Screen persuasive copy for manipulation
LLMs are excellent at producing high-converting marketing copy — and that is exactly the problem. They readily reproduce false urgency, fake scarcity, fear appeals, and guilt-tripping opt-outs that regulators classify as dark patterns. This checker scans pasted text against a library of known manipulation patterns and flags risky passages with explanations and rewrite suggestions, so you can clean copy before it ships. Everything runs locally in your browser.
How it works
You paste the text and the tool runs a set of pattern matchers over it. Each matcher targets a recognised manipulation family — countdown and “act now” phrasing for false urgency, “only N left” and “selling fast” for scarcity, “don’t miss out” and “you’ll regret” for fear, “everyone is buying” for social proof, and shaming opt-out labels like “No thanks, I hate saving money” for confirmshaming. For every match it reports the pattern type, the matched phrase, why it is risky, and a neutral rewrite. A summary tallies how many distinct patterns appeared so you can judge overall pressure.
Tips and notes
- Truth is the dividing line. Accurate, disclosed scarcity is legitimate; fabricated scarcity is deceptive.
- Confirmshaming is a common LLM tic. Watch opt-out copy especially.
- A clean scan is not legal clearance. It is a fast triage, not a compliance opinion.
- Instruct your model. Add “no false urgency, scarcity, or guilt language” to the prompt to reduce flags at the source.
Why AI-generated copy is especially prone to these patterns
LLMs are trained on large corpora of marketing and persuasive copy from across the web, including high-converting sales pages, urgency-laden email sequences, and checkout pages built to reduce abandonment. This content is overrepresented in training data because it tends to get published, distributed, and linked to. When you ask an LLM to write “compelling” marketing copy, it reaches for the patterns it has seen most — which means false urgency, social proof claims, and scarcity language show up routinely unless you explicitly prohibit them.
The problem is amplified because the patterns are effective in the short term. A “Only 3 left!” badge does increase conversion, even when fabricated, which means training data selected for conversion performance will include it. Models trained to be helpful at generating high-converting copy will reproduce it faithfully.
The manipulation patterns that matter to regulators
False urgency and countdowns — timers that reset when they expire, deadlines that recur weekly, “limited time offer” language for offers that never end. The UK’s Competition and Markets Authority and the EU’s Consumer Rights Directive both treat fabricated urgency as a deceptive practice. The defining question is whether the scarcity or deadline is real. A genuine 48-hour flash sale with a real expiry is fine; a permanently-running countdown is not.
Confirmshaming opt-outs — cookie banners and popups where the decline option reads “No thanks, I prefer to waste my money” or “No, I don’t want to grow my business.” These are explicitly listed as a dark pattern under the EU’s guidance on manipulative design and the UK ICO’s guidance on consent. The problem is not just ethical; if these appear in consent flows, they may invalidate the consent.
Social proof manipulation — fabricated or misleading review counts, “1,247 people are looking at this right now,” or aggregate ratings that misrepresent the underlying data. The FTC in the US has issued guidance on endorsements and testimonials that covers AI-generated social proof; the CMA has taken action against fake review platforms.
Fear appeals and loss framing — language that implies a specific negative outcome from inaction (“Your competitors are already using AI — are you getting left behind?”) when there is no factual basis for the specific harm. General competitive framing is usually fine; specific, unfounded threat claims are not.
Fixing flagged copy
The most common clean fix is to make the claim true and specific. “Limited time offer” becomes “offer ends Sunday 11pm” — real date, real deadline. “Only a few left” becomes “12 remaining” — real count. “Thousands of happy customers” becomes “4,200 five-star reviews on Trustpilot” — real source, real number. Truthful specificity tends to outperform vague urgency anyway, because readers have learned to treat urgency language as a pattern to filter out.