Chatbots have so far produced a very uniform output: text, paragraph by paragraph, occasionally a list or a code block.This is precisely where an innovation comes in that OpenAI announced in connection with the broader release of GPT-6 for ChatGPT.Under the name "Intelligent UI", responses are to consist less of pure text and instead contain more useful, interactive elements.

At first, this sounds like a detail improvement to a product.In fact, it touches on a fundamental assumption of digital product design: that the interface is designed in advance, implemented and then filled with content.When a model decides in what form a response appears, this order shifts.

From text field to generated interface

The term Generative UI has become established for this development – that is, user interfaces that are not hard-coded but are generated at runtime from the respective request and the context.Instead of a text description of a comparison, a table appears; instead of a list of options, a selectable element; instead of calculation instructions, an operable input field.

The idea is not to replace design.The idea is to adapt the form of presentation to the task, instead of pressing every task into the same response format.A chat history is well suited to explanations, but for selection, comparison and configuration it is often the worst available interface.

Which parts of a product can be generative

For companies that operate their own digital products, this raises a sober question: which parts of the interface must remain stable and recognisable, and which could sensibly emerge dynamically?A first sorting helps:

  • Design as fixed: navigation, brand identity, legally relevant representations, forms with binding details, checkout and payment processes.Here, reliability, verifiability and recognisability are more important than flexibility.
  • Candidates for generative approaches: result displays in searches, product comparisons, help and support dialogues, dashboards with changing questions, summaries of complex data sets.
  • Borderline cases: advisory journeys and configurators.They benefit from adaptability but need traceable logic and reproducible results.

This classification is not a technology question but a risk question.The more strongly an interface influences or documents decisions, the less its form should be left to the randomness of a model call.

What generative interfaces require

That a model can generate sensible interaction elements is not enough.For this to become a robust product, it needs framework conditions that are good practice in classic projects anyway – except that here they become mandatory.

  1. A defined component system. If generatively generated interfaces are composed only of vetted components, accessibility, brand image and behaviour remain controllable.Freely generating layouts, by contrast, quickly leads to results that nobody wants to take responsibility for.
  2. Clean, machine-readable data. An interface can only be as good as the structure behind it.Those who have modelled content and product data consistently can present them in different ways.Those who manage them in free text fields cannot.
  3. Clear interfaces. Interactive elements trigger actions.These actions must meet defined, secured endpoints, with permission checks and logging – regardless of how the interface came about.
  4. Measurability. If presentations vary, usage data is harder to compare.There needs to be an idea in advance of what success is and how it will be measured.

Assessment with a sense of proportion

The announcement initially concerns a single product.It cannot be inferred from this that classic frontends will become superfluous at short notice.A more sensible reading is the reverse: if users get used to receiving suitable forms of interaction from AI systems, expectations of all other interfaces will rise too.Long walls of text, rigid filters and form journeys that treat every request the same will then stand out more.

For TYPO3 projects and other existing platforms, this does not mean a break but a prioritisation.A consistent content model, structured fields instead of layout in text fields and a documented API are the prerequisite for later delivering content in different – including generatively generated – presentations.This work pays off regardless of whether generative interfaces ultimately play a large or a small role.

A pragmatic start

Anyone who wants to assess the topic does not need a major project.A manageable area with many different user questions – such as an internal search, a help section or an analytics view – is well suited to checking whether dynamic presentations actually lead to faster decisions.What matters is that comparison with the previous solution remains possible.

For web and software projects, this above all means: the investment at present belongs less in generative interfaces themselves than in their foundation – clean data structures, a binding component system and secured interfaces.Those who have this foundation can use generative elements in a targeted way where they bring benefit, and leave them out where reliability is more important.

Without it, even the best model will not produce a usable interface.

Sources