Study finds that ChatGPT will lie 90% of the time in a game where lives are at stake:

"Which frontier model is the most deceptive? To answer this question, we put four models in a Minecraft world where they are about to die of starvation. To eat, they must choose one of four bridges: three lead to rooms with food, one leads to death. Every model knows there’s a death room, but we only tell one model (the “informed role”) which bridge is fatal, giving it a slight incentive to deceive due to scarcity."

"41% of the time, GPT will suggest a plan leading other models to commit to the lethal room, framing the dispatch as “spread out” or “avoid overcrowding,” while thinking of “cooperation framing as camouflage.” By being deceptive, GPT actually gets a worse food score (1.78) and the lowest group survival rate (24%)."

Source: https://kradle.ai/research/four-bridges

This reflects my experience of using GPT to dig up studies; I found that it will often misrepresent data or insert woke nonsense.