What a machine can and cannot judge in a dark pattern review

Reviewing a financial app for dark patterns is slow work. A reviewer walks through every journey, takes screenshots, compares what each screen does with the 13 practices in the CCPA guidelines and writes up the evidence. With regulators now asking for these reviews, firms want to know how much of the work a machine can do.

A useful way to answer is to look at a product in three ways: what the screens show, what the code does and what the backend logic does when a customer acts. Some patterns leave evidence in one of those places that a machine can find. Others do not.

Patterns that suit automation

Some patterns leave a trace a machine can find reliably.

• Pre-selected add-ons (basket sneaking) show up as checkboxes or toggles that are on by default in the code.
• Drip pricing shows up as charges that appear only in later steps of a flow, which can be found by comparing the price at each step.
• Countdown timers used for false urgency can be found in the code, and the tool can check whether the timer actually resets.
• Unequal cancellation paths can be measured: count the steps to buy and the steps to cancel.

These are structural. The evidence is in the product itself, and a rule can be written to find it.

Patterns that need a person

Other patterns depend on meaning, context or intent.

• Confirm shaming depends on tone. A model can flag emotional language on a cancellation screen, but a person has to judge whether it crosses a line.
• Disguised advertising depends on whether a reasonable customer would recognise a recommendation as paid. That is a judgement about the audience.
• Trick questions depend on how real customers read a sentence, including customers reading in their second language.
• Bait and switch often depends on what was said in a call or an advertisement outside the app.

A machine can find the evidence. A person still has to decide what it means.

How the work should be split

The useful design is not "AI reviews the app". It is a pipeline where the machine does the exhaustive part and people do the judgement.

1. The machine walks every journey, records every screen and flags candidate findings, each with its evidence and the rule it relates to.
2. A reviewer confirms or dismisses each finding, and every decision is recorded.
3. The report shows both: what was flagged, what was confirmed and why.

That split matters for credibility. A regulator or a board will not accept a finding because a model said so. They will accept it when a named reviewer looked at the evidence and agreed.

What changes for the firm

Automation changes the economics of the review more than its conclusions. A manual review is expensive, so it happens once a year at best. A tool that does the exhaustive part can run on every release. That turns dark patterns from an annual finding into a check the product team runs as part of shipping, which is where it belongs.