Summary: Defaults in design are not simple shortcuts or starting points. They carry design decisions, bias and assumptions about what users need and want moment-to-moment. In AI interactions, and especially with AI agents, defaults are critical.

Why Defaults in AI Matter

Think of a default as a design decision made on behalf of someone who never got asked. In AI user interactions, that decision is made once and inherited by everyone downstream.

Defaults should be questioned in all designs, but in AI it matters more than in conventional UX. Here are a few reasons why:

• Most users never change their defaults. This is the oldest finding in behavioral design. For example, default enrolment in retirement (pension) schemes, organ donation opt-out or managing GDPR cookie banners are all designed with the intent of catering to a majority of users. But in AI user experience, the impact is stronger, because people don’t know there’s a setting. In non-AI enhanced design, you can see a checkbox on a form. You can’t see that the model was tuned to agree with you.
• Defaults are invisible, so they can’t be contested. A visible design decision can be debated; but a default passes as necessity. For instance, users won’t experience an AI assistant’s confident, fluent prose as a choice someone made. They will read it as inherent to how the thing works. This default that users do not see holds a design decision that often escapes scrutiny. That’s why questioning defaults has to be a deliberate act by your team. Nobody else is in a position to do it.
• Defaults inherit history. A bad default in a web form affects that form. A bad default in a foundation model, a system prompt template, or a speech stack is reproduced by every team that builds on it, most of whom will never know the decision existed. This is the scale difference that makes it a strategic issue rather than a usability one.
• Failure is silent and self-blaming. When your UK doctor’s AI voice LLM’s default is a Yorkshire accent and patients miss it, it doesn’t say “I wasn’t trained on people like you.” It acts like the the Yorkshire accent is the ‘normal’ default and you’re wrong. The user absorbs the failure as personal. Exclusion that produces no complaint also produces no backlog ticket, so it never gets fixed.
• Defaults encode who the “normal” user is. Every default is an answer to the question “Who is this for?”. “Normativity” is endemic in Product and Engineering teams, usually small and homogeneous groups, working under delivery pressure. The imagined user becomes the actual user, and everyone else is in the long tail by design.

In agentic systems, they’re not just preferences, they become authority: The default autonomy level, the default permission scope, the default definition of “done” determine what gets done to the world without a human in the loop. That’s a governance decision wearing a UX costume. See my Redefining Agentic AI with UX and Service Design (30 min Lightning Lesson, watch free).

• Defaults set the incentive, not the intent. Defaults are where business strategy turns into user behavior. Nobody has to argue for engagement in a design review. They just set the default, and most users never change it.

The practical upshot: defaults are the highest-leverage thing a design team controls. Changing one default changes the experience of everyone who never touches a setting, which is nearly everyone. That’s why questioning them is worth more than most of what sits on a design backlog.

Some of the Key AI Default Settings to configure

• Base Prompt (System Prompt / Persona): The core instruction that dictates how the chatbot behaves, its tone of voice, and its boundaries. Without a default base prompt, an AI chatbot will revert to generic behavior.

• Default Reply / Fallback Message: The catch-all response or behavior triggered when the AI doesn’t know the answer or when a user’s input falls outside its knowledge base.

• Temperature: A parameter (usually ranging from 0 to 1) that controls creativity. A default of 0 or low values is best for factual, grounded answers (like customer support or data-retrieval bots), while higher values allow for more creative text generation.

• Underlying Model: The default Large Language Model (LLM) such as GPT-4o mini, GPT-4o, or Gemini that powers the reasoning engine behind the bot.

Defaults in AI are assumptions shipped at scale

Defaults in AI aren’t settings. In conventional UX, a design default used to live in one product. A gender field, a font size, a screen-share menu. Get it wrong and you excluded the people who used that product. AI defaults don’t work like that. They sit in a model, a system prompt, or a voice stack that thousands of teams build on top of without ever seeing the decision that was made. One default, inherited everywhere, questioned by almost no one. Some examples:

• Voice assumes a speaker. Every voice interface has an opinion about who is talking. It has a default accent it hears best, a default pace, and a default silence length it reads as “you’ve finished.” Speak with a strong regional accent, a stammer, second-language rhythm, or a speech disability, and the system doesn’t say “I wasn’t designed for you.” It says “Sorry, I didn’t catch that,” and you conclude the failure is yours. The default voice it speaks back in carries assumptions too: the industry spent a decade defaulting to a female-sounding, endlessly agreeable assistant and called it neutral.
• Chat assumes a reader. The default tone of most assistants is confident, agreeable, and long. None of those are neutral choices. Confidence is a default that hides uncertainty from the people least able to spot the error. Length and reading level are defaults that quietly exclude users with cognitive disabilities, low literacy, or English as a second language. And plain language is almost never the default, it’s a thing you have to ask for. But agreeableness is a default because agreeable answers score well with raters, which is the AI version of “addiction is more profitable than connection.” Ask any assistant about a typical workday, a family budget, or a normal commute, and you’ll meet the imaginary user it was built around. They’re usually Western, able-bodied, neurotypical, salaried, and well-connected.

Agents raise the stakes: The defaults that matter most in agentic systems are the ones about autonomy and permission. How much does the agent do before it checks with you? What counts as “done”? Is cheapest the default definition of a good flight, or is it the one with connection times a wheelchair user can actually make?

Agents are typically set up with broad access to mail, calendar, files, and payment, granted once and never revisited. Least privilege is not always the default. Neither is a visible audit trail, an undo, or an escalation to a human when the agent gets stuck. It retries and guesses instead, and the people harmed by a confident wrong guess are the ones with the least slack to absorb it.

What to Notice

There are four things that are turned on when you first use an AI product, without you asking for them. Each one helps the company more than it helps you. They are: engagement, data collection, memory and autonomy. The defaults that protect people are switched off: plain language, confirmation steps, least privilege, opt-in memory, human handoff. That asymmetry isn’t a conspiracy. It’s what happens when nobody in the room is paid to ask who the default leaves out.

Four questions for any AI feature you’re designing. Who did we picture when we set this default? Who does it fail silently, in a way they’ll blame themselves for? What does the safe option cost the user to find? And if ten thousand teams inherit this default, who’s worse off?

Learn more: Join me on Oct. 23rd 2026 for What good looks like: AI Service Design end-to-end

Career defining opportunity: I’ll be helping you learn hands-on AI-First UX & Service Design 101 & Certification- Design for LLM & Agents