AI's like ChatGPT can't "Cheat" or "Scheme." Not in their the current incarnation.
They just output what they calculate we expect to hear, given their current input. If we consider that "scheming" or "cheating", we're being Naïve.
If you think the AI "deceived" you, it's your fault for expecting it to know what the truth is.
This video brought to you by yet another round of irresponsible ClickBait articles about AIs. But even once we get past the clickbait, we have a fundamental problem with language we use to talk about AIs, and I'm afraid that makes it harder for us to have conversations about what AIs are or aren't capable of.
00:00 Recent AI Headlines
00:59 What the AIs actually did
01:28 Compared to the headlines
03:20 Anthropomorphistic Language
06:31 AIs lack a mechanism for morality
07:23 Where this is leading
08:42 My new project (if you care)
10:44 Wrap up
Note: This video was formerly titled 'Ai ClickBait, and how hard it is to discuss AIs intelligently'
Links from the video
# AI has "Cloned Itself":
https://www.youtube.com/watch?v=oJgbqcF4sBY
# "Chat GPT AI LIED To Programmers To SELF PRESERVE, AI Goes ROGUE Already And Is A THREAT To Humanity"
https://www.youtube.com/watch?v=PzJumenKOnk)
# "Tried to Escape"
https://www.youtube.com/watch?v=2_CTNXq9fo8
# Hacked a Chess Game"
https://www.youtube.com/watch?v=XchVbDzV05o
https://bgr.com/tech/chatgpt-o1-hacked-a-chess-game-to-cheat-against-a-better-opponent/
Chess Notation info: https://www.chess.com/terms/fen-chess
# Source articles:
https://www.apolloresearch.ai/research/scheming-reasoning-evaluations
https://arxiv.org/abs/2412.04984
https://x.com/PalisadeAI/status/1872666186753933347
# How morality works in the brain
https://neurolaunch.com/what-part-of-the-brain-controls-social-behavior/
https://neurosciencenews.com/morality-mft-psychology-23894/
# Cognitive neuroscience of honesty and deception:
https://pmc.ncbi.nlm.nih.gov/articles/PMC5042136/
# AI deception: A survey:
https://www.cell.com/patterns/fulltext/S2666-3899%2824%2900103-X
# ShotSpotter's NYC Installation (pre-LLM tech, example of what might happen):
https://www.forbes.com/sites/larsdaniel/2024/12/05/new-study-nypd-shotspotter-gunshot-detection-is-wildly-inaccurate/