This post was written by a human.

So what? Did you need to know that? And does it matter?

As a product designer, I’ve never felt compelled to label the source of my visual work. If I produced a product image, I never specified whether it was a render made in KeyShot or a real photograph. Nobody asked, and nobody cared. The output was judged on its own merit.

So what makes AI-generated content different from Computer Graphics or “real” photos? Why, suddenly, does the process matter more than the result?

You probably can’t tell. And you might prefer AI.

Here’s something that might be uncomfortable. In March 2026, The New York Times ran a blind writing quiz, created by Kevin Roose and Stuart A. Thompson, that presented 86,000 readers with five pairs of writing samples spanning literary fiction, fantasy, science writing, historical fiction, and poetry. No bylines, no metadata, no hints. Just two passages side by side. Readers were asked which one they preferred. The result: 54% chose the AI-generated writing. In one pairing, the gap was 67/33 in favour of AI. The human passages were not obscure either. They included work by Cormac McCarthy, Ursula K. Le Guin, and Elizabeth Bishop. The methodology was designed to test literary prose specifically, not functional writing, and the AI was asked to craft original passages rather than imitate the human samples. These were not casual readers. This was a New York Times audience, arguably one of the most literate and discerning readerships available for that kind of test. And they still preferred the AI version more often than not.

The point becomes sharper when you look at what happens once people know the source. A 2024 study from UC Santa Barbara, published in the Findings of the Association for Computational Linguistics (ACL 2025), ran three experiments across text rephrasing, news summarisation, and persuasive writing. They recruited raters through Amazon Mechanical Turk and ran a three-condition test. In the blind condition, raters could not reliably distinguish AI writing from human writing at all, correctly guessing authorship roughly 50% of the time. But in a second condition, where texts were correctly labelled as “Human Generated” or “AI Generated”, people overwhelmingly preferred the human-labelled text, by more than 30%. The third condition was the most revealing: the researchers deliberately swapped the labels, marking AI text as human and human text as AI. The preference pattern followed the labels, not the actual authorship. People preferred whatever they were told was human, regardless of who had written it.

A separate 2025 study, covered by The Conversation, approached the same question from the perspective of consumer behaviour. Participants who were told a story was AI-generated rated it more harshly on dimensions like predictability, authenticity, and emotional depth. Yet when the researchers measured actual behaviour rather than stated opinion, those same participants were willing to spend the same amount of time and money reading the AI-labelled story as the human one. Nearly 40% said they would pay less for AI-generated work, but their reading behaviour and spending did not reflect that claim. As the researchers concluded, many consumers appear to value human labour in theory, yet show little willingness to act on that distinction when it actually costs them something.

The perception gap in practice

This gap between stated preference and actual behaviour is not just academic. It is already visible in public culture.

In November 2025, Coca-Cola released its Christmas ad, generated entirely using AI by studios Silverside and Secret Level. The backlash was immediate. Viewers called it “soulless” and criticised the company for prioritising cost-efficiency over the warmth and craft that people associate with Coca-Cola’s festive campaigns. The company’s head of generative AI defended the decision publicly, stating that the genie was out of the bottle, which only intensified the criticism. This was not an isolated reaction. The 2025 ad followed similar criticism of Coca-Cola’s 2024 AI Christmas campaign, yet the company doubled down, reportedly using even fewer people in production.

That same month, Apple revealed that its new Apple TV intro had been shot entirely in-camera using physical glass panels. No CGI. No AI. Every shimmer and colour shift came from real light hitting real surfaces, with a crew physically moving lights and rotating the glass during the shoot. Apple’s VP of marketing communications, Tor Myhren, said the approach reflected Apple’s commitment to “tactile detail and camera-centric aesthetics.” The behind-the-scenes footage went viral, and the public response was almost the inverse of Coca-Cola’s. People celebrated the craftsmanship.

The contrast is revealing. Both pieces of content achieved their functional goal. Yet one was praised because of how it was made, while the other was criticised for the same reason. The output was not the issue. The label was.

So why are we building systems to label it?

Despite this evidence that our preferences are largely shaped by perception rather than quality, there’s a growing push for formal AI content labelling. And it’s not without good reason.

The Dubai Future Foundation launched the world’s first Human-Machine Collaboration (HMC) icon classification system in July 2025. It defines five levels of collaboration, from “All Human” to “All Machine”, paired with nine functional icons covering stages like ideation, data analysis, writing, and design. The system is mandatory for Dubai government entities and freely available for anyone else. It’s the first standardised, visual approach to disclosing the degree of machine involvement in content production.

On the technical side, Content Credentials, developed by the Coalition for Content Provenance and Authenticity (C2PA), backed by Adobe, Microsoft, Google, the BBC, and others, embed tamper-evident metadata directly into files. They function as a kind of digital nutrition label: recording what tools were used, whether AI was involved, and what edits were made. Unlike watermarks, which can be cropped or compressed away, Content Credentials are cryptographically signed and travel with the file.

These are significant efforts. But the question remains: if people can’t tell the difference, and often prefer the AI version, what problem are we actually solving?

The real reason AI is different

I think the answer isn’t about quality. It’s about two things that genuinely separate AI from previous tools: the ease of generation and the ability to capture likeness.

It would take a tremendous amount of effort and exceptional skill for a CG artist to create a completely photorealistic video of a real human saying something they never said. With AI, it’s becoming trivially easy. That’s the deepfake problem, and it’s a risk that simply didn’t exist at scale with CGI or traditional photography.

For that specific use case, manipulating identity and fabricating speech, we absolutely need robust verification. Content Credentials, the Dubai HMC system, and technical watermarking all make sense as safeguards against deception.

But does that logic apply to everything else?

For a product render, an illustration, a brand animation, or a blog post, the deepfake argument does not apply. Nobody is being impersonated. Nobody is being misled about what is real. The content is doing what it was always meant to do: communicate an idea.

We do not label any other tool in our creative process. I have never credited KeyShot on a product render or disclosed that an image was composited in Photoshop rather than shot in a single frame. A photographer does not footnote their lens choice. A writer does not declare which sentences survived from their first draft and which were rewritten. We judge creative work by what it communicates, not by how it was assembled. AI is a tool. Yet we are not treating it like one.

Should we declare the process behind our work?

I don’t have a clean answer. For content involving likeness, identity, and claims about reality, yes, verification matters and the tools being developed are important. For everything else, I think we’re conflating two separate problems: the genuine risk of AI-generated deception and a vaguer discomfort with the idea that machines can do creative work well.

The research suggests our instinct to favour human-made work is more about the label than the substance. That doesn’t mean the instinct is wrong, but it’s worth interrogating.

This post was written by a human. At least, that is what I told you at the top.

The truth is, this article was written with AI. The ideas, argument, and point of view are mine. The research, structure, and prose were produced collaboratively between me and a machine.

Yet you still read this far. You engaged with the points. Maybe you agreed with some and pushed back on others.

Now that you know how it was made, does that change how you feel about what you just read?

Interested to hear your thoughts in the comments below.

References

• Kevin Roose & Stuart A. Thompson, Who’s a Better Writer: AI or Humans?, The New York Times, March 9, 2026

• Tiffany Zhu, Iain Weissburg, Kexun Zhang & William Yang Wang, Human Bias in the Face of AI: Examining Human Judgment Against Text Labeled as AI Generated, Findings of the Association for Computational Linguistics: ACL 2025, Vienna, Austria

• Reece Akhtar, David Harding & Gia Nardini, People say they prefer stories written by humans over AI-generated works, yet new study suggests that’s not quite true, The Conversation, March 2025

• Dubai Future Foundation, Human-Machine Collaboration Icons, launched July 2025

• Coalition for Content Provenance and Authenticity (C2PA), Content Credentials

• Apple TV identity, shot in-camera with practical glass effects by TBWAMedia Arts Lab, November 2025

• Coca-Cola AI-generated Christmas ads by Silverside AI & Secret Level, November 2024 & 2025

You present an early-stage wireframe, a deliberate structural blueprint, and the first response is: “Can we make that button blue?” It’s the designer’s version of showing someone the floor plan for their house and receiving feedback on the curtains.

This is not a client problem. It is a communication problem. Not everyone can look at a rough sketch and mentally project the finished product. Most people cannot. The real question is not whether clients focus on the wrong details. It is whether we, as designers, set them up to focus on the right ones.

After running design projects through Sans Studio, across brand and product, I have found that the gap between what a designer intends to communicate with a lo-fi deliverable and what the client actually interprets is significant. Most designers present low-fidelity design as show-and-tell. This is a mistake which can lead to poor decisions and unproductive conversations. Closing the gap is itself a design challenge.

The fidelity trap

Low-fidelity work is designed to invite broad, conceptual feedback. Research from the Nielsen Norman Group (NNG) supports this: when a design looks incomplete, stakeholders recognise the work is unfinished and feel more comfortable offering honest, structural critique. People are more willing to challenge a rough sketch than a polished mockup.

However, this only works if the reviewer understands what kind of feedback you are asking for.

Without that framing, low fidelity becomes a liability. Some clients cannot look past the rough appearance. They see grey boxes and placeholder text and conclude you have not done the work. Others fill in the visual gaps themselves — imagining colours, fonts, imagery — and provide feedback on a version of the design that exists only in their head.

Why rough feels wrong

There is a well-studied cognitive bias called the aesthetic-usability effect. First identified by Kurosu and Kashimura (1995), it describes our tendency to perceive attractive things as more functional — even when they are not. Our brains equate beauty with quality. It is a halo effect for design: if it looks good, we assume it works well.

This bias operates at the level of first impressions, often within milliseconds. It works in reverse too. When something looks rough, unfinished, or visually plain, people instinctively downgrade their perception of its quality — including the quality of the thinking behind it.

This is the invisible headwind every designer faces when presenting low-fidelity work. You know the grey boxes represent months of research, information architecture decisions, and careful user journey mapping. However, the client’s brain is pattern-matching against every polished app and website they have encountered. Grey boxes do not match those patterns. So, something must be wrong.

It is not that clients perceive laziness. Their brains are wired to associate visual polish with competence, effort, and trustworthiness. A sketchy wireframe triggers the opposite associations — unfinished, uncertain, not ready — even when the structural thinking is robust.

Consequently, you hear feedback that feels disproportionately emotional for what should be a logical, structural review. Comments like “it doesn’t feel right” or “I’m not sure about this” are often not about layout or flow. They are visceral reactions to the aesthetics of the deliverable, misread as signals about the quality of the work.

As designer Cheryl Platz observes, if something looks near-complete, people are less likely to provide the hard feedback you need. They assume you already know about the structural issues and do not want to offend. Rough work, counterintuitively, is more honest. However, only if you help people past that initial “this looks unfinished” reaction.

You cannot switch off someone’s cognitive wiring. However, you can do two things: explicitly name the gap between how the work looks and the thinking it represents, and create enough context and narrative that the client’s attention shifts from the surface to the substance. Give their brain something else to evaluate.

The real problem: presenting when you should be narrating

The single biggest mistake designers make with low-fidelity work is treating it as a presentation. You share a screen. You show wireframes. You wait for feedback. The client, left to interpret without guidance, gravitates toward whatever is most tangible — which at this stage is usually the wrong thing.

The fix is storytelling.

Instead of showing screens, walk the client through a user journey. Do not say “here’s the dashboard.” Say: “Sarah is a new manager. She has just completed her onboarding assessment. She opens the app and this is what she sees first. Her eye goes here because this is the most important thing we want her to understand at this moment.”

This achieves two things. First, it anchors every piece of feedback in the user’s experience rather than the client’s aesthetic preferences. Second, it gives the client a role in the story — they evaluate whether the journey makes sense, not whether the interface looks good.

Seven things that help

1. Set the rules before you show anything

Before a single wireframe appears on screen, tell the client what they are about to see and what kind of feedback you need. Be explicit:

“What you are about to see is a blueprint, not the building. We are here to discuss whether the rooms are in the right place, not what colour to paint them. The visuals are deliberately rough so we can focus on structure and flow.”

This is not patronising. It is professional. You establish a shared language for the session and give the client permission to think at the right level of abstraction.

2. Make things look deliberately unfinished

There is a subtle but important difference between something that looks rough and something that looks lazy. Tools like Balsamiq, with their hand-drawn aesthetic, signal “this is early” in a way that a clean Figma wireframe does not. When components have a sketchy, marker-pen quality, clients instinctively understand they are looking at thinking-in-progress rather than a near-final deliverable.

If you work in Figma, consider using a lo-fi component kit or hand-drawing key screens to photograph and present. The medium sends a message about the stage of the work.

3. Use real content, not lorem ipsum

This seems contradictory. You want things rough, but the content should be real? Yes. The Interaction Design Foundation (IxDF) makes the case clearly: placeholder text like lorem ipsum is no longer advised because copy and images that differ significantly from your placeholders will affect the final user experience.

You do not need final copy. You need realistic copy. If a heading reads “Your Progress Overview” instead of “Lorem Ipsum,” the client can evaluate whether the information hierarchy makes sense. The conversation remains about whether the right content is in the right place. This is precisely what this stage should address.

4. Prototype the flow, not the pixels

Even a rough clickable prototype fundamentally changes the quality of feedback. When someone can tap through screens in sequence rather than examining flat layouts, they shift from evaluating what it looks like to evaluating what it feels like to use. This prompts a different thought process entirely.

This does not need to be elaborate. A few linked frames in Figma is sufficient. The goal is to make the experience sequential rather than static.

5. Direct the feedback with specific questions

“What do you think?” is the worst question you can ask at this stage. It is an open invitation to comment on whatever catches the eye — which will inevitably be a visual detail that does not matter yet.

Instead, ask task-oriented questions: “Can you find where you would go to check your team’s progress?” “Does the order of information on this screen match what you would want to see first?” “If you were a new user, would you know what to do next?”

This keeps the conversation structural and functional.

6. Use your research as an anchor

If you have done user research, discovery interviews, or workshops, and you should have, reference those findings as you present. “Remember when three of our test participants said they felt overwhelmed by too many options? That is why we have simplified this to two paths.”

This shifts the authority for design decisions away from personal opinion — yours or the client’s — and toward evidence. It is significantly harder for someone to argue “I don’t like it” when the response is “the users told us this is what they need.”

7. Know when low-fidelity is not enough

Some clients genuinely struggle with abstraction. Not everyone can look at grey boxes and mentally project the finished product. For those clients, a mid-fidelity approach — greyscale but clean, with real layout structure, proper typography scale, and meaningful content — is the better option.

This is not a compromise. It is reading your audience. The goal of any deliverable is to facilitate the right feedback at the right time. If a slightly more polished wireframe achieves that more effectively than a hand-drawn sketch, use it.

Reframing the conversation

The deeper principle here is that presenting design work is itself a design problem. You are designing an experience — the experience of understanding and evaluating a concept that does not fully exist yet. That experience requires the same rigour you would give to any user journey: clear entry points, guided navigation, appropriate context, and a well-defined action at the end.

When a client fixates on details at the wrong stage, it is not because they are difficult. It is because we have not given them the right framework to engage with the work.

The wireframe is not the deliverable. The conversation is.

And like any good design, that conversation works best when it starts with the user.

iPhone Air adopts muted, premium colours and polished titanium finishes, once reserved for the Pro line. It’s positioned as a refined, design-led object - a statement piece.

iPhone 17 Pro reverts to aluminium, for thermal performance properties. The colours are bold and saturated. With fewer aesthetic constraints, Apple maximised raw capability. It’s positioned as a professional tool.

The contrast in CMF directly differentiates two ‘premium’ products. This diverging approach highlights a strategic shift: Apple differentiating form and function within its flagship portfolio.

For designers, it’s a reminder that CMF decisions go beyond aesthetics; they signal the relationship a user has with a product. Sometimes the most powerful design move is not to blend form and function into one device, but to separate them.

AI is rapidly redefining how we interact with software. Traditional graphical user interfaces (GUIs) built around nouns (menus, buttons, forms, and other static elements) are giving way to AI-driven interactions and automations focused on outcomes and actions. Instead of manually clicking through interfaces, users increasingly tell the computer what outcome they want and let intelligent algorithms handle the execution (Nielsen, 2023).

This shift raises new challenges and opportunities for designers: how to introduce novel AI capabilities while maintaining a predictable, trustworthy user experience. In this article, we explore four key principles shaping AI-native UI design: the move from noun-based to action-based interfaces, the balance between novelty and familiarity, the role of feedback and transparency, and the trade off between fidelity and immediacy, with examples from cutting-edge tools.

From Nouns to Verbs: A New Paradigm

Traditional software UIs have been dominated by visible objects (files, buttons, form fields) and the principle of direct manipulation — users achieve goals by performing step-by-step commands on these noun-like interface elements. In an AI-native interface, this paradigm is flipped. Interaction becomes verb-centric: users express what they want to do (often in natural language), and the system handles the low-level operations. Nielsen (2023) describes this as a new “intent-based” paradigm where “users tell the computer the desired result but do not specify how this outcome should be accomplished,” reversing the locus of control from the user to the AI (Nielsen, 2023). In other words, the user declares an intention, and the system figures out the procedure.

Real-world examples illustrate this shift. ChatGPT and other generative chatbots let the user simply request an action or output, rather than navigating menus or manually executing each step. A task that might have required many clicks and actions can now be achieved with one well-phrased prompt. This approach also reduces a technical barrier preventing people from executing their ideas; they no longer need to learn advanced software tools. For instance, imagine prompting an image AI to create a complex sci-fi illustration versus using a graphics editor like Photoshop; the difference is clear, as seen in figure 1 which highlights the difference in UI between Adobe Photoshop (traditional) and Adobe Firefly (AI-native).

GitHub Copilot, a coding AI, similarly exemplifies verb-driven interaction. As a developer types code or comments describing a goal, Copilot instantly generates suggestions to implement it (figure 2). It works “directly in your editor, suggesting whole lines or entire functions” based on the context (GitHub, n.d.). This represents a more recent change in AI interaction: shifting from separate apps to direct integration with the source application. In effect, the interface element (the code editor) becomes an intelligent assistant that responds with actions (code generation) rather than passive text. This makes the act of coding a higher-level, verb-oriented collaboration between human and AI.

This approach remains consistent across other modalities. Voice-based assistants allow users to simply say what they want, like “Schedule a meeting tomorrow with John”, without navigating a calendar UI. ChatGPT Advanced voice and multimodal models mean such conversations can become even more nuanced and reliable. However, voice alone has limitations, Singh (2025) notes that going “voice-only” for a day was exhausting, as it demands continuous articulation and doesn’t suit every context or mood.

When users are unsure what they want, or are merely exploring, many AI products now fight “blank-prompt anxiety” by surfacing contextual suggestions and step-wise controls. Flora AI, for instance, uses a canvas-based UI that affords and encourages exploration by using nodes/boxes with branching and prompt helpers (Flora, n.d.), as seen in figure 3. By scaffolding open-ended exploration in this way, AI interfaces convert vague intent into concrete direction while preserving user agency.

The future likely lies in hybrid interactions where voice, text, and GUI elements complement each other. Still, the trend is clear: whether via chat, intelligent automation or voice, AI-native UIs are moving away from exposing underlying data structures and controls and instead focus on helping users accomplish goals with minimal friction.

Balancing Novelty with Predictability

Designing the future of interfaces means treading a fine line: introducing enough novelty to harness AI’s potential, but not so much that users feel lost or mistrustful. A striking observation in current AI products is how uniform and minimal many of their interfaces are. This is a deliberate strategy. “Sameness is safe, and in the world of emerging tech, safe is strategic”, writes Olena Mostepan (2025), referring to the copy-paste design trend across AI tools. When users are confronting a fundamentally new capability, a familiar interface lowers the cognitive load. The chat prompt borrows the mental model of texting or search, something users have done countless times, so they can focus on what they want the AI to do, not how to operate the UI (Mostepan, 2025).

Crucially, too much novelty in UI can backfire. If an AI product introduced a radically new interaction paradigm on top of the new technology, users might be overwhelmed. As Mostepan (2025) puts it, “innovation demands novelty, but usability depends on familiarity”. For instance, Notion AI is embedded within the existing Notion UI rather than presented as a flashy new toy. Users invoke it by typing a familiar slash command within a page (figure 4), and the AI’s responses appear inline in the document. This design makes the powerful features feel like a natural extension of Notion, “taking full advantage of [Notion’s] incredible design to make [prompting] feel natural.” (Lee, 2023). That said, designers shouldn’t shy away from all new interface ideas, the key is to introduce novelty gradually and judiciously. As users become comfortable with basic input paradigms, products can start pushing the envelope.

Predictability in an AI-native product is inseparable from clarity: the interface must telegraph what will happen next and where new controls will surface. Adaptive UIs show how dynamic elements can enhance efficiency without disorienting users. When buttons hide, morph, or re-order, their location and affordance should remain familiar so users’ spatial memory still works; otherwise novelty feels like instability. Consistency of layout, colour, and motion cues reinforces trust, and a single, visibly focused input anchors the experience even as surrounding actions adapt. Designers should also expose multiple levels of abstraction, letting experts trigger granular controls while newcomers issue high-level intents, so people can predict outcomes whether they think in detailed steps or broad goals.

In the near future, we may see more experimental interface elements. The guiding principle is to ensure any new interaction still behaves in a predictable way. Users should be able to form reliable mental models: knowing when the AI will act, how to invoke it, and what to expect from it. Novelty for its own sake must never undermine clarity.

Feedback and Transparency

As AI becomes more autonomous, the risk is that users might feel unsure about what the system is doing or why. Feedback and transparency are thus paramount in AI-driven UIs. To foster trust, an interface should constantly communicate two things to the user: what is happening (or just happened) and why/how it happened (Eriksen, 2025). Classic usability heuristics like “visibility of system status” still apply, arguably even more strongly, in the age of AI (Eriksen, 2025). Users need to feel in control of the outcome, even if they delegated the task to an AI.

AI UIs should constantly provide feedback about the current system state. Effective progress indicators blend animation and clear text to reassure users that the system is active. For brief waits, a simple spinner with copy such as “Generating draft…” can suffice, but once delays exceed a few seconds designers should switch to determinate cues that expose steps or percentages (e.g. transcribing → translating → summarising) so users can see where the time is going (Sherwin, 2020). By considering what the user might be doing whilst waiting for the AI, Deep Research mode in ChatGPT executes progress indication well, as seen in figure 5. Where possible, add an ETA, research shows explicit time estimates reduce user anxiety (Rosala, 2019). Finally, progressive disclosure, first showing a low-resolution preview, then swapping in a high-fidelity render, keeps users engaged and gives them a chance to abort, preserving both patience and resources.

However, feedback is a two-way street. Not only should the system inform the user, but the user must be able to give feedback to the system, and good AI-native UIs give clear indications of AI actions. For example, after an AI generates content or executes a command, the result might be highlighted or accompanied by a subtle label. Gemini in Google Workspace, when it produces an answer or a paragraph, immediately offers controls for the user to accept or reject that output, shown in figure 6. This not only lets the user decide whether to keep the AI’s contribution, but also reassures them that nothing is applied irreversibly without their consent. Response feedback serves a dual purpose, it reassures the user that the developers care about quality, and it gives the user a small sense of agency. Such mechanisms embody a critical mindset: AI suggestions are proposals, not commands, and the human is still the ultimate decision-maker.

Transparency also means explaining or exposing a bit of the AI’s “thinking” when appropriate, making it clear why the AI made a decision. Since an AI may use hidden data or heuristics, the interface should surface relevant context so the user isn’t left guessing. The search tool Perplexity is notable for this. When it returns an answer, it doesn’t do so in a black-box manner, instead it returns answers with in-line citations. In practice, similar systems should cite the reference material that the answer is based on. By giving users a trail to follow, the UI increases hallucination confidence: the user can verify and understand how the AI arrived at its output. Users are more forgiving of AI errors when they can see the rationale and correct it. Indeed, Nielsen (2023) cautions that when people “don’t know how something was done, it can be harder for them to identify or correct the problem”. Hidden behaviour might wow users in demos, but in daily use it can erode trust.

Communicating that an AI has genuinely grasped a request is almost as critical to trust as the eventual result. Conversational assistants borrow a page from human active-listening, paraphrasing the user’s intent before acting (e.g. “Sure, here are some hotel options”) a technique shown to boost perceived understanding and confidence in the system (Lau, 2023). Multimodal experiences complicate that handshake: when a user shares an image or video they cannot easily tell what the AI perceives that image to be, unlike in human communication where we all see the same thing. Products such as Google Lens cope by overlaying bounding boxes and lightweight labels on the detected object while the search runs, so users can instantly confirm that the AI is “looking” at the right thing, as shown in figure 7. Emerging patterns propose heat-map previews, tap-to-highlight regions, or transient “what-I-see” ghost layers that flash before the model answers, giving people a chance to nudge the AI’s focus. Developing such visual acknowledgements will be essential for multimodal assistants.

Ultimately, designing for transparency means accepting that AI systems have probabilistic and sometimes unpredictable natures, and making that visible and manageable for the user. This could even mean inventing new UI conventions: how do you display an AI’s confidence level or uncertainty? How do you indicate that an answer might be incomplete or speculative? How do you communicate prompt adherence? These questions are prompting innovation in interface design.

Fidelity vs. Immediacy: Controlled Illusions

Generative AI processes take time, which leads to a tradeoff between fidelity (how good the output is) and immediacy (how quickly the output is returned). Designers should determine which to prioritise based on context, and sometimes stage the experience so it feels fast and ends up high-quality.

Model choice is often the first lever. Google’s new Gemini 2.5 family illustrates how providers now publish separate variants for speed-sensitive versus quality-sensitive jobs: Flash is optimised for “high-volume, latency-sensitive tasks”, while Pro is recommended for “high-complexity reasoning” or richer multimedia outputs, even if responses take longer (Google, 2025). Product teams should therefore map user stories to model tiers.

In creative workflows, it is important to find a balance between fidelity and immediacy. For example, avatar video generation platforms such as Argil AI render the output in a low-fidelity draft (seen with a blurred overlay in figure 8), which allows the user to make changes quickly (Y Combinator, 2025). Only once the user is satisfied is the high resolution, computationally heavy video processed, saving both user patience and GPU bills. Interleaving quick feedback loops with inexpensive refinement removes a common pain point: waiting minutes for a perfect render only to discover an error.

Conversely, some queries demand slower, multi-step orchestration because correctness trumps speed. OpenAI’s Deep Research mode in ChatGPT plans and executes web searches, source evaluation and summarisation in the background; answers take 5–30 minutes, but users get a citation-rich report suitable for due-diligence or academic work (Lawler, 2025). Interface copy and progress bars manage expectations so that a longer wait is interpreted as “thorough, not broken”.

In voice-enabled use-cases, designing initially for fidelity becomes paramount, because even a 600 ms delay can shatter the illusion of natural conversation. Voice pipelines therefore cheat, by streaming the first token from a speed-optimised model, then back-fill with higher-quality synthesis. OpenAI’s Realtime API streams audio outputs as they’re generated and auto-handles interruptions, making dialogues feel live (OpenAI, 2024). ElevenLabs recommends its Flash TTS models and streaming to cut “time-to-first-byte”, noting that chunking plus pre-roll fillers or micro-pauses can mask longer back-end processing (ElevenLabs, 2025). Human-computer-interaction studies even show that synthetic “um… let me see” fillers measurably reduce perceived wait times for 8–16 second delays (Abbas et al., 2021, 2–14), a reminder that illusion management is an important part of product design.

In practice, teams rarely favour fidelity or immediacy outright, they layer them. Light models or previews keep the human-AI interaction snappy, while heavyweight reasoning or high-resolution passes run once the user signals commitment. The decision lies in choosing the break-points, and in signalling them clearly, so that users feel in the loop even when the system is quietly working overtime in the background.

Conclusion: Designing for an AI-First Future

The rise of AI-native user interfaces represents a profound shift in interface design, arguably the biggest since the advent of mobile touchscreens. We are moving into an era where the interface is less about controls and more about possibilities: software can already generate content and perform complex actions on behalf of the user.

Looking ahead, we can expect AI to enable experiences that are more adaptive and context-aware than ever. But even in such futures, users will always need clarity, predictability, and a sense of mastery over their tools. As we design the next generation of AI-native products, success will come from marrying machine intelligence with human-centered design, achieving true human-machine-collaboration. The interfaces of the future should feel like natural extensions of our intentions, empowering us to do more, with less effort. For designers and product builders, the challenge is to harness this power without overwhelming or alienating people.

References

Abbas, T., Gadiraju, U., Khan, V.-J., & Markopoulos, P. (2021, 10 4). Making Time Fly: Using Fillers to Improve Perceived Latency in Crowd-Powered Conversational Systems. Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, 9(1), 2-14. https://doi.org/10.1609/hcomp.v9i1.18935

ElevenLabs. (2025, March 6). Conversational AI latency with efficient tts pipelines. ElevenLabs. Retrieved July 4, 2025, from https://elevenlabs.io/blog/enhancing-conversational-ai-latency-with-efficient-tts-pipelines

Eriksen, M. (2025, April 7). Designing AI User Interfaces That Foster Trust and Transparency. UXmatters. Retrieved July 2, 2025, from https://www.uxmatters.com/mt/archives/2025/04/designing-ai-user-interfaces-that-foster-trust-and-transparency.php

Flora. (n.d.). FLora. FLORA. Retrieved July 4, 2025, from https://www.florafauna.ai/

GitHub. (n.d.). Getting code suggestions in your IDE with GitHub Copilot. GitHub Docs. Retrieved July 1, 2025, from https://docs.github.com/en/copilot/how-tos/completions/getting-code-suggestions-in-your-ide-with-github-copilot

Google. (2025). Gemini models. Gemini API. Retrieved July 4, 2025, from https://ai.google.dev/gemini-api/docs/models

Lau, E. (2023, June 8). From Chatbots to Confidants: The Evolution of Conversational AI. UX Primer. Retrieved July 4, 2025, from https://articles.ux-primer.com/from-chatbots-to-confidants-the-evolution-of-conversational-ai-607fdd9a3aa1

Lawler, R. (2025, February 2). ChatGPT’s agent can now do deep research for you. The Verge. Retrieved July 4, 2025, from https://www.theverge.com/news/604902/chagpt-deep-research-ai-agent

Lee, L. (2023, June 1). Building the AI × UX Scenius — with Linus Lee of Notion AI. Latent Space. Retrieved July 2, 2025, from https://www.latent.space/p/ai-interfaces-and-notion

Midjourney. (n.d.). Discord Overview – Midjourney. Midjourney. Retrieved July 4, 2025, from https://docs.midjourney.com/hc/en-us/articles/33330535666445-Discord-Overview

Mostepan, O. (2025, June 16). Understanding the Design Trend That’s Taken Over AI. BuiltIn. Retrieved July 2, 2025, from https://builtin.com/artificial-intelligence/ai-minimalist-design

Nielsen, J. (2023, June 18). AI: First New UI Paradigm in 60 Years. Nielsen Norman Group. Retrieved July 1, 2025, from https://www.nngroup.com/articles/ai-paradigm/

OpenAI. (2024, October 1). Introducing the Realtime API. OpenAI. Retrieved July 4, 2025, from https://openai.com/index/introducing-the-realtime-api/?utm_source=chatgpt.com

Rosala, M. (2019, February 3). Status Trackers and Progress Updates: 16 Design Guidelines. Nielsen Norman Group. Retrieved July 4, 2025, from http://www.nngroup.com/articles/status-tracker-progress-update

Sherwin, K. (2020, September 11). Mask Interaction Delays with Progress Indicators (Video). Nielsen Norman Group. Retrieved July 4, 2025, from https://www.nngroup.com/videos/progress-indicators

Singh, P. (2025, March 21). Rethinking Interfaces for AI Era. Why Buttons and Text Boxes Won’t Cut It Anymore. Medium. Retrieved July 2, 2025, from https://medium.com/@pacificleo/rethinking-interfaces-for-ai-era-why-buttons-and-text-boxes-wont-cut-it-anymore-1af101405284

Y Combinator. (2025, February 27). AI Interfaces Of The Future | Design Review. YouTube. Retrieved July 3, 2025, from www.youtube.com/watch?v=DBhSfROq3wU&t=7s