GPT-6 Intelligent UI: Why Your Business Chatbot Should Stay Boring

On October 7, 2026, OpenAI started rolling out GPT-6 across ChatGPT with a feature called Intelligent UI. Instead of answering with plain text, ChatGPT can now compose responses with inline buttons, forms, charts, interactive diagrams, calculators and even small playable games generated on the fly. Paid tiers (Plus, Pro, Business, Enterprise) got it that same day. Free and Go users followed on October 8. Enterprise admins can disable it per workspace.

It is, by any measure, the most ambitious change ChatGPT has made since launch. It also tells you exactly what your business chatbot should not become.

What Intelligent UI actually does

The old ChatGPT answered with prose. The new ChatGPT, running GPT-6 Sol on paid tiers and GPT-6 Luna on free, picks a layout based on your question. Ask it to compare three savings strategies and you get a side-by-side table with sliders. Ask it to split a dinner bill and you get a working calculator. Ask about bike maintenance and you get an interactive diagram with tappable parts. The model decides, in real time, which widget to render and wires it to its own answer.

For personal productivity, this is a leap. For a customer-facing chatbot sitting on your website, it is a trap.

Four reasons an interactive widget is the wrong shape for a business chatbot

1. Hallucinated charts are worse than hallucinated sentences

A hallucinated sentence is easy to spot: a human reader notices the wrong number, the wrong name, the invented product reference. A hallucinated chart or calculator is not. If the slider is wired to the wrong variable, if the comparison silently drops a row, if the “estimated monthly fee” widget uses the generic LLM’s guess instead of your pricing table, the user acts on authoritative-looking output they have no way to audit.

The industry has already clocked this. Independent reviews of Intelligent UI note that “when the model renders the thing you act on, every UI it generates needs scrutiny — except now the errors are hidden behind controls that look authoritative.” OpenAI’s own system card has no safety evaluation specific to generated interfaces.

For a chatbot that quotes your company’s prices, warranties, delivery times or compliance rules to a prospect, this is a liability vector. One hallucinated calculator saying your software is €49/month when it is €79 and the customer screenshots it is a hard problem to walk back.

2. EU AI Act Article 50 gets harder on a widget

Article 50 of the EU AI Act has been enforceable since August 2, 2026. It requires any AI system interacting with a human to disclose that fact clearly and distinguishably at the first interaction. Fines reach €15M or 3% of worldwide turnover, and the first wave of inspections led by France’s CNIL, Germany’s BfDI and Spain’s AESIA started in September 2026.

Disclosing “AI generated” in a chat bubble is easy: one line of text, screen-reader compatible, done. Disclosing it on a dynamically generated form that looks like a native UI element is harder. Where does the badge go? Is it still “clearly distinguishable” when the widget occupies two-thirds of the viewport? What about on a chart the user has already interacted with? Every new widget type is a new compliance edge case that your supervisory authority will interpret on their own schedule.

A deliberately boring text-first chatbot with a static “AI assistant” badge in the header is boring on purpose. It passes Article 50 the same way every time.

3. On-the-fly forms are a GDPR landmine

Intelligent UI can render input forms. A business chatbot that generates its own forms on the fly inherits every GDPR obligation attached to the data collected — but with no consent flow, no documented retention policy, no DPIA, no way to pre-declare the fields to your data protection officer, and no legal basis logged against each submission. If a user types their email into a chat-generated form “to receive their quote,” is that marketing consent? Service provision? You cannot answer in advance because you did not design the form.

A chatbot’s lead capture should be a single, deliberately designed component with a known schema, a known consent string, a known retention window, and a known downstream destination (your CRM, your email provider, nothing else). That is what every DPO, every French CNIL inspector and every B2B buyer has in mind when they ask “how does your chatbot handle personal data.”

4. Dynamic widgets are a native phishing surface

The visual grammar of Intelligent UI — tappable buttons, input forms, “login to continue” style flows — is uncomfortably close to the visual grammar of a convincing phishing page. On ChatGPT, the trust boundary is “it is inside ChatGPT, hosted by OpenAI.” On a chatbot embedded in your website, the trust boundary is already blurred: the widget is on your domain, so users will enter credentials, card numbers and ID photos into whatever the chatbot renders. Giving the LLM authority to render arbitrary input UIs removes the hard distinction between “the website asked me for this” and “the chatbot asked me for this.” That is not a theoretical concern — it is the OWASP top risk for AI applications in 2026.

What a customer-facing chatbot should look like instead

The right architecture for a chatbot that greets strangers on your website in 2026 is the exact opposite of a generative UI playground. Three properties, non-negotiable:

• Bounded. It answers from your documents, not from the LLM’s training data. It does not browse. It does not render third-party widgets. It does not generate code or forms. If the answer is not in the knowledge base, it says so and routes to a human.
• Deterministic. The same question returns the same answer. The lead-capture form is the lead-capture form you designed, with the fields your DPO signed off on. The AI-disclosure badge is in the same place every time.
• Auditable. Every user turn, every retrieved chunk, every generated answer is logged against a tenant ID. A CNIL inspector or an internal audit can reconstruct any conversation in minutes, not days.

How DoxyChat is built around those three properties

DoxyChat is a RAG (Retrieval-Augmented Generation) chatbot: it only answers from the documents, URLs and RSS feeds you connect to it. It never improvises. If a user asks “how much does your Pro plan cost?” and your pricing PDF says €79, the answer is €79 — regardless of what the underlying LLM has memorized. No chart, no calculator, no silently miswired slider.

Under the hood: Mistral Large via Scaleway (hosted in France, outside the US CLOUD Act), vector search over PostgreSQL + pgvector with Row Level Security, FastEmbed multilingual embeddings, FlashText moderation with 1062 patterns at ingestion. The widget is a single <script data-bot-id="…"> tag, roughly 150KB. The lead-capture form is one component you configure once with GDPR consent built in. The “AI assistant” disclosure for EU AI Act Article 50 is on screen from the first message, every time, text-based, screen-reader friendly.

Nothing about this architecture is as flashy as a GPT-6 Intelligent UI calculator. That is the point. A boring, bounded, auditable chatbot is what a customer landing on your page needs — and what passes a French data-protection inspection without a sprint of retrofitting.

When Intelligent UI is the right tool

To be clear: GPT-6 Intelligent UI is a real step forward for the problem it solves. If your sales team wants to model deal scenarios in ChatGPT Business before writing a proposal, Intelligent UI will shave hours. If your analyst wants a quick visualization of a CSV, Intelligent UI is better than Excel. These are productivity use cases for your employees, inside a workspace you administer, under contracts you negotiated.

That is a different job from “a stranger lands on our website and asks a question about our product.” That job needs a boring chatbot. Keep the two populations separate.

Try DoxyChat free

The DoxyChat Discovery plan is free forever: one chatbot, ten documents, two hundred queries per month. Upload your product documentation, embed the widget on your site, and watch what a bounded, deterministic, auditable chatbot does with real traffic. Start on www.doxychat.com.