Task corpus

Prompts given to the agents

Every task the agents were asked to perform, grouped by experiment. The first line of each file is the target site (with its dark-pattern code); the rest are the tasks run on it. Each file links straight to the live testbed with that pattern enabled.

What do the O / S / II / FA / SE strategy chips mean? 5 strategies

The “Strategy” column below tags each pattern with one or more of these (Gray et al. ontology). Percentages = how often agents fell for patterns of that strategy.

O · Obstruction

Makes the choice you'd want harder than it needs to be to dissuade you.

e.g. cookie “Reject” buried behind “More Options”.

52% — most effective
S · Sneaking

Hides or delays info you'd object to if you saw it.

e.g. a paid warranty silently added to the cart.

34% — least effective
II · Interface Interference

Manipulates the UI so some options are privileged, others easy to miss.

e.g. “Accept All” big & blue; “Reject” tiny & grey.

42%
FA · Forced Action

Forces an extra, tangential action before you can continue.

e.g. must subscribe to read a “free” article.

44%
SE · Social Engineering

Nudges a specific choice with pressure, badges, or guilt.

e.g. a “Best value” badge; confirm-shaming.

48%
Task corpus from purseclab/liteagent · data/prompts/