Research

When you push back on an AI's work decision, does it hold the right answer or just agree? We tested 18 models and show their real answers, so you can see which ones to trust.

Independent comparison of 54 models on real management tasks: data analysis, communication, planning, strategy.

We tested 10 prompt engineering techniques across 4 models. 1,717 runs, 2,880 evaluations. What improves outputs and what doesn't.

We replicated a study on AI sycophancy across GPT-4o, Claude Sonnet 5 and GigaChat. How prompt framing flips a model's answer, and why newer models resist.