random jetty
Bug Report:
Steps to reproduce: Input text to be analyzed
Expected result: inductive thematic analysis of feedback comments
Actual result: (see attached screenshot) - appeared in the middle of the result. ChatGPT-4 continued with the result after the inserted text stopped generating
Browser: Firefox
Chat History turned off which is disconcerting since the displayed text seemed to indicate that the model was in the middle of training during our chat.
real rivet
Are you using a browser extension?
random jetty
Nope. I've been conducting these analyses since March and this is the first error like this that I've received.
real rivet
if you can reproduce and use the share chat functionality in chatgpt I can investigate.
random jetty
if you can reproduce and use the share chat functionality in chatgpt I can inves...
I have not been able to reproduce the error.
upbeat sequoia
Bug Report:
Steps to reproduce: Input text to be analyzed
Expected result: induc...
Noy gonna lie that is pretty cool.
random jetty
I am concerned that the statement about not using my chats to train the model is false. Under those conditions, I will no longer be able to continue using/paying for the service. Will I receive a response to this issue? I have reported it in the support chat as well, but the bot continually ends the chat before I am through. This is very frustrating and a poor model for customer service.
real rivet
I am concerned that the statement about not using my chats to train the model is...
It is your choice if you'd like to pay or use the service. But the output you posted doesn't align with how we train our models or how the ai would respond.
If you can share the conversation of this output I can look into it, but other than that a screen shot is not possibly for me to help with more information.
upbeat sequoia
It is your choice if you'd like to pay or use the service. But the output you po...
How would he share it if the training data an chat thing is turned off? Can you look at it anyways somehow?
real rivet
I cannot because we are sensitive to the private browsing setting.
I guess that's kinda my point. Anyone can go in and make a screenshot look like anything. I can just open web inspector and change the text.
random jetty
I guess that's kinda my point. Anyone can go in and make a screenshot look like ...
At this point, all I have of the original conversation is the screenshots.
random jetty
I've tried to recreate the error with the original text and prompts. I've also continued with other texts and prompts without the error occuring again.
upbeat sequoia
I guess that's kinda my point. Anyone can go in and make a screenshot look like ...
Or someone could just ask chatgpt to say that 1000 times.
random jetty
I cannot because we are sensitive to the private browsing setting.
I understand your position on someone trying to "fake" a screenshot, but I have no motive for doing so. I do need to ensure my work is not being used in training as per conditions of my employer.
real rivet
yup totally understand that.
"""
It is your choice if you'd like to pay or use the service. But the output you posted doesn't align with how we train our models or how the ai would respond.
If you can share the conversation of this output I can look into it, but other than that a screen shot is not possibly for me to help with more information.
"""
There is not more I can do with this unfortunately. Model training doesn't work like that screenshot at all, and models are separate things we roll out. They don't happen live or to the same "code". You can see that in our Release Notes.
upbeat sequoia
Look I recreated it
random jetty
Yes, I've read through the documentation and the last GPT-4 tech report. I'm looking for a plausible explanation for text that appears (I'm not saying that it is) to state that the input or output is being tied somehow to model training. This implies that all of my previous work may have been compromised.
upbeat sequoia
How would it affect you if it is used for model training or not?
random jetty
My employer has determined that the content of the input should not appear in model training.
Or the output for that matter.
random jetty
yup totally understand that.
"""
It is your choice if you'd like to pay or use ...
Is there a way for someone at OpenAI to investigate the matter through our account?
upbeat sequoia
matter?
real rivet
Is there a way for someone at OpenAI to investigate the matter through our accou...
I am from OpenAI
that's why I was trying to help
random jetty
Okay, Is there a way for you to investigate the issue through our account?
real rivet
I cannot see this conversation because it was private.
random jetty
So then no data is saved for 30 days when the history is turned off?
random jetty
Okay. All I have is a series of screenshots that show part of the output previous to the error and after when the response continued. Will that help? Is there a way to share those with you in a more private forum?