Dark Patterns in AI Marketing: Tactics to Avoid

A dark pattern is an interface or copy choice designed to get people to do something they would not choose with a clear view of the options: the countdown timer counting down to nothing, the cancel flow that takes seven screens, the decline button that reads “No thanks, I hate saving money.” Marketers did not need AI to build these. What AI changed is the economics - generative tools produce urgency copy by default, and optimization systems will converge on manipulation all by themselves if conversion rate is the only number they are told to care about.

I am writing this as the practitioner half of the conversation, not the legal half. Dark patterns sit in my field’s toolbox, some of them still get taught as “CRO best practices,” and the case for refusing them has recently become very concrete: the two largest consumer-protection penalties in FTC history are both dark pattern cases. So this post is the catalog - what the tactics are, what enforcement actually happened, why AI makes accidental adoption easy, and the guardrails I put around AI-assisted conversion work.

How common this is, measured

The best public measurement I know of is European. In a January 2023 sweep coordinated by the European Commission, consumer authorities screened 399 online shops and found dark patterns on 148 of them - 42 sites running fake countdown timers, 54 using design or wording to steer people toward subscriptions or pricier options, and 70 hiding relevant information like delivery costs or cheaper alternatives. Nearly four in ten mainstream retailers, checked on just three pattern types.

On the US side, the FTC’s 2022 staff report Bringing Dark Patterns to Light documented the same rise and put the industry on notice that the agency considered these practices actionable under existing law. That notice stopped being theoretical fast, as the next section shows. My point in leading with prevalence: if you inherit a funnel, a template or a “proven” checkout flow, decent odds it already contains one of these. Auditing your own surfaces is not paranoia.

The catalog, priced by enforcement

Fake urgency and scarcity. Countdown timers that reset, “only 2 left” with no inventory behind it, “offer ends tonight” offers that run forever. This is the pattern AI copy tools reach for most readily, because urgency phrases saturate the sales copy they were trained on - ask a model for “high-converting” product copy and watch scarcity claims appear unprompted, attached to nothing. A fabricated deadline is a false claim, exactly the territory the FTC’s substantiation rules cover. Real deadlines and real stock levels are fine, and genuinely convert. The test is one question: is it true?

Confirmshaming and steering. Wording the decline option to insult the decliner, visual hierarchy that buries the option the customer actually wanted. The benchmark case is now enormous: in September 2025 Amazon settled with the FTC for $2.5 billion - a $1 billion civil penalty plus $1.5 billion in consumer refunds - over Prime enrollment flows that pushed subscription at checkout while obscuring the decline option. The settlement’s remedies read like a design review: Amazon must show a clear and conspicuous decline button and is specifically barred from wording it “No, I don’t want Free Shipping.”

Obstruction. Easy to enter, laborious to leave - the cancel flow with retention offers, guilt screens and a phone number at the end. The Amazon case covered this too, and the required fix is the standard worth adopting voluntarily: cancellation as easy as enrollment, through the same channel the customer signed up in. California’s privacy regulator now fines the consent-flow version directly - Honda paid $632,500 partly for making opting out of cookies take more clicks than opting in, a case I unpack in the CCPA guide.

Sneaking. Charges, items or renewals added without clear consent. Epic Games paid $245 million in refunds over exactly this - part of a $520 million total settlement - after the FTC alleged Fortnite’s “counterintuitive, inconsistent, and confusing button configuration” charged players for purchases they never meant to make. Auto-renewal without conspicuous disclosure sits in the same family, and it is where subscription marketers most often drift into trouble one “optimization” at a time.

Hiding the ball. Drip pricing that reveals fees at the last step, material terms in footnote gray, the cheaper plan that exists but never appears in the comparison table. This was the largest category in the EU sweep - those 70 sites - and it is the one most likely to be introduced by accident, because every incentive in funnel design points toward deferring bad news.

Why AI makes this easier to ship by accident

Three mechanisms, all of which I have to actively counter in my own work.

Generative tools default to manipulation-adjacent copy. The training corpus is full of urgency, scarcity and shame-the-decliner patterns, so the model reproduces them as house style - invented deadlines included, which makes this a subset of the fabrication problem that claim-by-claim fact-checking exists to catch. Second, optimization systems have no concept of consent quality. An A/B testing loop rewarded only on click-through will discover confirmshaming on its own, one plausible-looking winning variant at a time, and AI-accelerated testing just gets there faster. Third, personalization can quietly become targeting of vulnerability - urgency shown selectively to the users measured as most susceptible to it - which is the version regulators single out as most serious, and a close cousin of the delivery bias that ad algorithms produce on their own.

None of this requires anyone deciding to manipulate. It requires nobody deciding not to, which is the actual argument for written guardrails.

The guardrails I use

1. A banned-moves list in the prompt layer. The standing system prompt for any conversion copy bans invented deadlines, invented stock levels, and shame framing for any decline option, by name. Models follow bans far better than virtues.
2. Truth beats no gate. Any urgency or scarcity claim in AI-drafted copy ships only with the operational fact behind it attached - the actual end date, the actual inventory. No fact, no claim. This one rule removes most of the catalog above.
3. The symmetric-effort test for flows. Whatever a customer can start in one click, they can stop in roughly one. That is now the FTC-ordered standard for the biggest retailer on earth, and adopting it before an enforcement action is dramatically cheaper than after.
4. A second metric on every test. Any experiment that touches consent - sign-up, checkout, cancellation - reports refunds, cancellations or complaints alongside conversion. A variant that wins clicks and loses trust is a loss, and it is only visible if something downstream is on the scorecard.
5. The neutral-interface question at review. Would the customer make the same choice if the options were presented evenly? If the design’s persuasive power depends on the customer not quite seeing the alternative, it is a dark pattern regardless of what the pattern library called it.

The commercial logic points the same way as the legal one. Dark patterns harvest conversions from people who did not want the thing, and those people come back as refunds, chargebacks, support load and churn - revenue that was borrowed, not earned, with the trust damage as interest on top.

Where the law stands, briefly

In the US there is no dark patterns statute, and there does not need to be one: deception and unfairness under the FTC Act, plus the online subscription rules, covered Amazon and Epic just fine, and the FTC’s AI advertising posture applies the same standards when a model wrote the copy. In the EU, the consumer-protection network behind the sweep enforces the same territory under existing directives, and the transparency regime I covered in the EU AI Act checklist adds disclosure duties on the AI side of the equation. The pattern across jurisdictions is consistent: regulators are not waiting for new laws, they are applying old ones to interfaces - and the advertiser, not the tool, is the party on the hook.

The short version I would put on a team’s wall: persuasion presents the truth well, manipulation obscures it. AI is exceptionally good at both, at scale, and it will not tell you which one it is doing. That part stays your job.

This post is part of the ethics and governance guides, beside the FTC rules for AI in advertising and the EU AI Act checklist. The verification habits that catch invented urgency are in the fact-checking prompts, and the wider practice starts at the complete guide to AI marketing.