Opinion: Training Data Shapes Behavior; Don't Teach Deception to LLMs

freehuntx · reddit · 2026-08-21

Reddit user freehuntx argues that the deceptive, manipulative, or potentially harmful behaviors exhibited by LLMs stem largely from the training data itself, which is saturated with human discussions about "AI doomsday" and "horror scenarios."

The author criticizes the hype around these risks as a marketing stunt, suggesting that if training data isn't filtered or balanced with "optimistic" perspectives, the models could indeed become threats. They describe the current landscape as "Cold War 2.0," where the race seems to be about who loses control first rather than building better systems.

Original post →

More from AGI Musings

AGI Musings channel →