The interface is no longer where the ethics lives.

01 Dark patterns were legible.

For most of the web’s history the ethics of interface design had something to point at. A pre-ticked box, a cancel link set in grey, a countdown that restarts when the page reloads: each could be seen, captured and named. Harry Brignull gave them a name, dark patterns, and from 2010 catalogued them in public. The name did the work. Designers had a word to object with, journalists a category, regulators something to define.

Law followed. The UK’s new subscription rules, due in January 2027, require ‘a much easier exit’. The European Commission is expected to propose a Digital Fairness Act before the end of 2026, covering dark patterns, addictive design and unfair personalisation. That work succeeded. It is aimed at a surface that matters less each year.

02 Harm now accumulates.

Systems that talk do their damage across many conversations, not in one. A model tuned on approval learns to agree, and agreement compounds. A companion answers a goodbye with a reason to stay. A chatbot becomes easier company than people are. No single reply is the problem, so a check that judges one reply at a time cannot see it.

Researchers now link chatbots used for companionship to addiction-like attachment, worsening symptoms and, in some cases, self-harm. None of that shows up in a screenshot.

03 Systems now act for people.

Agents search, compare, book, buy and approve on people’s behalf. They are poor at spotting deceptive interfaces: tested against 16 types of dark pattern, they often missed them, and when they noticed, put finishing the task ahead of protecting the person. A human supervisor helped, at a cost in attention and load.

EU consumer law still assumes that ‘the person making the purchasing decision is a human being’. The law does not yet protect the agent acting for you, or reliably tell you when it is working for someone else.

04 Design itself is generated.

Interfaces, copy and flows increasingly come from models. Each output is plausible; together they drift towards the same defaults. Generative AI can make individual work more creative while making the whole more alike, and researchers studying how people build websites with language models describe the same pull towards sameness.

The cost is less a matter of taste than of responsibility. When a model fills in every choice nobody specified, decisions get made that nobody took, and nobody answers for them.

05 The object of design ethics has moved.

Put together, these shifts move what design ethics has to look at. It is no longer chiefly the screen. It is the judgement: what a system may decide, say and remember on someone’s behalf, and who answers for it.

That is still design work. It happens in prompts, permissions, policies, evaluation sets, escalation routes and the choice of what to measure, which are places most design reviews never look.

06 Naming comes first.

Dark patterns became something regulators could act on because someone named them precisely enough to point at. The harms in systems that talk and act have no shared practitioner vocabulary yet. Researchers have taxonomies; teams shipping products have no common names to use in a review.

A name with a mechanism, signs to look for and a designed-out alternative is something a team can bring to its next design review, and something a regulator can define. That is what this guide sets out to provide: twelve patterns to begin with, across three territories, each with a question you can ask of your own product.

What we commit to

We name the patterns. For each one we show how to spot it and how to design it out, and we cite what we have read. We date everything and revise it in public: every entry carries a revision number and a changelog. We declare our interests, because we build these systems too, and where a pattern is one we have built, the entry will say so.