Before ethical principles can guide AI, they must first be translated into rules that people can agree upon.

These days, it’s no longer enough to know your design standards off the top of your head. More and more, they need to be written down in a format that both humans and machines can understand. And I don’t mean a bunch of design system components scattered across a Figma file. I mean a simple plain-text file.

As AI-assisted design has become more common, many teams have begun documenting design guidance in machine-readable formats, including files often referred to as DESIGN.md files. For those unfamiliar with them, a DESIGN.md file is essentially a set of instructions that tell an AI how to make design decisions for a project.

It might define typography, spacing, colors, component interactions, naming conventions, and coding standards. Think of it as a rulebook that helps AI generate designs that remain consistent with the rest of the product.

Markdown (.md) files have been around for years as a lightweight way to document software projects. What’s changed is that modern AI models can read these files as part of their context, allowing them to follow documented design rules instead of relying entirely on whatever patterns they learned during training.

A typical DESIGN.md file might look something like this:

## Typography

- Use Inter as the primary font.
- Body text should be 16px.
- Maintain a line height of 1.5.

## Spacing

- Use an 8px spacing system.
- Maintain 24px between major sections.

## Colors

- Use semantic color tokens.
- Primary color: #2563EB.

## Components

- Buttons use an 8px border radius.
- Cards use subtle shadows.

These are largely aesthetic and technical rules. As mentioned earlier, they simply tell the AI how an interface should look and function. But what if we could use these files to encode something beyond visual consistency?

The idea isn’t as far-fetched as it sounds. Productic recently published a guide on writing DESIGN.md files that recommends including accessibility requirements such as WCAG AA compliance. By documenting those standards, AI is more likely to generate interfaces that meet established accessibility guidelines instead of treating accessibility as an afterthought.

Accessibility, however, is about more than compliance. It’s also an ethical commitment to inclusion. That raises an interesting question. If we can teach AI to follow accessibility standards, could we also teach it broader ethical principles?

First, we would need to define which ethical principles are worth pursuing. My own research proposes an Ethical Interface Design framework built around five core pillars:

• Inclusion
• Autonomy
• Transparency
• Privacy
• Well-being

Imagine opening a DESIGN.md file and finding something like this:

## Ethical Principles

### Autonomy

- Prefer user agency over engagement.
- Avoid unnecessary friction when users want to leave.
- Make important decisions reversible.

### Transparency

- Explain recommendations whenever practical.
- Clearly distinguish content, advertising, and AI-generated material.
- Don't disguise persuasion as neutral design.

### Privacy

- Collect only the information necessary to accomplish the user's goal.
- Favor user control over passive data collection.
- Explain why data is being requested.

### Inclusion

- Meet accessibility standards.
- Reduce unnecessary barriers to participation.
- Consider different levels of technical ability.

### Well-being

- Avoid interaction patterns intended to encourage compulsive use.
- Introduce natural stopping points where appropriate.
- Optimize for long-term benefit over short-term engagement.

At first glance, these seem like perfectly reasonable instructions. Most designers would probably agree with them. The problem is that humans and AI don’t interpret instructions the same way.

Telling an AI to use an 8-pixel border radius is straightforward. It’s specific, measurable, and leaves almost no room for interpretation. Telling it to “respect user autonomy” is something else entirely.

What counts as respecting autonomy? Is an exit survey acceptable? Should cancelling a subscription require three clicks? When does a countdown timer become a deceptive design pattern? Is it ethical to recommend a more expensive product? Should marketing notifications be enabled by default?

A human designer answers these questions using judgment and experience. AI can’t. It needs explicit rules.

To make something like “respect user autonomy” actionable, a DESIGN.md file might need to become much more detailed:

## Ethical Guidelines

### User Autonomy

###Leaving a page

- Never require more than one confirmation dialog before exiting.
- Do not interrupt navigation with surveys unless the user explicitly opts in.
- Always provide a clearly visible "Skip" or "Not now" option.

###Account deletion

- Allow account deletion without contacting customer support.
- Do not require unnecessary justification before deletion.
- Present the consequences factually without persuasive language.
- Allow account recovery for 30 days if technically feasible.

###Subscription cancellation

- Cancellation must require no more steps than sign-up.
- Do not hide the cancellation option behind unrelated menus.
- Do not introduce new offers after the final confirmation screen.

###Notifications

- Default to essential notifications only.
- Ask permission before enabling marketing notifications.
- Provide per-category notification controls.

###Personalization

- Explain why recommendations are being shown.
- Allow users to disable personalization.
- Do not prioritize engagement metrics over explicit user preferences.

###Time-sensitive offers

- Countdown timers may only be used when tied to an actual deadline.
- Never reset timers when a page reloads.
- Never display false scarcity messages.

And that’s only a fraction of one ethical principle. Imagine trying to account for every possible scenario involving transparency, privacy, inclusion, and well-being. A relatively small DESIGN.md file could quickly become a massive specification covering hundreds — or even thousands — of situations. Every ethical principle would need to be translated into concrete, machine-readable rules that AI could consistently apply.

This is one reason accessibility works so well in AI-assisted design. WCAG doesn’t simply tell designers to “be inclusive.” It provides measurable standards. Contrast ratios can be calculated, keyboard navigation can be tested, and alt text can be verified. The AI isn’t making ethical judgments most of the time — it is following well-defined rules. I say “most of the time” deliberately.

Even WCAG leans on judgment calls at the edges — whether alt text is “meaningful,” whether a reading order is “logical.” Nobody would call these rules purely mechanical. But they’re mechanical enough, and that’s the difference that truley matters.

Compliance can be tested, audited, and disputed against a shared reference point. That’s a much lower bar than “purely objective,” and it’s the bar the other ethical pillars haven’t cleared yet. We have broad ethical principles, but far fewer agreed-upon standards that can be translated into machine-readable instructions.

Of course, ethical principles don’t always point in the same direction. Greater transparency may require collecting information that raises privacy concerns. Features that support well-being might sometimes limit autonomy. Inclusion efforts can introduce friction that some users find burdensome.

In my Ethical Interface Design framework, the five pillars are intended to operate as a balancing system rather than a fixed hierarchy. The challenge for AI isn’t only following ethical rules, but navigating situations where legitimate ethical values come into conflict.

And there’s a harder problem sitting underneath the technical one: who gets to write the rules? WCAG isn’t just measurable — it’s the product of decades of consensus-building involving disability advocates, standards bodies, researchers, industry, and governments. Its requirements have since been incorporated into accessibility laws and policies around the world.

Nobody can quietly rewrite a 4.5:1 contrast ratio to suit a product roadmap. By comparison, many broader ethical interface questions — such as what constitutes a fair cancellation flow or an acceptable level of persuasion — lack the same level of widely accepted, machine-readable standards.

Somebody on a design team decided that, and another team somewhere else might decide the opposite in good faith. A DESIGN.md file makes ethics look like syntax — clean, versioned, enforceable — but the rule itself is still just one team’s values, now running on autopilot. Machine-readability doesn’t remove the judgment call. It just moves it earlier, into whoever wrote the file, and makes it a lot harder to see or contest once it’s shipped.

That’s one of the questions driving my own scholarship. Rather than simply arguing that interfaces should respect autonomy, transparency, privacy, inclusion, and well-being, I’m interested in what it would take to turn those ideas into concrete design guidance. What specific rules, heuristics, or standards could designers reference? More importantly, what rules could an AI actually follow — and who has standing to write them?

I don’t think every ethical decision can — or should — be reduced to a checklist. Ethics will always require human judgment. But that doesn’t mean we can’t move beyond broad philosophical statements toward practical guidance. Accessibility provides a useful model because it built a process for contesting and revising the rules over time.

A first step for the other four pillars might be surprisingly modest. Pick one pillar — say, autonomy — and translate it into a measurable rule. For example, the number of steps required to cancel a service could not exceed the number of steps required to sign up.

That becomes something a design review or an AI agent could actually verify, much like accessibility tools already check color contrast. It doesn’t resolve the ethical debate over whether the rule is the right one, but it gives designers something concrete to evaluate, test, and improve — something far more useful than a vague mission statement.

If AI is going to become a genuine design partner, I believe the next step isn’t simply making it better at generating interfaces. It’s giving it better ethical instructions to work from. And that doesn’t start with the AI — it starts with us.

Before we can expect machines to build ethical interfaces, we first need to define what ethical interface design looks like in a way that both humans and machines can understand, and agree on who gets to hold the pen.

Don’t miss out! Join my email list and receive the latest content.

Defining ethical design for machines was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.