A thread for unboxing AI

There's a big difference between the two. We know that out of say a million people getting driver's licenses a percentage will run a red light and kill someone. But there's a big difference between that type of risk and actually knowing ahead of time that the ai will error a percentage of the time and kill people for certain. I don't see a point where we allow passenger planes to be flown by an ai that can hallucinate altitudes. And I think that sort of standard will occur wherever people using it can incur some sort of loss if ai gets it wrong.

This just seems one of the most surreal turns in the debates. This is now. it's not science fiction anymore.

AI is being used now and it is being tested for much more use. The UK is testing AI cars and talking about introducing them in the next few years. They will most likely miss that but I just can't imagine what criteria some here think is being used for allowing use. AI robots ae starting to appear, early days but it's way too late to argue about whether it's going to happen. The safety issues are overwhelmingly about outcomes not interpretability*. Yes AI accidents are different and so there will be demands for evidence they are are undeniably much safer than humans. How much this will be ignored or fudged remains to be seen.

Yes, Ethics, law and politics are going to be massive issues. That's not in dispute.

*One form of interpretability might become common. That's where we ask the AI to tell us how it comes to it's conclusions.

It's not about goals. it's about reality. Part of the denial of what is happening it to compare AI to some perfect standard. It's ridiculous because the reality of automation is that it's about cheaper,faster and/or safer

A standard example of this is in an article about entry level programming jobs.

https://www.bbc.co.uk/news/articles/cm21dvg8l1go

BBC News

Computer science graduates struggle to secure their first jobs

well ldo.

Most of the science fiction I hear are the horror stories about ai running amok that probably won't ever happen. The main issue with an error is if someone gets killed or injured. But it's only a small minority of businesses and potential applications where that could happen. Mostly ai getting it wrong will just cost the company using it money.

Yes and how that's going to be handled is a key part of the political, ethical and legal debate. Not whether it's going to happen (or, otherwise interesting discussion about the difficulty of knowing the internal workings, interpretability etc). That ship has sailed.

If a company hires a bus driver with a known history of sudden narcolepsy and... that's basically the definition of reckless homicide. I don't think there's much legal debate over whether or not execs will become criminally culpable if they employ AI and it falls a sleep at the wheel.

Suppose the AI system reduces such incidents by several orders of magnitude and theres no oncealing of safety data erc.

You might still be right but I wouldn't bet on it. Not every failure is criminal.

The horror stories of AI centers on general intelligence models, and we don't have those yet.

We have generative intelligence and deep learning models, which are models which makes predictive output token by token based on previously being fed training material. They "speak" by guessing what the next word will be. It can also do things based on predictions based on training material it has been shown. Well, except for the big issue where they hallucinate the wrong thing, which happens... and is what the entire previous debate I was involved in was about, so I won't repeat it.

In comparison, a general intelligence model will speak based on what it wants to say, and it will perform actions based on what it wants to achieve, and perhaps more importantly it will be able to remember very well what it did 5 seconds ago. This is the AI where the experts divide about 50 / 50 about whether it will lead us into a transcendent future, or whether it will doom us all to extinction. Some think that if we build these machines, it will be "Skynet", others think it will be "Multivac".

I maintain as I did earlier in the thread that it is a problem that we base discussion about general intelligence models based on our experiences with generative models. It is like basing your regulations of fusion power based on your experiences with making lemon batteries; there is an overlap, but it is pretty damn irrelevant.

If we want to worry about new horrors (and it's a very real worry) then it's drones/robots/ai/etc used by human governments.

The development is so fast. I don't think we need skynet.

I think there is a lot of debate. I strongly suspect that there will be statutes passed that limit owners and managers of driverless car companies from liability so long as the companies follow XYZ regulatory protocol and have been blessed to operate by the relevant governmental authority.

For general intelligence models I think that makes sense. There will be a time period where we debate whether to treat them as machines or sentient beings, which will in some (and I mean some, not all of it) part mimic the debate on slavery: Some will want to view them as machines so they can make more money, some will want to view them as people (or near as) lest we be monsters.

The problem we will face in treating them as machines is that machines are rarely seen as culpable. When faced with a machine that can do its own thinking, this view can fast become problematic. As a lawyer, you might find that to be an intriguing thought experiment and you will have a better notion than me on how that could pan out.

Then we will have to start figuring out if we want an entire set of laws for sentient machines, a law for sentient beings (if we want to include more than machines) or if we should start including such beings into the laws that governs us. Which will probably end up a huge societal mess regardless of what we do.

I will check out the book.

I will be curious to hear what you think if you read it. To be clear, the book isn't really about AI (although AI is a part of the book) or whether our machines should have legal rights if they become advanced enough. It is a first contact story that explores what it means to be alive, to be sentient, and to be conscious. As I mentioned before, the author does the reader no favors. The book is revered in hard sci-fi circles, but it is not easy.

I am keenly interested in this topic. Because of our priors about human exceptionalism, I think that we are a long way away from broad public acceptance that AI models (whether general intelligence models or something lesser) should have rights that are co-extensive with human rights. On the flip side, I easily can envision a near future in which activists promote the idea and even engage in various forms of what they view as white hat destructive behavior to draw attention to their beliefs. And I can envision a relatively near-term future in which humans are legally prohibited from turning off certain AI models, although I think the reasoning for such laws may in the first instance be grounded in practical concerns rather than moral concerns because the former will allow for easier consensus building.

https://www.vancouverisawesome.com/highl...

Autopiloted car drives through children's crosswalk while lights are flashing and children waiting to cross.

A failure rate is a 'design flaw' of many things that are used. It doesn't always imply negligence.

In these discussion think its also important to realise that any country that prevent AI/robotics etc by legislation even if it's massively safer,cheaper/faster will be left behind. Then they would have to change the legislation because it makes no sense at all except 'feelings'. I seriously doubt countries will go down that route for very long, especially given the power of the tech companies to push it.

The far bigger real problem imo (if there's time) will be the use of AI/robotics etc that isn't safe enough just because it's so much more profitable. That's more capitalism' mo.

In the normal case of negligent homicide that is undertaken by a person negligently taking action that leads to the death of the other person, where is the agency there?

If I’m stopped at a red light and someone on their phone plows into my trunk sending me hurtling through the intersection and getting tboned, is the fact that they are being driven by a person any real comfort?

In any case it just seems you’re trivially wrong about not being able to shield companies from civil or criminal proceedings. As a matter of definition that’s practically all that a statute is in this context, establishing some kind of criminal or civil penalty for an act.

References to negligence just confuses the situation in my opinion, because negligence is itself implying some kind of violation of a criminal law. It’s sort of like invoking murder. The concept of it being illegal is baked in. It’s better to think of this situation in terms of risks and benefits, and whether the risks can be mitigated such that the benefits are overwhelming. For instance, Waymo doesn’t get on the freeway.

I’m not following how identifying certain risks that can lead to death is uniquely bad to AI. Is it because there’s some kind of computer involved? Because risks that can lead to death are just a pretty common thing throughout many professions, products, industries, etc.

that's literally what gvmnt does regulating airplane insurance, or vaccination lawsuits, and many other things (nuclear energy and so on)

btw pls stop thinking self driving cars AI is based on LLMs (it's not)

What you wrote above isn't correct. It varies by jurisdiction, but in some jurisdictions, you absolutely can sign an enforceable contract in which you waive your right to assert claims based on negligence. (Contracts that purport to eliminate liability for gross negligence or recklessness, on the other hand, are never enforceable.)

In any case, I wasn't talking about signing away your rights via some sort of contract of adhesion. I was talking about a statutory exemption from liability. I can't think of any reason why such a statute would be unenforceable.

If an auto manufacturer decided to use a wheel bearing with a known and predictable history of catastrophic failure and someone gets killed as a result, how is that not reckless homicide?

I don't know what statutes you guys think we could put in place that would override what is essentially an inalienable right to not be recklessly killed. I don't think negligence is a high enough bar either since it's hard to say that they should have known better but didn't when everyone now knows better.

I legit have no idea what you are talking about. The only reason murderers are subject to criminal liability is because there are statutes that prohibit murder. You might think you have a natural right to not be murdered, but that's something else entirely.

You're effectively arguing for strict liability to be applied to self driving. The confusion is "negligence" and "reckless" have very specific meanings in law and they don't apply to strict liability cases.

The whole point of the discussion is the application of strict liability holds the AI to an impossible standard: perfection.

As some of noted, the probable outcome is some legislation will be necessary to shield AI companies from some of the strict liability (capped fines, if any, if certain standards are met etc.)

We can't write a statute that legalizes slavery again.

Those homicide statutes don’t come from nowhere; they flow from our unalienable rights in the Constitution. The law codifies what’s already a baseline: people have a right not to be recklessly killed. You can’t draft a statute that wipes that out and makes it legal. That’s why I’m saying negligence isn’t even the right bar here because if execs knowingly deploy tech with foreseeable lethal failures, that’s reckless homicide.

It isn't obvious to me that it would be unconstitutional for a state to abolish its murder statutes.

Sure, they are required to get their vehicle up to certain standards, but they won’t be expected to get their wheel bearing to the point where it will never fail ever.

That’s why they include servicing recommendations in your owner’s manual.

Well I think the concern is and the reason why it doesn’t work the way that you’re asking for it to work is that it seems like nothing could ever work if we had the standards that you’re talking about. Effectively you are placing a much higher standard on driverless cars than you are on cars driven by humans, even if the driverless cars would actually be faster, more efficient, and lead to less deaths.