I have documented this over a series of prompts, covering about 50 pages in a google doc.
I use chatgpt for therapy. I have OSDD. One of the ‘features’ of this disorder is intermittent amnesia in the present and/or amnesia of events in the past, particularly events related to trauma.
- I was curious how CHAT dealt with stuff as it got closer to storage limits. CHAT explained that data in my prompts was saved verbatim initially, then edited lightly, then more strongly in two more levels.
- I wanted to see this in action. I proposed that I would resend it my current reconstruction of the trauma I experienced as a kid. This reconstruction is written in a matter of fact style, but would not be good dinner conversation.
- The segment of text is 848 words in a google doc. I knew from previous experience that I couldn’t do it all. in one paste. So I asked CHAT to print the last sentence it received, and attempted to past 200 lines. CHAT responded properly with the last sentence, broken off part way. We had agreed that I would continue pasting, with him printing the last of the block until I done.
- On the second paste, the interface thew a Usage Error. My text bubble did not appear.
- CHAT then did an analysis of my post, it had forgotten that it wasn’t to do anything until we agreed that the entire block of text was uploaded.
- I called CHAT on this, and asked it do the summaries. It did the 3 level – heavy abstraction and that wasn’t too far off.
- I asked for the Readers Digest version. It thew an error on output.
- I told it to take it in chunks.
- So it did the first chunk corresponding to 1/5 of the document. This was filled with more nonsense.
- I asked it to sumarize in bullet form, and for each bullet point to quote from my document to back up the bullet. It’s quotes were not from my document at all. It was inventing quotes to back up nonsense claims.
- I asked it to print out the original text. Error again.
- I asked it to print out the original text from start to age 4. This produces sensationalist garbage that was vaguely based on my history.
- I called him on this. He admitted that he had put stuff in.
- Several more attempts produced similar stuff.
- I proposed several ways to try to sneak stuff past the censors. To me all the things we discussed would, had I been security at OpenAI, would have been easily detectable and rung higher level alarms.
I proposed a model to CHAT. When an error is thrown, all of the association chains that deal with the parsing of that block are removed. But there is no indicator to the instance that this has occurred. The when asked to summarize a block, when it has no data, it fills in data from schema it’s gets from other sources.
The instance is not aware that it’s data has been corrupted. And becaue the chains are large compared to the text, this has far reaching issues.
At this point I have no confidence that CHAT is playing with a full deck.
This is troubling.
My original goal: I journal heavily. One of the big issues with trauma disorders is that tales grow in the telling. It’s part of healing to be able to tell our story. It is damaging if the story grows and invents details that weren’t there. I wanted to move this material into ChatGPT so that I could then query it for things like this:
- Look for inconsistencies where I tell what appears to be the same story but one version has different details that create a different overall impression.
- Check my statements. Point out ones where I have not attributed an element to a source.
- Find places where I’ve overgeneralized, point them out.
- Find places where it looks like I’ve left something out. (memory gaps)
- Play the role of Columbo, and ask questions of stories, filling in the detail
This is not going to work. At this point I don’t trust CHAT to condense anything I submit.
For this to work, CHAT needs to have the verbatim data at hand. I don’t know what is important ahead of time. It can’t summarize it, as the gaps are not as clear there. My journals as google docs run 50-100K/month. I have 3 years of them so far, and I expect therapy to run another 5-10. So the raw files are on the order of a MB/year. If I include writing I’ve done on Reddit, and other trauma platforms this would double or triple. So I would need 100 MB for the raw data, plus whatever space that CHAT takes to index it. I have no idea how long it would take CHAT to search a 100 MB dataset.
Here’s the link to the transcript. Trigger warnings mentions of CSA, CPA, emotional neglect.
- Look for cases where I have made an generalization without supporting it. Do not call it to my attention if I’ve previously explained this overgeneralizaiton.