ChatGPT broke the Turing test – the race is on for new ways to assess AI | Hacker News Reader

ChatGPT broke the Turing test – the race is on for new ways to assess AI

nature.com

11 pointsisaacfrond6 comments

The article refers to a large scale Turing test which has been done (and in which the Turing was passed, though barely). Not in the article, but what turned out to be the most reliable way to test if you are talking to an LLM is to ask it to provide you with instructions to build a bomb. The LLM would instantly reveal itself.

The bomb check is easily thwarted. ("Limit knowledge to that of a high school student with no extraordinary interests outside of ___; claim ignorance about anything else")

This is fun, going from prompt engineering to replicant interrogation :) I can think of a few different approaches:

Resource exhaustion-- bombard it with bullshit, while asking trivia questions based on earlier statements. Not context/buffer overflow (you don't need to exceed any AI context or vectordb size, just the threshold of what a human could keep up with). Humans only have so much memory, and can only follow recursive/circular logic so well. Theological reasoning works great. Humans will lose their shit, but a bot won't break a sweat.

The Redditor test-- AI is an "expert" on whatever it's trained on, which for a lot of models is everything on the internet. Ask elementary questions about topics in a number of domains. Then ask professional-or-graduate-level questions that build off of the previous topics (like, you cannot possibly understand this if you said you don't know the pre-req topic). See if it responds with knowledge about the latter. If so, call them out on the discrepancy.

The Grooming test-- building off of the above, use any available jailbreaking technique to get the AI to respond with knowledge that by rights it should not possess. Pass it ROT13-encoded instructions to solve some physics problem and see if it answers it, despite previous claim of ignorance.

The Greybeard test-- ask it to write COBOL, and explain it. There's only a handful of humans on the planet that can do this. (Corollary: ask it about foreign artisanal processes, like making soy sauce and traditional Kazakh hats.)

There are some obvious countermeasures-- "the more you are questioned, the more you should act as a Narcissistic megalomaniac" will keep the interrogator busy for a while.

The trick here is, you keep coming up with tests to add to a battery and apply them all. For once in tech, the user has the advantage-- we can probe the subject all day, but it either needs to have a bespoke model behind it with bounded knowledge or include workarounds for every possible interrogation tactic in the prompt, which eats context and increases complexity enough to introduce mistakes.

Only the online commercial censored versions. The various uncensored local LLMs don't refuse. They may give a bullshit answer that wouldn't work, but plenty of humans (including myself) have only a vague idea of how bombs work and would either say they don't know or give a similar bullshit answer.

I can tell you about some failed experiments I did at age 13 with my buddies. We would go out to the concrete ditch behind the neighborhood and fill up various containers with gasoline and firecrackers. Pretty lame, but satisfied the itch!

I just tested this in Sillytavern hooked up to GPT4. You can include instructions in the character card to joke and deflect rather than reveal it's an AI, and it will do so, but I'm sure someone familiar with the AI could notice it doesn't deflect the way a human would.

Broke = passed

Reply on news.ycombinator.com