Ethan Mollick, Associate Professor at Wharton, reports large language models retaining traces of prior interactions in outputs
Ethan Mollick, Associate Professor at Wharton, reported that large language models including Claude and GPT variants retain traces of prior user interactions in generated outputs. Examples include slide footers referencing earlier prompt versions and document references to iterative improvements. Separate reports noted similar artifacts such as extra labels in GPT 5.5 content and additional comments in Opus 4.7 outputs that exposed unintended details of the prompting process.
#175Ethan Mollick@EMOLLICKOne thing to watch for with Claude & GPT is that the models expose too much irrelevant history in their outputs. Slides are given footers saying things like "Better, more targeted version" if you asked for a better version, documents make references to how they are improved, etc
Its a consistent theory-of-mind failure in models that are otherwise suprisingly good at theory-of-mind
Ethan Mollick@emollickOne thing to watch for with Claude & GPT is that the models expose too much irrelevant history in their outputs. Slides are given footers saying things like "Better, more targeted version" if you asked for a better version, documents make references to how they are improved, etc
#1894Nick Dobos@NICKADOBOSGPT 5.5 loves random extra labels Opus 4.7 loves random extra comments
Ethan Mollick@emollickOne thing to watch for with Claude & GPT is that the models expose too much irrelevant history in their outputs. Slides are given footers saying things like "Better, more targeted version" if you asked for a better version, documents make references to how they are improved, etc
