Hacker News Reader

Comment by mrfinn | Hacker News Reader

Which risk? I mean I saw "Terminator 1" and "Terminator 2" (great movies) and I played "Horizon Zero Dawn" (great game). But what proof do we have of those risks exist apart from wild exercises of imagination?

View on HN

You cannot prove that the risks exist, and you cannot prove that the risks do not exist, so this approach to the problem is clearly not useful to make any progress on the question "is AI dangerous or not".

Of course you can, there's a thing called research. This narrative is being pushed WITHOUT any significant research behind. So this is only about economics and politics, not science.

How would you design an experiment to definitely prove or disprove that AI is (not) dangerous?

Well I'm not being paid to do that, but my professional instinct is, like you do with any other software, you set up a sandbox. Pretty sure an AI specialist could have much better ideas than me. What I don't get is how this narrative is being so wildly pushed without any research sustaining it, even just to say "hey we tried to investigate this huge issue, here you can see the great amount of papers we wrote, but we concluded is actually impossibly to establish the risks because etc etc". This is first class BS.

I am not sure how you came to the conclusion that there is no research on this topic. On Google Scholar you can literally find millions of research papers discussing the risks of AI, its ethics, alignment with human values, and so on, for example:

- Scherer, Matthew U. "Regulating artificial intelligence systems: Risks, challenges, competencies, and strategies." Harv. JL & Tech. 29 (2015): 353. (934 scholar citations)

- Araujo, Theo, et al. "In AI we trust? Perceptions about automated decision-making by artificial intelligence." AI & society 35 (2020): 611-623. (334 scholar citations)

- Osoba, Osonde A., and William Welser IV. "An intelligence in our image: The risks of bias and errors in artificial intelligence." Rand Corporation, 2017. (278 scholar citations)

- Gabriel, Iason. "Artificial intelligence, values, and alignment." Minds and machines 30.3 (2020): 411-437. (216 scholar citations)

- Bryson, Joanna, and Alan Winfield. "Standardizing ethical design for artificial intelligence and autonomous systems." Computer 50.5 (2017): 116-119. (214 scholar citations)

- Yu, Han, et al. "Building ethics into artificial intelligence." arXiv preprint arXiv:1812.02953 (2018). (200 scholar citations)

- Munoko, Ivy, Helen L. Brown-Liburd, and Miklos Vasarhelyi. "The ethical implications of using artificial intelligence in auditing." Journal of Business Ethics 167 (2020): 209-234. (180 scholar citations)

- Christian, Brian. "The alignment problem: Machine learning and human values." WW Norton & Company, 2020. (178 scholar citations)

- Manheim, Karl, and Lyric Kaplan. "Artificial intelligence: Risks to privacy and democracy." Yale JL & Tech. 21 (2019): 106. (178 scholar citations)

- Russell, Stuart, et al. "Ethics of artificial intelligence." Nature 521.7553 (2015): 415-416. (164 scholar citations)

- Risse, Mathias. "Human rights and artificial intelligence: An urgently needed agenda." Hum. Rts. Q. 41 (2019): 1. (135 scholar citations)

- Turchin, Alexey, and David Denkenberger. "Classification of global catastrophic risks connected with artificial intelligence." Ai & Society 35.1 (2020): 147-163. (108 scholar citations)

- Taylor, Jessica, et al. "Alignment for advanced machine learning systems." Ethics of Artificial Intelligence (2016): 342-382. (95 scholar citations)

- Santoni de Sio, Filippo, and Giulio Mecacci. "Four responsibility gaps with artificial intelligence: Why they matter and how to address them." Philosophy & Technology 34 (2021): 1057-1084. (84 scholar citations)

Oh c'mon... you know what I'm talking about, I'm not talking about theory and more theory, I'm talking about actual research done on top of an LLM, giving it the apparent freedom to go rogue and see what tries to do then. Specifically on top of OpenAI's GPT, as they the ones carrying this message along the world. This is the type of research I'm talking about. I saw Altman's presentation to the US congress and not a word, about any study, was said.

Well, there are plenty of real-world examples of harms caused by AI-powered systems (https://incidentdatabase.ai). For LLMs in particular, I think it's just too early to have convincing independent research on their harms, but coincidentally this just happened a few days ago:

National Eating Disorder Association shuts down A.I. chatbot it planned to use to replaces humans saying it 'may have given' harmful information (https://fortune.com/well/2023/05/31/neda-ai-chatbot-harmful-...)

> You can't prove [claim] doesn't exist

I'm sure there is a name for this logical trick you are trying to use.

OK I will claim there is a grand spaghetti monster. You can't disprove the existence of the spaghetti monster. Therefore, the spaghetti monster must exist.

The burden of proof is on the AI fanfic crowd to prove that terminator and AGI is just around the corner.

> I'm sure there is a name for this logical trick you are trying to use.

Your example is an instance of the fallacy of the inverse (https://en.wikipedia.org/wiki/Denying_the_antecedent).

But that is not the type of reasoning I was doing. Following your example, you also cannot prove that spaghetti monsters do exist. Since neither side can offer definitive proof of their stance, the only way to settle the debate is to have less stringent requirements than proofs, such as, for example, circumstantial evidence.

Nobody has ever seen spaghetti monsters, therefore it seems unlikely that they exist. Not a definitive proof, but it does allow a certain degree of certainty. The point is to reach probabilistic conclusions in light of the available evidence, rather than seeking definitive proofs, which are impossible for most real-world questions.

> The burden of proof is on the AI fanfic crowd to prove that terminator and AGI is just around the corner.

No, this is an argument from ignorance (https://en.wikipedia.org/wiki/Argument_from_ignorance). Each side of the debate is supposed to provide proof, evidence, or reasons for their perspective, otherwise they do not have much clout, besides prejudice.

Since there are plenty of examples of risks and harmful consequences of AI applications (see my other replies in this thread), it is quite reasonable to assume that new AI technologies pose new risks and that we should at least consider what those risks may be. Now the burden is yours to argue that no, despite all these past failures, we're still totally safe.

Myself and many others have plenty of evidence from daily life that AGI does not yet exist. The burden of proof is not on me.

None of the existing AI technologies we examine have AGI. There are hundreds of examples you can use.

Reply on news.ycombinator.com