Every week I read another framework for designing AI into products. Transparency. Explainability. Confidence indicators. Graceful failure. Human-in-the-loop. They're all correct, and if you've been shipping AI features you already know the anxiety underneath them: the surface used to be static and predictable, full of signifiers of invisible affordances, and now it talks back, guesses, lacks signifiers leaving affordances invisible, and sometimes gets it wrong.
So here's the short version of everything the best teams are publishing right now — and then the part nobody says out loud, which is that you probably already know how to do all of this.
The "New Rules" for Designing AI Into Your Systems
Tell people it's AI, and tell them what it can't do. Set expectations before the first prompt, not after the first failure.
Show your work. Surface reasoning, sources, and a confidence signal so output can be trusted or challenged.
Keep the human in control. Every AI action needs an obvious override, edit, and undo. Suggestion, not seizure.
Design for being wrong. The AI will be wrong. Recovery is a primary flow, not an edge case.
Fit the existing workflow. Bolt AI onto what people already do; don't make them learn a new religion.
Make the wait legible. Streaming, progress, and "still thinking" states, because AI is slow and non-deterministic.
Close the loop. Let people rate, correct, and steer — and show that their feedback changed something.
Protect data and dignity. Be explicit about what's collected, what's used for training, and how to opt out.
That's the consensus, distilled from Google's PAIR team, Microsoft Research, Apple, Anthropic, OpenAI, IBM, and a dozen sharp practitioners. Print it, tape it to your monitor.
But hold up now…
We Wrote These Rules in 1994
Read that list again. Then read Jakob Nielsen's 10 Usability Heuristics for User Interface Design — published in 1994, refined over thirty years, and still one of the first things most of us learned. It's almost a one-to-one match. The "new rules" for AI aren't new. They're Nielsen's heuristics wearing a black hoodie.
Here's the mapping I keep coming back to:
1. Visibility of system status → Show what the AI is doing. Nielsen said keep users informed through timely feedback. That's your typing indicator, your streaming response, your "searching 4 sources," your confidence score. Latency and uncertainty just raised the stakes on a heuristic we've always had.
2. Match between system and the real world → Speak the user's language. Conversational AI is the most literal expression of this heuristic ever built. The whole promise of natural-language interfaces is that the system finally speaks human instead of forcing humans to speak system.
3. User control and freedom → Override, edit, undo. Nielsen called it the "emergency exit." Every AI-agency framework — Google's autonomy principle, Microsoft's "support efficient correction," Anthropic's human checkpoints — is this heuristic restated. Don't automate away the exit.
4. Consistency and standards → Predictable AI behavior. Users build a mental model of your AI even if they don't begin with one. If the same prompt yields wildly different shapes of answer, or the "AI" badge means three different things across your product, you've violated a 30-year-old rule. Standardize your AI patterns like you'd standardize a button.
5. Error prevention → Guardrails and good prompts. Better than a good error message is a design that prevents the error. That's input validation, example prompts, scoped capabilities, and Apple's built-in safety guardrails. Stop the bad generation before it happens.
6. Recognition rather than recall → Suggest, don't interrogate. Blank-prompt anxiety is a recognition failure. Give people conversation starters, templates, and visible options so they don't have to conjure the magic words from memory. The empty chat box is the new blank canvas problem.
7. Flexibility and efficiency of use → Serve novices and experts. Progressive disclosure of AI: gentle defaults for newcomers, parameter controls and shortcuts for power users. IBM's tiered explanations ("recommended because…" → full reasoning) are pure Nielsen — accelerators for the experienced, safety rails for the rest.
8. Aesthetic and minimalist design → Don't drown people in AI. Not every feature needs a sparkle icon and a disclosure panel. The best AIs — heck, the best UIs in general — often recede. Surface only what advances the task, which is exactly what OpenAI tells its own app builders.
9. Help users recognize, diagnose, and recover from errors → Graceful failure. This is the heuristic everyone's rediscovering as "design for hallucination." Say what went wrong in plain language, explain why if you can, and offer a next step or a fallback. Blame the situation, not the user, and never a raw stack trace.
10. Help and documentation → Onboard people into the AI. Just-in-time guidance, example use cases, and clear capability descriptions. People need to learn what your AI is good at — and Nielsen told us how to teach them decades ago.
Ten for ten. Hopefully by now you're going: "Oh, THAT's all this is. Yeah, I know how to do that."
So…Did Anything Actually Change?
If the heuristics still hold, why does AI feel so different to design for? One word: ambiguity.
Classic interfaces are deterministic. Click the button, get the result, every time. From the abacus in 3000 BC all the way through to tomorrow's hot new app, UX assumed a predictable machine you were helping people operate. AI breaks that assumption. It's probabilistic. The same input can produce different outputs. It's confidently wrong. It operates on intent. Nielsen's own term for the shift is "intent-based outcome specification," where users say what they want, not how to get it, and the system figures out the path.
That's the real design problem, and it's why "designing for ambiguity" is the honest name for this whole discipline. We're no longer designing a set of predictable states. We're designing the scaffolding around an unpredictable partner — the trust signals, the correction paths, the graceful exits that let a probabilistic system be useful without being reckless.
But notice: the scaffolding is still made of the old lumber. Managing uncertainty is just harder-mode error prevention and system-status visibility. Keeping users oriented around a system that improvises is user control and recognition-over-recall, turned up to eleven. The problem got harder. The principles didn't change.
The Practical Takeaway
You don't need to throw out your craft and relearn design from an AI influencer's carousel. You need to apply the fundamentals to a probabilistic medium. Next time a stakeholder asks how you'll "design the AI experience," walk them through ten heuristics that predate the smartphone and show how each one answers a live problem in your product.
The frameworks flooding your feed are valuable — you should probably read them. But they're a modern gloss on a durable foundation. Master the foundation, design for the ambiguity, and the "new rules" stop being intimidating. They become a checklist you've been running your whole career.
Further Reading
Some of the best current thinking on AI UX, if you want to go deeper:
Jakob Nielsen — 10 Usability Heuristics for UI Design and AI: First New UI Paradigm in 60 Years (Nielsen Norman Group). The source material — and Nielsen's own take on intent-based interaction.
Google PAIR — People + AI Guidebook. The most complete human-centered AI design toolkit, from Google's People + AI Research team.
Microsoft Research — Guidelines for Human-AI Interaction (Amershi et al., CHI 2019). 18 validated, still-canonical guidelines.
Apple — Human Interface Guidelines: Generative AI and Machine Learning. Practical patterns for attribution, control, and safety.
Anthropic — Building Effective AI Agents. Transparency over cleverness, simplicity, and human checkpoints — from the makers of Claude.
OpenAI — Apps SDK UX Principles. How OpenAI tells builders to design inside ChatGPT: clarity, native fit, no ornamental UI.
IBM — Everyday Ethics for Artificial Intelligence (Pribić & Rossi). Accountability, explainability, data rights, and fairness, made actionable.
Emily Campbell — The Shape of AI. The best living pattern library for AI UX.
Microsoft — Guidance for Generative AI Application UX. Concrete patterns for dynamic, Copilot-style experiences.
Weisz et al. — Design Principles for Generative AI Applications (IBM Research, CHI 2024). The academic backbone for designing around generative variability.
LogRocket — AI-driven UX design patterns and UXmatters — Why GenAI Demands a UX Revolution. Practitioner syntheses of where the field is heading.
