2025
2025
02
02
Interfaces
Interfaces
Behavior
Behavior
Designing AI is not designing screens.
Designing AI is not designing screens.
Designing AI is not designing screens.
Most designers who work on AI products are doing the same job they've always done. They're mapping flows. Building components. Designing states. Refining typography. Making things look considered and feel smooth.
Most designers who work on AI products are doing the same job they've always done. They're mapping flows. Building components. Designing states. Refining typography. Making things look considered and feel smooth.
Most designers who work on AI products are doing the same job they've always done. They're mapping flows. Building components. Designing states. Refining typography. Making things look considered and feel smooth.
That work still matters. But it's not enough anymore. And in most teams, nobody has noticed yet.
That work still matters. But it's not enough anymore. And in most teams, nobody has noticed yet.
That work still matters. But it's not enough anymore. And in most teams, nobody has noticed yet.
What "designing for AI" actually means today
What "designing for AI" actually means today
What "designing for AI" actually means today
Ask a designer what they're working on and they'll tell you about the chat interface. The prompt input. The response container. The loading state. The error message.
Ask a designer what they're working on and they'll tell you about the chat interface. The prompt input. The response container. The loading state. The error message.
Ask a designer what they're working on and they'll tell you about the chat interface. The prompt input. The response container. The loading state. The error message.
These are screens. They are not the product.
These are screens. They are not the product.
These are screens. They are not the product.
The product is the behavior of the system behind those screens. How it responds to ambiguous input. What it does when it's uncertain. How it handles the gap between what the user asked and what the model understood. How it communicates the boundaries of what it can and can't do.
The product is the behavior of the system behind those screens. How it responds to ambiguous input. What it does when it's uncertain. How it handles the gap between what the user asked and what the model understood. How it communicates the boundaries of what it can and can't do.
The product is the behavior of the system behind those screens. How it responds to ambiguous input. What it does when it's uncertain. How it handles the gap between what the user asked and what the model understood. How it communicates the boundaries of what it can and can't do.
That behavior is designed too. It just rarely gets treated as a design problem.
That behavior is designed too. It just rarely gets treated as a design problem.
That behavior is designed too. It just rarely gets treated as a design problem.
Most AI products have beautifully designed screens sitting on top of completely undesigned behavior.
Most AI products have beautifully designed screens sitting on top of completely undesigned behavior.
Most AI products have beautifully designed screens sitting on top of completely undesigned behavior.
Why classical UX patterns break on AI
Why classical UX patterns break on AI
Why classical UX patterns break on AI
UX as a discipline was built on a foundational assumption: deterministic systems. You click a button, something specific happens. You submit a form, you get a predictable response. The designer's job was to make that predictable response feel clear, fast, and trustworthy.
UX as a discipline was built on a foundational assumption: deterministic systems. You click a button, something specific happens. You submit a form, you get a predictable response. The designer's job was to make that predictable response feel clear, fast, and trustworthy.
UX as a discipline was built on a foundational assumption: deterministic systems. You click a button, something specific happens. You submit a form, you get a predictable response. The designer's job was to make that predictable response feel clear, fast, and trustworthy.
AI systems are not deterministic. The same input produces different outputs depending on context, model state, conversation history, and factors the user never consciously controlled. There is no single correct response to design around. There is a distribution of possible responses, each requiring a different kind of handling.
AI systems are not deterministic. The same input produces different outputs depending on context, model state, conversation history, and factors the user never consciously controlled. There is no single correct response to design around. There is a distribution of possible responses, each requiring a different kind of handling.
AI systems are not deterministic. The same input produces different outputs depending on context, model state, conversation history, and factors the user never consciously controlled. There is no single correct response to design around. There is a distribution of possible responses, each requiring a different kind of handling.
The patterns don't transfer. Error states designed for a system that failed in a specific, known way don't work for a system that produced something plausible but wrong. Confirmation flows designed for irreversible actions don't account for actions the system took before the user realized they could be irreversible. Onboarding flows designed to teach features don't prepare users for a system whose outputs will change as it learns more about them.
The patterns don't transfer. Error states designed for a system that failed in a specific, known way don't work for a system that produced something plausible but wrong. Confirmation flows designed for irreversible actions don't account for actions the system took before the user realized they could be irreversible. Onboarding flows designed to teach features don't prepare users for a system whose outputs will change as it learns more about them.
The patterns don't transfer. Error states designed for a system that failed in a specific, known way don't work for a system that produced something plausible but wrong. Confirmation flows designed for irreversible actions don't account for actions the system took before the user realized they could be irreversible. Onboarding flows designed to teach features don't prepare users for a system whose outputs will change as it learns more about them.
Applying deterministic UX thinking to non-deterministic systems doesn't just produce bad design. It produces design that actively misleads users about how the product works.
Applying deterministic UX thinking to non-deterministic systems doesn't just produce bad design. It produces design that actively misleads users about how the product works.
Applying deterministic UX thinking to non-deterministic systems doesn't just produce bad design. It produces design that actively misleads users about how the product works.
What non-determinism changes for the designer
What non-determinism changes for the designer
What non-determinism changes for the designer
When a system can produce different outputs from the same input, the designer's job shifts from designing the response to designing the relationship between the user and a system they can't fully predict.
When a system can produce different outputs from the same input, the designer's job shifts from designing the response to designing the relationship between the user and a system they can't fully predict.
When a system can produce different outputs from the same input, the designer's job shifts from designing the response to designing the relationship between the user and a system they can't fully predict.
That's a fundamentally different problem.
That's a fundamentally different problem.
That's a fundamentally different problem.
It means designing for calibrated expectations. Users come to AI products with assumptions shaped by everything they've used before, from search engines to calculators to forms. Those assumptions are mostly wrong. The designer has to correct them early, gently, and continuously, without making the product feel unreliable.
It means designing for calibrated expectations. Users come to AI products with assumptions shaped by everything they've used before, from search engines to calculators to forms. Those assumptions are mostly wrong. The designer has to correct them early, gently, and continuously, without making the product feel unreliable.
It means designing for calibrated expectations. Users come to AI products with assumptions shaped by everything they've used before, from search engines to calculators to forms. Those assumptions are mostly wrong. The designer has to correct them early, gently, and continuously, without making the product feel unreliable.
It means designing for recovery, not just for success. When the system produces something unexpected, the user needs a clear path forward. Not an error message. Not a dead end. A way to redirect, correct, or try again that feels like a natural part of the interaction, not a breakdown.
It means designing for recovery, not just for success. When the system produces something unexpected, the user needs a clear path forward. Not an error message. Not a dead end. A way to redirect, correct, or try again that feels like a natural part of the interaction, not a breakdown.
It means designing for recovery, not just for success. When the system produces something unexpected, the user needs a clear path forward. Not an error message. Not a dead end. A way to redirect, correct, or try again that feels like a natural part of the interaction, not a breakdown.
Designing for AI means designing for a system that will surprise its users. The question is whether those surprises feel like features or failures.
Designing for AI means designing for a system that will surprise its users. The question is whether those surprises feel like features or failures.
Designing for AI means designing for a system that will surprise its users. The question is whether those surprises feel like features or failures.
The questions you should be asking at every screen
The questions you should be asking at every screen
The questions you should be asking at every screen
Most design reviews for AI products focus on the same things they've always focused on. Is the layout clear? Is the hierarchy right? Does the copy make sense?
Most design reviews for AI products focus on the same things they've always focused on. Is the layout clear? Is the hierarchy right? Does the copy make sense?
Most design reviews for AI products focus on the same things they've always focused on. Is the layout clear? Is the hierarchy right? Does the copy make sense?
These are necessary questions. They're not sufficient ones.
These are necessary questions. They're not sufficient ones.
These are necessary questions. They're not sufficient ones.
The questions that matter for AI products are different. What does the user believe is happening right now, and is that belief accurate? What happens when the system produces something unexpected, and does the interface make recovery obvious? How much of the system's reasoning is visible, and is that the right amount? What signals is the user using to decide whether to trust this output, and are those signals reliable?
The questions that matter for AI products are different. What does the user believe is happening right now, and is that belief accurate? What happens when the system produces something unexpected, and does the interface make recovery obvious? How much of the system's reasoning is visible, and is that the right amount? What signals is the user using to decide whether to trust this output, and are those signals reliable?
The questions that matter for AI products are different. What does the user believe is happening right now, and is that belief accurate? What happens when the system produces something unexpected, and does the interface make recovery obvious? How much of the system's reasoning is visible, and is that the right amount? What signals is the user using to decide whether to trust this output, and are those signals reliable?
These questions don't have standard answers. They require judgment about how much transparency is useful versus overwhelming, how much control gives users confidence versus creates cognitive load, how much explanation builds trust versus undermines the feeling that the system knows what it's doing.
These questions don't have standard answers. They require judgment about how much transparency is useful versus overwhelming, how much control gives users confidence versus creates cognitive load, how much explanation builds trust versus undermines the feeling that the system knows what it's doing.
These questions don't have standard answers. They require judgment about how much transparency is useful versus overwhelming, how much control gives users confidence versus creates cognitive load, how much explanation builds trust versus undermines the feeling that the system knows what it's doing.
That judgment is design work. It's some of the most consequential design work in any AI product. And it happens on almost no design team's agenda.
That judgment is design work. It's some of the most consequential design work in any AI product. And it happens on almost no design team's agenda.
That judgment is design work. It's some of the most consequential design work in any AI product. And it happens on almost no design team's agenda.
What this asks of the designer
What this asks of the designer
What this asks of the designer
Designing for AI doesn't require becoming an engineer. It requires understanding enough about how these systems work to ask the right questions about how they should behave.
Designing for AI doesn't require becoming an engineer. It requires understanding enough about how these systems work to ask the right questions about how they should behave.
Designing for AI doesn't require becoming an engineer. It requires understanding enough about how these systems work to ask the right questions about how they should behave.
It requires being comfortable with ambiguity at a level that most UX processes are explicitly designed to eliminate. Personas help. Journey maps help. But they were built for users making decisions in systems that behave predictably. AI products need design tools that account for systems that don't.
It requires being comfortable with ambiguity at a level that most UX processes are explicitly designed to eliminate. Personas help. Journey maps help. But they were built for users making decisions in systems that behave predictably. AI products need design tools that account for systems that don't.
It requires being comfortable with ambiguity at a level that most UX processes are explicitly designed to eliminate. Personas help. Journey maps help. But they were built for users making decisions in systems that behave predictably. AI products need design tools that account for systems that don't.
It requires thinking about the user's mental model of the system, not just their path through it. A user who fundamentally misunderstands what an AI product is capable of will misuse it, distrust it, or both. Correcting that misunderstanding is a design responsibility, not a marketing one.
It requires thinking about the user's mental model of the system, not just their path through it. A user who fundamentally misunderstands what an AI product is capable of will misuse it, distrust it, or both. Correcting that misunderstanding is a design responsibility, not a marketing one.
It requires thinking about the user's mental model of the system, not just their path through it. A user who fundamentally misunderstands what an AI product is capable of will misuse it, distrust it, or both. Correcting that misunderstanding is a design responsibility, not a marketing one.
It requires caring about what happens after the output, not just before it. Most UX thinking is front-loaded: the research, the flow, the first interaction. AI products live or die on what happens in the fifth session, the twentieth, the moment when the system does something the user didn't expect and the user has to decide whether to stay.
It requires caring about what happens after the output, not just before it. Most UX thinking is front-loaded: the research, the flow, the first interaction. AI products live or die on what happens in the fifth session, the twentieth, the moment when the system does something the user didn't expect and the user has to decide whether to stay.
It requires caring about what happens after the output, not just before it. Most UX thinking is front-loaded: the research, the flow, the first interaction. AI products live or die on what happens in the fifth session, the twentieth, the moment when the system does something the user didn't expect and the user has to decide whether to stay.
The designer who understands this is not doing a different job. They're doing a deeper version of the same job, applied to a layer most designers haven't reached yet.
The designer who understands this is not doing a different job. They're doing a deeper version of the same job, applied to a layer most designers haven't reached yet.
The designer who understands this is not doing a different job. They're doing a deeper version of the same job, applied to a layer most designers haven't reached yet.
Designing a behavior, not a screen
Designing a behavior, not a screen
Designing a behavior, not a screen
There's a useful distinction that most design teams haven't made yet.
There's a useful distinction that most design teams haven't made yet.
There's a useful distinction that most design teams haven't made yet.
Designing a screen means deciding what information appears, in what order, with what visual weight. Designing a behavior means deciding how a system responds to a human being across time, across uncertainty, across the full range of things that can go wrong between intent and output.
Designing a screen means deciding what information appears, in what order, with what visual weight. Designing a behavior means deciding how a system responds to a human being across time, across uncertainty, across the full range of things that can go wrong between intent and output.
Designing a screen means deciding what information appears, in what order, with what visual weight. Designing a behavior means deciding how a system responds to a human being across time, across uncertainty, across the full range of things that can go wrong between intent and output.
Screens are artifacts. Behaviors are systems. And AI products, at their core, are behavioral systems that happen to have screens on top.
Screens are artifacts. Behaviors are systems. And AI products, at their core, are behavioral systems that happen to have screens on top.
Screens are artifacts. Behaviors are systems. And AI products, at their core, are behavioral systems that happen to have screens on top.
The teams that understand this are building AI products that users trust, return to, and integrate into how they actually work. The teams that don't are building beautiful interfaces for systems that consistently fail to meet the expectations those interfaces create.
The teams that understand this are building AI products that users trust, return to, and integrate into how they actually work. The teams that don't are building beautiful interfaces for systems that consistently fail to meet the expectations those interfaces create.
The teams that understand this are building AI products that users trust, return to, and integrate into how they actually work. The teams that don't are building beautiful interfaces for systems that consistently fail to meet the expectations those interfaces create.
That gap is a design problem. It has always been a design problem. The discipline just hasn't fully caught up to it yet.
That gap is a design problem. It has always been a design problem. The discipline just hasn't fully caught up to it yet.
That gap is a design problem. It has always been a design problem. The discipline just hasn't fully caught up to it yet.
Designing AI is not designing screens. It's designing the space between what a system can do and what a human being can understand, trust, and use. That space is where the real work is.
Designing AI is not designing screens. It's designing the space between what a system can do and what a human being can understand, trust, and use. That space is where the real work is.
Designing AI is not designing screens. It's designing the space between what a system can do and what a human being can understand, trust, and use. That space is where the real work is.
Raphaël D. - Head of Product Design, designing at the intersection of AI infrastructure and human experience.
Raphaël D. - Head of Product Design, designing at the intersection of AI infrastructure and human experience.
Raphaël D. - Head of Product Design, designing at the intersection of AI infrastructure and human experience.
Design for AI
Thinking through the design problems that AI products create.
Not how to use AI as a designer. How to design for it.
© 2026 Design for AI. All rights reserved.
Design for AI
Thinking through the design problems that AI products create. Not how to use AI as a designer. How to design for it.
© 2026 Design for AI. All rights reserved.
Design for AI
Thinking through the design problems that AI products create. Not how to use AI as a designer. How to design for it.
© 2026 Design for AI. All rights reserved.