Using ISO 24495 to assess AI generated content

Generative AI can produce fluent text, but fluency does not establish that the information is clear, accurate or useful. A draft may sound convincing while leaving out a condition or assuming knowledge the reader does not have. To assess it properly, a team needs an agreed basis for deciding what the communication should achieve. The plain language principles give us a useful framework for that work.

At Narratology, we use the four principles of ISO 24495-1: relevance, findability, understanding and usability. They help us examine the information readers need, how it is organised and whether people can use it. They also help us decide what an AI system needs to be told before it generates a draft. The standard itself does not establish that a particular AI system is suitable for the task.

Relevance: decide what information belongs in the explanation

Relevance begins with a selection decision. Before asking a model to write, establish what the reader needs to know and why that information matters in their circumstances. If the model does not receive that context, the draft may rely on assumptions that sound plausible but miss the point. The person preparing the brief needs evidence about the audience and a clear account of the task.

Useful information includes the decision the reader faces, their existing knowledge and the circumstances in which they will encounter the communication. Include relevant accessibility or language needs where these are supported by evidence. Do not ask the model to invent a typical customer or infer vulnerability from uncertain signals. Where information is missing, the workflow needs a way to identify and resolve the gap.

The same applies to product facts. A brief needs approved information about the conditions, costs and actions the explanation must cover. If the organisation has not settled a policy question, a model cannot resolve it by producing a smooth sentence. Someone responsible for the product needs to establish what the organisation can accurately tell the reader.

Findability: specify how the information should be organised

A generated draft can be grammatically sound while making the main message difficult to find. Models may add background or a friendly introduction that delays information the reader needs immediately. Decide the order of the explanation before asking for a particular tone. This gives the model a structure to follow and gives a reviewer something specific to check.

That structure should reflect the purpose of the communication. Put the main message where the reader can recognise it early, then provide the detail needed to understand its effect. Use headings that explain their sections, and keep important conditions close to the decision they affect. Reusable templates can help apply these choices consistently, provided the template is appropriate for the task.

The surrounding journey still needs review. A well-organised page may be difficult to reach, or the link leading to it may not explain why it matters. Metadata can help describe its subject and author, but cannot ensure that a search or AI system will interpret it correctly. The team needs to consider the route to the information alongside the structure of the page itself.

Understanding: control terminology and sentence structure

AI systems can introduce variation even when a team has good reasons to use the same term throughout. In financial and legal content, related expressions may describe different things. An overdue balance and an outstanding amount, for example, are not necessarily interchangeable. Give the model approved terms, definitions and usage examples so its choices can be compared with the organisation’s intended meaning.

Sentence guidance needs to address structure as well as length. Ask for clear subjects and actions, and examine clauses that interrupt the main point or create ambiguity. A shorter version can still be difficult if it compresses several relationships into a small space. Review the whole explanation to check that the reader can follow how each condition relates to the outcome.

Voice and tone need a similar level of definition. A word such as “friendly” leaves considerable room for interpretation across cultures and situations. Practical guidance should explain which language choices suit the context, supported by examples the team has reviewed. This makes it easier to assess a generated draft against something more specific than personal preference.

Give numbers enough context to keep their meaning

Copying a figure accurately does not guarantee an accurate explanation. The reader also needs the relevant unit, period and condition. A monthly interest rate loses essential meaning if the draft removes “per month”, while a minimum payment can be misrepresented as a full balance. The review therefore needs to compare the surrounding wording with the source, alongside checking the digits.

Agree how dates, currencies and percentages should appear, then include those conventions in the writing guidance. A table may help when readers need to compare related information, provided its structure remains manageable. Check how the explanation works on the devices and with the assistive technology people will use. These presentation decisions affect whether the figures are useful to the reader.

Usability: find out whether the information works for readers

The final principle asks whether people can use the information for the purpose it was written to support. A model’s assessment of its own draft cannot establish this. Important communications need testing with intended readers, taking account of the circumstances in which the information will be used. Ask people to find relevant information, explain it and use it in a realistic task.

AI may help organise research material or flag recurring queries for investigation. Those outputs still need to be checked against the original evidence, with appropriate protection for participants’ information. After publication, support queries and other indicators can reveal further questions. A change in contact rates needs interpretation, because a completed process or fewer complaints does not by itself demonstrate understanding.

Build the framework into the writing process

Writing teams are increasingly designing the conditions under which content is generated, as well as writing the content itself. That makes agreed source material, review responsibilities and testing plans particularly important. The ISO principles help organise these decisions around what readers need. Their value comes from applying them to the actual communication and checking what the process produces.

We help teams connect their writing guidance with practical drafting, review and evaluation. Explore our AI writing training and plain language training to discuss how these skills fit your organisation’s work.

Source and further reading

ISO 24495-1:2023, Plain language: governing principles and guidelines. The AI workflow described here is Narratology’s application of the principles.