@elonmusk@simonw Additional observation: GPT-6 showed the same source-attribution failure with different content. It recalled facts correctly but misidentified their origin and claimed verification without evidence. Controlled replication is still needed.
Couldn't sleep, so I tested ChatGPT's Chat and Work modes.
Work made up a Japanese phrase. Chat later claimed it had introduced that same phrase.
In reverse, Work attributed a different phrase to Chat via a conversation-search summary.
@elonmusk Would Grok do the same?
@elonmusk@simonw A curious ChatGPT source-attribution error: Work coined a Japanese phrase, but Chat later claimed it was its own. In reverse, Work correctly attributed Chat's phrase via a search summary. English translations above. Could this be a provenance issue?