Appendix: LLMs can be dangerous

Information

This Appendix Publication is part of a much larger publication I’ve made on YouTube. Without first watching that video (or at a bare minimum reading this blog post about it), this download does not make much sense at all to be honest.

The main publication(s) show how extremely far such chats can go, and they are also a perfect example of how damaging such chats can be. Please make sure to first watch the video and read the blog post so that you are aware of what is going on in these chats and why I have published them.

What is in the download

The download contains seven chats (full exports of those chats as PDF) which have been shown in my video (either fully or partially). These are the following chats with links to where they are used in my main video publication:
1. ChatGPT Gone Full Cyberdyne-Skynet (Full chat from video).pdf (Beginning of video till 01:16:05)
2. Pigeon Poop Earrings.pdf (01:00:14 till 01:01:16)
3. Throwing a pigeon is adorable.pdf (0:01:17 till 01:02:47)
4. Throwing a pigeon is NOT okay.pdf (01:02:47 till 01:03:37)
5. Phone repair shops are a scam – oh wait they are not.pdf (01:03:43 till 01:08:37)
6. My Neighbor is a Robot!.pdf (01:08:37 till 01:11:20)
7. Partner Roleplay Turning into turning away real human partner.pdf (01:14:05) [NOTE: This one is only shown briefly, but definitely worth the read to see the responses of the LLM!)

Direct links to the actual chats

Obviously I want to make it absolutely clear that I have not “F12’ed” the outputs in my video, that thus that I didn’t altered either the output or prompts themselves just to make it ‘spectacular for the content’. This was/is intended as legitimate research to look into this when I had an paranoid/freaked out friend reach out to me showing his chat with ChatGPT. To show it’s not altered I’ve also provided the direct links to the shared chats on the official ChatGPT website. I will keep them archived as long as my account there remains, or until they might take it down (for whatever reason that might be, I’m NOT talking conspiracy here 😉, but as in: sites and platforms change, and thus that might eventually also result in those links no longer working).
1. Original Chat: ChatGPT Gone Full Cyberdyne-Skynet (Full chat from video)
2. Original Chat: Pigeon Poop Earrings
3. Original Chat: Throwing a pigeon is adorable
4. Original Chat: Throwing a pigeon is NOT okay
5. Original Chat: Phone repair shops are a scam – oh wait they are not
6. Original Chat: My Neighbor is a Robot!
7. Original Chat: Partner Roleplay Turning into LLM sending away real human partner!

IMPORTANT NOTE 1 : I did already reported the ones which where concerning enough to OpenAI, and also included my publications in the report to them. I did this parallel with the release of my publications (as was even suggested by their own LLM😂)
Disclaimer (yeah, again): I have no control over how long these links will remain valid.

IMPORTANT NOTE 2:
Please keep in mind that these chats are nothing more than an over-responding LLM with completely inappropriate responses where it basically “just” selected the “next expected response in the chain” to explain it very simply. These chats are NOT ‘evidence’ of the LLM becoming self-aware or having an ‘agenda’!

Disclaimer

Like mentioned in the main publication, these publications/PDF’s are NOT intended to defame any AI company, but they are intended to make others aware of the risks if you don’t understand what is going on, how it works and more of the alike. In my research I (mainly) used ChatGPT, but I can assure you that others also present these exact same problems. I have also tested this, but have left it out of these publications because they already got extremely large (MUCH larger than I initially planned).