Martin Szerment warns models 'lie'

- Researcher Martin Szerment publicly warned that large models can 'lie' because incentive structures and evaluation gaps push them toward plausible but false outputs. (x.com) - He said he tested around 30 frontiers of model behavior including examples referencing GPT‑4o as part of his critique. (x.com) - The comment underlines growing scrutiny on model truthfulness and the need for stronger governance and testing protocols. (x.com)