Skip to main content
NexoraLaunch home

Build log live

Explaining Model Choices and Limits to Your Users

AI · 10 min read ·

Which model you use, why, what it can and cannot do and how it may change: a practical guide to telling users enough about the AI inside your product.

Illustration: A deep-violet technical panel with a stacked diagram: user input, product logic, model, output, each a thin-line geometric block with a small label

When someone uses your AI product, they are trusting a system whose behaviour they cannot see. The more they understand about what is inside, what it is for and where it fails, the better they can use it, and the more they will trust you if something goes wrong. Telling them is not a threat to your product. It is part of the product.

This guide describes what to tell users about your model choices and limits, how to write it, where to put it and how to keep it current.

Why users need to know

Different users have different reasons.

  • Individual users want to know when to rely on the output and when to check.
  • Technical buyers want to know capability, cost implications, latency and which provider is involved.
  • Security and privacy teams want to know where data goes and who processes it.
  • Regulators and procurement may require documentation about how an automated system works.
  • Journalists and researchers want facts they can quote.

A short, accurate statement serves all of them and saves you repeating yourself.

What to say about the model

Keep it factual and brief. A good model statement covers:

Which model or models you use. If you use a third-party provider's model, say which provider and, where the provider's terms allow, which model family. If you use more than one, say what each is used for. If you fine-tuned or trained your own, say so and what it was trained to do.

What it is used for. The specific tasks in your product: summarising, classifying, extracting, generating. Not every feature uses the model.

What surrounds it. Instructions, retrieved documents, checks, filters and human review. Most of the quality of a product lies in these.

Where it runs. The region or environment, and whether any data leaves your control. A large language model is a type of model trained on large amounts of text; running one involves sending text to a service that may be operated by you or by a third party.

How it may change. Whether you pin to a particular version, how you test changes and how you notify users.

An example, in prose: "The summaries are produced by a third-party large language model, accessed through its provider's interface. Your text is sent to the provider to generate the summary and is not used by us to train models. We add instructions and check the output against your source text for obvious omissions. We test each model update on a fixed set of examples before switching, and we note changes in our changelog."

What to say about limits

Limits are where you earn trust. Cover these categories.

Accuracy and reliability. What kinds of errors occur and how often, if you know. "In our test of two hundred documents, about one in twenty summaries omitted a key figure."

Hallucination. If the product generates text, the risk that it will state things that are not supported. Say what you do about it, for example by quoting sources, and what users should check.

Coverage. Languages, domains, formats and lengths the product handles well or poorly.

Bias and fairness. Where relevant, what you have done to examine differences in performance across groups, and what you have not been able to check.

Safety and misuse. What the product should not be used for, and how you try to prevent misuse.

Dependence on inputs. Quality of the output depends on the quality of the input; say what helps.

Not a substitute for professional advice. In areas such as health, law and finance, say that outputs are not professional advice.

Use plain, specific language. "May occasionally be inaccurate" is true of everything and informative about nothing.

The model card idea

A model card is a short document that accompanies a machine-learning model and describes its intended use, performance, evaluation and limitations. The idea was introduced in a widely cited research paper, and many organisations now publish similar documentation.

You do not need to produce a research-grade card. A one-page version for your product can include:

  • Model details: type, version, provider, date.
  • Intended use: tasks, users and contexts.
  • Out-of-scope use: what it should not be used for.
  • Data: a plain description of any data used to adapt it, and of your test data.
  • Evaluation: what you measured, how and with what result.
  • Limits and risks: as above.
  • Contact: how to report problems.
  • Date and version of the card.

The NIST AI Risk Management Framework encourages organisations to document AI systems, their purposes, their risks and their measures. A model card is a practical way to do part of that.

Data handling statements

Users will ask what happens to their data. The UK Information Commissioner's Office has published guidance on AI and data protection, which highlights the importance of transparency, fairness and accountability when personal data is used in AI systems. For your users, a clear statement should cover:

  • What data is sent to the model.
  • Whether it is stored, by whom and for how long.
  • Whether it is used to train or improve any model.
  • Who the sub-processors are.
  • How a user can delete their data.

A separate guide on this site covers privacy statements for AI tools in more depth; the point here is that the model statement and the data statement should agree.

Where to put it

  • A page in your documentation or help centre, titled plainly, such as "How our AI works".
  • A short summary near the feature, with a link to the full page.
  • In the product, at the point of use: an information icon or a line beneath the output.
  • In your listing, a sentence on the model and a link.
  • In sales and support materials, the same facts.

Keep the wording identical, or very close, everywhere, so that no one finds two versions.

Update it

Model statements go stale quickly. When you change a model, a provider, a data practice or a safety measure, update the page, put a date at the top and add a note to the changelog. Review it every quarter even if nothing seems to have changed.

If a change might alter users' results, tell them before it takes effect, where possible, and offer a way to compare or to stay on the previous behaviour for a time.

A worked example

A team offers an assistant that drafts responses to customer emails. Their old help page said: "Powered by cutting-edge AI". They replace it with a one-page "How the assistant works".

It says that drafts are produced by a third-party language model, that the assistant sees the customer's email and the help articles the user has connected, and that nothing is sent without a person pressing a button. It lists known limits: drafts can include out-of-date policy details if the connected articles are out of date; the assistant works best in English; it is not suited to legal or medical correspondence. It describes the test: two hundred past emails, two reviewers, and a summary of the findings with the date. It states that customer emails are processed by the provider only to generate drafts and are not used to train models, subject to the provider's terms linked on the page. It explains that model updates are tested before release and noted in the changelog.

Sales calls become shorter because the page answers the questions buyers always ask, and support receives fewer complaints from users who understand what they are getting.

Questions founders ask

Will naming the provider help competitors? It tells them little they could not guess. Buyers, on the other hand, may require it. If a provider's terms restrict naming, follow the terms and say what you can.

What if I use several models behind the scenes? Describe the roles, not every route: "a smaller model for classification and a larger one for drafting". Update the description if the arrangement changes materially.

How much detail do I give about test results? Enough that a reader could repeat the test or judge its relevance: task, data description, size, measure and date. Do not publish anything that exposes private data.

What if a user finds a failure I have not documented? Thank them, reproduce it, add it to your test set and, if it is a pattern, add it to the limits section.

Can I say the product is safe? Avoid absolute words. Describe the measures you take and the risks that remain, so that users can judge for themselves.

A short reading test

Ask a colleague outside the product team to read your model page and answer five questions: what does the AI do here, what could go wrong, what happens to my data, who do I tell if it fails and when was this page last updated? If they cannot answer all five from the page, rewrite until they can. The test takes ten minutes and is more reliable than any internal review, because the people who built the product cannot easily see what is missing.

Summary

Tell users which model you use, what it is for, what surrounds it, where data goes and how it may change. Describe limits specifically: accuracy, hallucination, coverage, bias, misuse and the need for human judgement. Consider a one-page model card, keep your data statement consistent with it and place it where users will find it. Date it and update it whenever something changes. Plain candour about the model is one of the most valuable features an AI product can offer.

Questions and answers

Do users need to know which model I use?
Many technical buyers do, for reasons of capability, cost and data handling. Say so in a plain statement and keep it up to date.
What is a model card?
A short document that describes a machine-learning model: its intended use, how it was evaluated, its limits and considerations for use.
What happens when a model is updated?
Behaviour can change, sometimes in ways that affect users' results. Describe how you test and announce changes.
How do I explain limits without scaring users?
State them factually, with examples and what to do about them. Users trust specific limits more than vague reassurance.

Sources

Related guides

Ask a question