When the answer becomes an interface

Facts are attributed and linked to their sources. The analysis and opinions are the author's own.

What it is

On 7 October OpenAI brought GPT-6 to all ChatGPT users, which it says number more than 1.2 billion people a week. The headline feature is Intelligent UI: OpenAI says it trained GPT-6 to compose responses from text, visuals and interactive elements, so that answers can include graphics, tappable buttons, forms, charts and interactive experiences used directly in the conversation.

Paid tiers (Plus, Pro, Business, Enterprise) get GPT-6 Sol; Free and Go tiers get GPT-6 Luna, rolling out from the following day. OpenAI notes that the models powering its Work and Codex products are not changing as part of this release, and that Enterprise availability depends on workplace admin settings.

Two days earlier OpenAI also set out its approach to text watermarking in response to the EU AI Act: an invisible watermark on eligible ChatGPT and Codex text in the EU over the coming weeks, opt-in watermarking for API customers globally (off by default), and a detector initially limited to approved researchers.

Why it matters

For years we evaluated language models on what they said. An interface is different: it has defaults, it nudges, it hides some options and foregrounds others. A generated chart has an axis someone chose. A generated form decides which fields are required. Every one of those is a small design decision now being made by a model, at the scale of a billion weekly users.

A generated chart has an axis someone chose. Now the someone is a model.

That shifts the measurement question. Accuracy of text is no longer enough; we need to ask whether the generated interface represented the options fairly, whether the chart's scale exaggerated a trend, whether the button the model put first is the one the user would have chosen. These are the counter-metrics an interface needs.

There is a practical upside that is easy to miss. Many of the questions people bring to a model are really small analytical tasks: compare these options, show me how this changes over time, help me fill this in. A model that answers with a working chart or a short form can be far more useful than three paragraphs. For learning, an interactive explanation often beats a static one; I built this site's Learn section on that bet.

But usefulness raises the stakes on correctness. A wrong sentence is easy to spot and discount. A polished, interactive chart carries an authority that prose does not, and most people will not check the axis. The better the interface looks, the more scrutiny the data underneath deserves.

For teams deploying this inside an organisation, I would treat generated interfaces like any other analytical output: they should state their data source, their assumptions and their date, and they should be reproducible. If a model builds a tool in the moment, someone should be able to see what it computed and why.

What I'd watch

• Admin controls in enterprise settings. OpenAI already ties Enterprise availability to admin settings; the useful next step is policy on which interactive elements are allowed in regulated workflows.
• Provenance for interfaces, not just text. Watermarking text is a start; knowing which parts of a generated tool were model-composed will matter for audit.
• Whether teams start testing generated UIs the way they test human-built ones: accessibility, dark patterns, and whether the default choice serves the user.

Sources

1. GPT-6 and Intelligent UI for everyone. OpenAI, 7 Oct 2026.
2. Our approach to EU text provenance rules. OpenAI, 5 Oct 2026.