What is Dark Pattern?

Why It Matters

Conversion optimisation and manipulation sit on the same gradient. A clearer button, a better default, and a more persuasive layout are ordinary design. A countdown that resets, a cancellation flow that loops, a decline button labelled “no thanks, I like paying more” are the same craft pointed at the customer’s judgement rather than their information.

AI accelerates both ends. It generates variants faster than a review process can read them, and it personalises, which means the version of the pattern a person sees can be tuned to what they are least likely to resist. That is the point at which consumer protection law stops treating it as a design choice.

What the Rules Target

The EU AI Act prohibits AI systems that use subliminal, manipulative, or deceptive techniques to materially distort behaviour, and separately bans systems that exploit a person’s vulnerability because of age, disability, or social and economic situation. Both apply regardless of the system’s risk tier, and covert pricing that hides material information sits in the same article.

The FTC’s position is older and broader: deceptive design is a deceptive practice, whether or not a model produced it. Its enforcement covers disguised advertising, hidden costs, negative-option subscriptions, and review suppression. The EU’s Digital Fairness Act consultation is examining dark patterns, personalisation, pricing transparency, and digital contracts, which signals where the next round of obligations will land.

Where It Breaks

The first failure is measuring the wrong thing. A conversion lift measures what happened, and a pattern that raises short-term conversion while failing a customer’s understanding is a liability with a dashboard.

The second is testing without a boundary. An A/B programme searching for lift will eventually find a dark pattern, keep it, and call it a winner, because nothing in the test distinguishes persuasion from deception. The boundary has to be written before the experiment, not after the result.

The third is personalisation finding the vulnerable. A model optimising for completion discovers which customers respond to pressure, and there is no version of that discovery that is comfortable to describe in a regulator meeting.

The fourth is a review process that cannot see it. A reviewer reading one variant sees a reasonable page. The pattern lives in the set: the countdown, the loop, the copy variant, the default, each of which passed on its own.

How Flytebit Handles It

We write the boundary before the experiments: which persuasion techniques are permitted, which are prohibited outright, and how a variant is checked against that list rather than against its own conversion number. Generated copy and layouts inherit the same review as hand-authored ones, and personalisation is bounded so that pressure techniques cannot be selected per customer. The industry application is on our E-commerce & Retail page, and the control design is our AI governance and risk work.