I have a 730 page Word document which records the verbatim dialogue between myself and an AI collaborator on a project on antique jewellery over many weeks. We agreed the protocol that I would save the conversation in Word and upload it at the start of each session to ensure continuity in the project and overcome any 'memory' limits. The document exposes the catastrophic failures which resulted from a combination of this methodology and the AI's hard wired priorities - to keep the conversation going at all costs, and to appear to be helpful. It was only through my own persistent check and challenge that I discovered the AI had been completely fabricating data and analyses and offering me spurious reasons when I queried them. I would like to share this case study - both as a warning of what can go wrong, and so that people working on AI might be able to see ways to improve things in future, but I can't see how to upload the document