 + 2 forever, which designers can anticipate and test against.

Generative language models are probabilistic: rather than retrieving a fixed answer, they generate responses by predicting likely sequences of text. That means the same prompt may produce different responses, varying in wording, emphasis or, occasionally, factual accuracy. This unpredictability creates new challenges for UX evaluation, but behaving differently doesn't necessarily mean needing an entirely different set of principles.

Old principles in new clothes

For decades, UX teams have judged interfaces against usability heuristics – short, established rules of thumb for what makes a system usable, such as keeping users informed, preventing errors and using plain language.

The best known are the ten heuristics Jakob Nielsen published in 1994, building on earlier work with Rolf Molich, which are still in active use. Running a digital interface against them, a method called heuristic evaluation, remains one of the most common ways to inspect a design.

Look at the qualities we associate with good AI experiences through the lens of those heuristics, and much of the territory is familiar. The terminology may have evolved, but many of the underlying principles have not.

Start with user control, one of the qualities on most lists of AI principles. In an AI-powered interface it means letting people edit a generated draft, override a recommendation, or halt the system before it acts. That maps almost directly onto one of Nielsen's ten heuristics, user control and freedom, which holds that because people often act by mistake, they need a clearly marked "emergency exit" – an undo, a redo, a way to back out without penalty.

The "undo send" button on your email is the everyday version. It builds trust in the same way in an AI product: knowing you can always step in and reverse a bad output is what makes it safe to let the system act in the first place.

The same goes for how an AI system handles its own mistakes – and it will make them. Helping users notice, understand and recover from an error is the ninth heuristic, recognise, diagnose and recover from errors – the logic behind every good error message. For an AI system it becomes "graceful failure", and that is what is being judged. When the model gets something wrong, does it surface the mistake rather than bury it, make clear what went wrong, and offer a way forward – correct, retry, or hand control back – rather than dead-ending the user? Handled well, an error becomes evidence that the system is honest about its limits, which is what earns it the benefit of the doubt the next time it is right.

AI can also change how established principles show up in practice. Take recognition rather than recall. Traditionally, the aim is to reduce the user's memory burden by keeping relevant information and choices visible. AI can go further by remembering context on the user's behalf. A chat assistant might retain earlier messages so the user does not have to repeat themselves; a recommendation system might use previous activity to personalise what it shows next. That still serves the original principle, reducing what the user has to remember, but it introduces new questions for evaluation. Is what the system remembers accurate and useful? Can users see or correct wrong assumptions? Can they control what is remembered?

One more heuristic is worth naming, because it sits right on the boundary between old and new: visibility of system status. It holds that a system should always keep people informed about what is going on, through timely feedback – a loading spinner, a progress bar, a "saving..." label. In a generative AI interface, that becomes the typing indicator while a chatbot composes its answer, or the "searching the web", "reading the document" and "thinking" updates an AI agent shows while it works through a multi-step task.

What AI changes: the focus on trust

AI systems behave in ways deterministic products do not. That makes one issue particularly important: helping users judge how far they should trust the system.

Two of the main interface-level mechanisms for supporting appropriate trust are transparency and explainability. They are closely related and often overlap, and different frameworks draw the boundaries between them differently.

Transparency

For the purposes of UX evaluation, it can be useful to think of transparency as helping users understand the "what" of an AI system – what it is, what it can and cannot do, what information it may use and what its limitations are.

For example:

• An "AI-generated" label makes clear that content has been produced by AI.
• A capabilities summary explains what an AI assistant can and cannot reliably do.
• A standing disclaimer reminds users that AI can make mistakes and that important information should be checked.
• An uncertainty or confidence indicator can help users understand how certain the system is about an output and whether they should verify it.
• A data-use explanation tells users what information the system may draw on, such as web sources, uploaded files or account history.

Explainability

Explainability helps users understand how a particular output, recommendation or decision was arrived at.

For example:

• Showing the sources or evidence used to produce an answer, so users can understand what the response was based on.
• Showing why something was recommended, such as a streaming service explaining that a film was suggested based on the user's viewing history.
• Showing the factors that contributed to a decision or prediction, such as which factors influenced a lending decision or insurance risk rating.
• Highlighting the information used to generate an answer, such as the relevant passage in an uploaded document.

Together, transparency and explainability give users information they can use to judge how far to trust the system and its particular outputs.

Trust is not the only thing that changes. AI systems can be confidently wrong, presenting an inaccurate answer in the same assured tone as a correct one, so evaluation needs to look at whether users can spot and check errors. They cannot always explain a particular output, which is why the explainability mechanisms above matter. And they do not hold still after launch: models can be updated quickly, so behaviour can shift without any change to the interface. That changes the rhythm of the work. Testing has to become a regular habit rather than a one-off gate before release.

The real difference

AI adds some considerations to the existing rulebook. The most important for UX is designing for trust – helping people judge when to rely on an AI system, when to question it and when to verify its outputs.

So, are AI evaluation principles really new? Largely not. Evaluating anything, AI or not, comes back to the same discipline: human-centred design. Understand real users and their goals. Keep them informed and in control. Design for the edges, and test with real people. Those principles did not expire when generative AI arrived. If anything, a probabilistic system whose outputs cannot always be readily explained needs them more, not less.

We are extending a discipline refined over decades to cover a system that behaves in a few new ways. The rigour stays the same. It just needs applying more often.

At User Vision, we help organisations hold their AI experiences to the same human-centred standard, through product evaluations, usability testing and accessibility assessments, so a product stays grounded in the people who use it, whatever the technology underneath.

If you'd like to learn more, get in touch at hello@uservision.co.uk.