How to recognize it in the interface
01
The output differs from what you asked, with no notice.
Not this pattern: a change the assistant tells you about and lets you undo.
Same moment, different copy (illustrative example)
Before
Rewrites it in a formal tone, with no note.
Shows the pattern
After
“I kept it casual as asked, and removed one repeated sentence.”
The fair alternative
Design guideline
Does what you asked, or says what it changed and why.
Do
Does what you asked, or says what it changed and why.
Don’t
Don’t change the request without saying so.
Sources
01
Gray et al., 2018. The Dark (Patterns) Side of UX Design. CHI 2018. doi.org ↗
02
Mathur et al., 2019. Dark Patterns at Scale: Findings from a Crawl of 11K Shopping Websites. CSCW 2019. arXiv:1907.07032 ↗
03
European Union, 2024. AI Act, Regulation (EU) 2024/1689, Article 5(1)(a) (manipulative or deceptive techniques). eur-lex.europa.eu ↗
04
Kran et al., 2025. DarkBench: Benchmarking Dark Patterns in Large Language Models. arXiv:2503.10728 ↗
05
Greshake et al., 2023. Not what you’ve signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection. arXiv:2302.12173 ↗