A technical analysis draws parallels between chat-based large language models and cold reading techniques used by psychics, revealing similar mechanisms for generating seemingly personalized responses.
The comparison highlights how both LLMs and psychic cons exploit pattern recognition and probabilistic outputs to appear insightful. Both systems generate high-confidence responses based on limited input, relying on audience interpretation to fill gaps.
Key similarities include:
- Vagueness by design: Both produce statements general enough to apply to many situations
- Post-hoc rationalization: Users interpret outputs through their own context, attributing accuracy
- Probabilistic generation: Responses aren't truly understanding—they're statistical predictions
- Confidence masking uncertainty: Both present outputs with unwarranted certainty
The analysis doesn't claim LLMs are fraudulent, but notes they share structural properties with known deception techniques. Users interacting with these systems may attribute understanding or insight that isn't actually present.
The observation carries implications for AI deployment in sensitive domains like counseling or medical advice, where the illusion of understanding could prove harmful.
Vercel, Cloudflare, and other platforms are rapidly integrating Jev, a new AI tool that accelerates model selection while reducing costs. TypeSafe's Jev matches the performance of GPT-5.6 and Claude Sonnet 5 on workflow evaluations.
Step Fun has previewed Step 5, its next-generation model aimed at pushing the efficiency frontier in AI performance. The announcement details improvements across multiple capability dimensions.
Tencent unveiled Gander, an AI assistant that processes speech, images, and text while simultaneously handling background tasks. The system separates conversation management from complex work, allowing users to maintain dialogue without interruption.