Why GenAI Still Looks Like a Terminal Window: an interface analysis
- David Turner
- Aug 28
- 7 min read
Why GenAI Still Looks Like a Terminal Window
GenAI has permeated everyday life. The go-to answer for millions of people faced with any number of different tasks. So why does it have such a basic monochrome interface?
Open a browser or app for any of the mainstream GenAIs, and you’ll be faced with some variation of a black page with a box for entering white text into. You can turn the mic on, upload a file or image, but those are ultimately context-driven sideshows. The text box is the start-point for almost all GenAI workflows.
If you’d never heard of GenAI, and someone explained its impact and capabilities to you, you’d naturally imagine something with a complex, immersive interface. Animations, interactive displays, sounds, colours, even spatial or 3D visuals. Then you’d open a browser window and see what looks like a Terminal console. It makes SAP look like a Matisse.
A GenAI Interface Analysis
Developers live inside Terminal, so it’s maybe natural that their latest creation replicates what they’re used to. With pressure to ship fast, rolling out features, updates, and new models, the UI just needs to work. So Terminal it is, albeit plain language rather than bash/zsh.
More importantly, a text interface has some very practical advantages - flexibility being chief amongst them. Fundamentally, a text interface doesn’t require the product to know in advance what the user wants to do.
The usual interface for a program has to make assumptions about the task. Excel has one set of controls, Lightroom another, Google Maps another still. A general-purpose AI system can’t make the same assumptions. One user or session may want to analyse a spreadsheet, the next to write some Python, the next to plan a holiday, and the next to ask about a dodgy knee.
Why the text box won a war over aesthetics
So the text box is first of all a utilitarian compromise. It gives the user a general-purpose, low-latency entry point without imposing a fixed information architecture onto the product, while remaining accessible to screen readers, adaptive tech, and low-bandwidth connections.
But there must be more to it - tech companies know all about interfaces and have no shortage of designers. It’s a long stretch to imagine Google overlooking product interface design on the basis that a text box is ‘good enough’. Consider too that it’s already been four years since the public release of ChatGPT - there’s been more than enough time to ship an MLP design.
It’s certainly not a technical limitation; AI companies have the most powerful silicon in the world at their disposal, knocking back millions of gallons of water a day to cool off after a tough day at the data centre (NB: when I ran this piece through three different AIs for fact checking, ChatGPT was the only one that told me I should remove any references to energy or water usage, with all the tact of a gardener who’d been caught watering their lawn during a hosepipe ban).
Maybe then we should be looking at real-world user behaviour to complete the picture.
What users are really typing into it
Research into consumer prompt data, such as HBR’s analysis of real-world AI usage, reveals a top-10 list dominated by highly personal, unstructured interactions: companionship, troubleshooting, relationship advice, fan fiction, and general life guidance. Similarly, a study published in Nature Health analysing over 500,000 health conversations on Microsoft Copilot found that over half related directly to personal symptom checks, emotional wellbeing, and medical education.
And it’s not hard to understand why. There’s no human judgement, no waiting for an appointment, you get unlimited numbers of questions and almost unlimited time in the multi-turn conversation. And importantly you can type and converse silently, avoiding the attention of eavesdroppers in the dentist waiting room or top floor of the bus. Who knew the societal impact of AI would extend to creating a shortage of overheard gossip for those clandestine networks of old ladies on their way to Tesco?
We can imagine that as these personal use cases dominated user-data, product teams naturally prioritised back-end model capability and discreet interfaces over aesthetic panache.
So the text box became the favoured interface for GenAI through a combination of flexibility, privacy, accessibility and familiarity.
The UX of GenAI, however, is where the really interesting stuff is concentrated. Instead of redesigning the home screen, Product and Engineering teams across the board have focused on building agentic infrastructure: Model Context Protocols, API hooks, tools, and sub-agents. Workflows that allow developers and creatives to connect tools and systems, create loops and orchestrate agents until the output is iterated, tested, documented and committed. Where visual innovation has occurred, it has taken the form of split-screen workspaces that support these flows. Think Claude Artifacts or OpenAI Canvas showing code or interactive frontend panes as they’re generated.
Generative UI: the next visible shift
This leads us to Generative UI, the phase of GenAI interface currently emerging. To solve the problem of not knowing what domain a user or session is going to focus on, the solution is dynamic.
Ask for an analysis and it might produce a table or chart. Ask to compare two products and it might produce a side-by-side comparison. Ask for a calculator and it might produce one. Ask to edit an image and the relevant controls and filters can appear. Once finished, the interface can disappear again.
Google has already demonstrated and begun deploying generative interfaces in its products, while tools such as Vercel's AI SDK and open-source projects such as CopilotKit are building the infrastructure needed to support this model. Anthropic launched Claude Design earlier this year, although in my capacity as an erstwhile graphic designer and artist, I think it’s largely underwhelming in a creative sense - hackneyed, regression to the mean stuff, the equivalent of AI-generated text. The interfaces and connections to workflows of course mean it will become a fixture in plenty of enterprises, but it does still highlight the difference between shipping a good-enough product versus shipping a good one.
GenUI creates significant implications for product design. The traditional software model is to build a fixed interface around a defined set of tasks and requirements. The emerging GenUI model, however, is to provide a general-purpose conversational entry point, and then construct parts of the interface around the task as it develops. So the text box may not disappear, but transition into a start-point before stepping back to allow dynamically-generated interfaces to complete tasks. The text box also provides a technical fallback for when GenUI compute costs and latency become too much.
Ambient AI and the privacy problem it can't dodge
The most likely phase to follow Generative UI is what’s variously described as ambient, invisible, or embedded AI. Not invisible in the literal sense, but more an assistant that appears when needed within the context of a user’s task. A simple example is the predictive text that appears in Google Docs and Gmail, offering to complete sentences as you write.
The AI isn’t a distinct application or destination in this case, but ultimately an omnipresent assistant baked into operating systems. Google is already making moves in this arena with Gemini intelligence embedded into Android 17, adding Gemini Nano Integration and System-Level Task Automation. Although it has to be said their marketing use-cases aren’t convincing (‘snagging a front-row bike for your spin class’).
There is however a substantial obstacle to this type of AI interface design, and that's privacy.
An AI embedded in an operating system or application can potentially understand far more about the user's activity than today’s chatbots do. An ambient assistant will have access to documents, messages, calendars, location, conversations and even the physical environment.
The oft-mooted move towards wearable and audio-based AI makes this issue particularly contentious. An always-on device listening to its surroundings might be useful to its wearer, but compromises privacy for everyone in the vicinity. Recording or processing conversations in a public place raises questions about consent, data protection and social acceptability. Which means shortening the odds on the owner of such a device getting punched in the face. A problem you don’t tend to get with a text interface.
So with audio recording and ambient AI posing compliance risks, what about the idea of spatial computing as an interface?
Virtual and augmented reality have been discussed as the next big thing for decades, but neither has managed to displace conventional screens where consumers are concerned. Headsets are usually expensive, and almost always look ridiculous. Not to mention the practical limits around displays, input and nausea.
AI however may succeed in changing the economics of spatial interfaces, as generating a three-dimensional environment no longer necessarily requires somebody to design every component manually.
Google DeepMind's Genie 3, for instance, is a world-model capable of generating interactive environments from plain descriptions. Its immediate applications are more obvious in areas such as simulation and training embodied systems than in everyday consumer interfaces, but the underlying technology points towards a different possibility; software environments that are generated rather than designed in advance.
That’s potentially significant for AI interfaces, because the distinction between an application and its output begins to evaporate. If an AI can generate the environment needed to complete a task, the application no longer has to exist as a fixed collection of screens.
But none of this means that the current chat interface is about to disappear.
Text will almost certainly remain important because it’s flexible, precise, simple, and as close to private as reasonably possible. Voice will remain useful where hands-free interaction matters, continuing to be a matter of user or cultural preference otherwise. Conventional graphic interfaces will remain useful where users need to inspect information or understand a complex state.
So the more likely outcome is that these modes become interchangeable. A user might describe what they want in text, inspect the result visually, manipulate it directly, ask a question by voice and then allow the AI to execute the resulting task. The interface will be assembled dynamically from whatever combination is appropriate.
That represents a more significant change than simply making today's chat window look sexier, and leads us toward a conclusion of this GenAI interface analysis.
The next phase of AI interface design is unlikely to be about replacing the text box with a more aesthetically pleasing fixed interface. It’s more likely to be about making the interface responsive to the task itself.
Some of that has happened through side panes, workspaces, tools, APIs and generated components. Generative UI is already extending the idea further, allowing parts of an interface to be created as needed. Embedded and ambient AI could eventually remove the need to open an AI application at all.
For Product, Design and Engineering domains, the primary question has evolved from ‘which LLM should we use’ to one of how the software around that model should behave, where the interaction should take place, what context the system should have access to, and how much of the interface should be fixed in advance.
Models will continue to improve, but the more interesting change is likely to be in how users interact with the software around them.
David Turner is the founder of Kói, an independent technology consultancy advising investors, founders, and boards on AI strategy and emerging AI interface design.
You can reach him at: enquiries@dkoi.design
© Kói Holdings Ltd 2026. All Rights Reserved.
