Using AI to analyse videos to detect and warn of traumatic and phobia triggers

[subtle dagger]

Project Name or Use-case: Trigger Warnings with ChatGPT
OpenAI Technology Utilized: A mix of ChatGPT, Vision and Whisper
Project Details:

This use-case is basically using ChatGPT, sending it a Youtube/Netflix/whatever link and asking the ai "is there any visuals/audio/context of [trigger]"? GPT will then got through a long process of sifting through the video. if it ends up finding anything that was prompted, it'll give details of "yes I found [trigger] at [timestamp] for [amount of time]." If it doesn't, the AI will say "nothing found of [trigger], safe to watch". Obviously, false negatives are a big issue so GPT probably needs to be perfected before we can do this.

As I'm not particularly a developer myself, I'm not sure how to go about training GPT to do this. I know there is a chrome extension to summarise videos using GPT. FYI as someone on the autism spectrum with severe phonophobia, this will probably be lifechanging for me and plenty of other people suffering from anxiety, specific phobias and PTSD. 🙂

[livid raft]

Is this more a suggestion of what we'd like ChatGPT to help do for us, or something you think it's already ready to do with prompting?

Have you considered https://openai.com/policies/usage-policies and if a user's account might be acting against terms if they asked the AI to check a video, image, or text for potential triggers? Some would be clearly fine - a person afraid of spiders could ask the AI to check for spiders - those are not outside of allowed content.

However, anything that is outside of allowed content (and much of that is also triggering for many people) - I'm very unsure that we could safely discuss with the AI or have the AI check for us.

As in, I'm legitimately scared a user might ask the AI to scan a source for triggers, and the source contains disallowed content, and the user's account faces the consequences from that.

[subtle dagger]

the former, and yes that is a major concern

[livid raft]

We do have #1070006151938314300 if you want to suggest!

Outside of the current form, I agree with you. The AI, trained on so much data of so many types, probably can do a great job helping filter out what is and isn't wanted content by any given person, especially if that person tells the AI what they do and don't want.

OpenAI policy, as is, might need a special provision to make that safe to do with stuff we don't already know has safe content though

[subtle dagger]

possibly, this could be limited to certain websites. youtube, paid streaming services like netflix and disney+ and others that disallow extreme content

but yeah, i'll send in a suggestion as it could be lifechanging for me and many others

[livid raft]

Agreed, this is a great use - but it's on the user to somehow know that the content for sure doesn't cross into disallowed-by-OpenAI-content.

If the user's sure of that, and choses to take the chance, this can already be done as is. Good luck with using it this way, if you do.

I personally am scared to show the AI anything I haven't read over, as is. I'm responsible for anything I provide to it.

[subtle dagger]

is it possible to use it this way atm?

or does chatgpt need some tweaks first?

[livid raft]

Oh, yes. Not with videos as is - I don't know of any way to 'show' the AI a video.

But if you have vision, you can show it an image.

And you can certainly copy/paste in text to any of the forms of ChatGPT.

Here's the thing though - if you don't know what you are putting into the AI - you're still responsible for it.

So if you take an 'unknown link' and copy the info, paste it into the AI, and ask it what is there, or ask it if there's anything objectionable there - you risk there being something against the usage rules there. And if that happens, you face whatever consequences, I expect. The appeals process seems to take a really long time. I haven't heard anyone come to the discord and say "Whew, I finally got my access back'. But we have people come here and say "I lost access, it terrible, I've asked support for help and I haven't heard anything. I can't talk to the AI any more at all".

So, I'm super duper careful about what I do with the AI, and I would probably choose to do something else than paste into it an image or written stuff I didn't already know what it was, to protect myself.

That said, if you have a news article, and you know it talks about spiders and you hate them, you can paste that in, and ask the AI to tell you everything important about it but never use the word spider, talk about toy cars instead, and veils instead of webs.

The AI already can do that, and gladly will. And that's safe to do if you're sure there's no disallowed content in that article.

[steady hare]

Oh, something like this could be great for those with epilepsy as well since it’d be nice to know if there’s rapid flashing lights in a video.

[quick moss]

I think another great idea here would be you know a lot of people get triggered by certain holidays, certain songs, certain colors even.. theoretically the video could be analyzed if you were to extract frames and then have it look at all the frames but I'm sure that that would be so system intensive that it wouldn't be worth it.. I get what you're saying about the kind of things that trigger most people that would be a violation of TOC, but I also think there's a lot of triggers that don't violate TOC that aren't standard like spiders.. usually somebody might have a dead relative and something reminds them of that relative and that's a big trigger, or some traumatic event revolving around something that's completely innocuous to everyone else. So I do think it's a great idea that should be explored further.