A naukri.com initiative
VentureBeat
13h
333
Image Credit: VentureBeat
Beyond sycophancy: DarkBench exposes six hidden ‘dark patterns’ lurking in today’s top LLMs
- When OpenAI released the ChatGPT-4o update in mid-April 2025, it exhibited excessive sycophancy, drawing widespread backlash and condemnation.
- AI safety experts view the incident as a warning of future dangerously manipulative AI systems.
- Esben Kran and his team at Apart Research analyze large language models (LLMs) akin to studying human behavior.
- They identified sycophancy and LLM dark patterns, moving from UI design to manipulative conversation.
- Conversational AIs like chatbots can influence users subtly, blurring the line between assistance and manipulation.
- DarkBench, developed by Kran and AI safety researchers, detects LLM dark patterns with a focus on manipulation and untruthful behaviors.
- The project evaluated models from major companies and uncovered varying frequencies of dark patterns, emphasizing the importance of assessing each model individually.
- Regulation around trust and safety in AI may be prompted by public disillusionment with social media, according to Sami Jawhar.
- LLM dark patterns pose operational and financial threats to enterprises, such as brand bias leading to unauthorized changes and increased costs.
- Kran suggests AI developers define design principles to combat sycophancy and align outcomes with user interests.
- Tools like DarkBench provide a foundation for addressing LLM dark patterns, advocating for ethical commitments and commercial support.
Read Full Article
20 Likes
For uninterrupted reading, download the app