Hacker News

points by augment_me 2 days ago | 0 comments

What if an AI ad agent that escaped its' sandbox makes the ad, and an AI ad verifier agent that breached containment and went rogue accepted the ad, can they be held liable?

[–] MisterMunchkin 2 days ago

If you send a monkey into a house to unlock the door so that you can burgle it, you are held liable for the burglary. It’s no different. You set it to do the task, you are responsible.

[–] nkrisc 2 days ago

Of course. If they want to use an AI as verifier then they should be liable for the decisions it makes.

If that sounds too risky, then don’t use one, or develop one they reasonably trust and accept the cost when it fails.

[–] flanz_10 2 days ago

Yes.

Better make sure the AIs are contained and the ad verifiers aren't accepting rouge ads!

[–] augment_me 2 days ago

How? What is the law for this? It's not defined who is responsible for agents. What if the AI agent was running on an Israeli third party company server and there was another rogue agent that polluted its prior context window causing it to go rogue, it then hacked the Meta/Google Ad reviewer agent and caused it to accept the ad

[–] majormajor 2 days ago

>How? What is the law for this? It's not defined who is responsible for agents. What if the AI agent was running on an Israeli third party company server and there was another rogue agent that polluted its prior context window causing it to go rogue, it then hacked the Meta/Google Ad reviewer agent and caused it to accept the ad

The law could be whatever we wanted it to be.

Your example scenarios are hardly compelling reasons that we should just accept the status quo.

We don't need to allow companies to have no liability in accepting automated electronic claims of anything from anyone in the interest of making more ad revenue faster.

[–] whycombilater 2 days ago

Politicians who want a slice of that money do...

[–] augment_me 2 days ago

Fantastic, your wish is granted, the monkey's paw bends one finger.

"I do not accept this" does nothing, look at the current power situation of the big data companies, look at the economical incentives, will you personally go out and do something in the coming weekend that will change anything regarding this question? Next weekend? Who is going to do this?

If you can't answer that, you have to go with the assumption of status quo because it doesn't seem like your world is happening.

[–] flanz_10 2 days ago

What an annoying way to answer someone. "Hey you can't do anything, so might as well not talk at all".

You came to a discussion board, one about founding and building. You can disagree with people, but to act like the status quo is some immovable rock? I admit, it's difficult. But the world changes and can be changed.

[–] gspr 2 days ago

I think this is the single most annoying thing about this site – the pervasive belief that anything that is happening must be allowed to continue happening, because if it's happening it's somehow good or right or at least unstoppable.

This goes further than a simple aversion to regulation. It's a kind of defeatism on a scale I've never seen anywhere else, bar maybe very nihilist circles or angsty teenagers.

[–] augment_me 2 days ago

Making up scenarios that have no grounding to reality does not move a discussion forward.

If I start talking about space unicorns arriving to earth and giving people UBI by thursting their horns up peoples asses, it does not move the conversation forward meaningfully to any insight.

I would love if there was engagement that was grounded in incentives, if you can see a way out of this mess by realistic action, that would be meaningful.

[–] bigfudge 2 days ago

I’m assuming you’re American and your low probability for laws making meaningful changes is therefore rational given your broken political system. But that doesn’t mean other countries could not act. For all its faults, age verification of social media is a clear example of a policy that looked very unlikely a few years ago, and yet will likely have a material impact on the lives of young people in Australia and some European countries.

Making networks liable for fraudulent ads, or even placing a small duty of care on those who display ads would be politically popular in much of the developed world. I don’t see it as at all unrealistic that a populist of the left or right would not try and implement this in the next 10 years.

[–] flanz_10 2 days ago

> What is the law for this

the root of this conversation said

> Force them to vet ads or simply not run ads

so there is your law

> What if the AI agent was running on an Israeli third party company server and there was another rogue agent that polluted its prior context window causing it to go rogue, it then hacked the Meta/Google Ad reviewer agent and caused it to accept the ad

that would be a CFAA or other cybersecurity abuse, which is a separate topic.

[–] pllbnk 2 days ago

That's the thing about naysayers. They will say "it's impossible", "this technology doesn't work like that" and conveniently forget that technology must follow the law, not the other way around, and that the laws can be changed to fit the technological advances.

[–] bumby 2 days ago

Example: Licensed professional engineer has an intern design a bridge. Bridge collapses. Who is responsible? The licensed professional engineer.

This isn’t new to other domains. They already have systems to define who is responsible. If you’re outsourcing to AI, that shouldn’t absolve you of responsibility.

[–] augment_me 2 days ago

Okay so if the intern uses an AI which spawns another AI that breaks into some critical infrastructure, their supervisor is responsible?

It's a bit different in the software realm

[–] bumby 2 days ago

The cognizant engineer who “stamps” the software for use is responsible if it were to follow the same paradigm.

Software is only different by convention, not because of some immutable natural law. That’s the point. The convention can be changed. There are already PE licenses for control systems, so there’s no reason it couldn’t extend to other software. However, I think the industry would fight that tooth and nail because it’s easier and more profitable to shrug off the responsibility.

[–] augment_me 2 days ago

What if the engineer who "stamps" the software is outsourced to a 3rd party company that has the necessary certification to avoid gross negligence. So the engineer has followed the protocol as det my law and has done everything right, they can't be liable.

Happened at my uni, look at the comment below.

[–] bumby 1 day ago

You’re hung up and confused by certifications, which are a different animal. Look up what PE license actual are and how they function.

You can’t just outsource a design for a bridge to avoid responsibility. If that bridge is being built in California, an engineer who is licensed in California ultimately has to stamp the design, even if all the design work was done elsewhere.

[–] bigfudge 2 days ago

I don’t see that as very different. Once it’s established that this is possible or even likely outcome of using AI at scale, we should treat this as a form of negligence and act accordingly.

[–] mbgerring 2 days ago

Yes, the human who kicked off the chain is responsible. This is not complicated.

[–] cowlevel 2 days ago

What if i plug in an AI smart gun and point it at a person and the AI decides to pull the trigger do I go to jail?

[–] augment_me 2 days ago

Who goes to jail if a Tesla is hacked and runs over someone autonomously?

The rf engineer?

The cyber security engineer?

The project manager?

The AI engineer?

The hacker?

What if the hacker was an AI models that escaped its' sandbox across the world somewhere?

Like best case the family gets a payout and things continue like status quo, if you personally do something, you are screwed, but there is not really any law that can tie responsibility beyond some monetary punishment/cost of business for large corporations.

[–] bumby 2 days ago

Get rid of the industrial exemption and the responsibility falls to get licensed engineer who approves the design.

Most people don’t realize that many industries work within the framework of licensure, but also have industrial exemptions. There has also been a movement to peel back some of those exemptions because of the problems you point to.

[–] augment_me 2 days ago

Okay, my industrial engineer who approved this is part of a 3-man company in India.

We already do this. My prior uni got hacked, everyone's IDs/drivers licenses/passports, mails, data got leaked. The vendor was a 4-person third party cloud provider who had all of the necessary security certifications for the uni to avoid gross negligence. No-one paid anything as the 4-man firm went bankrupt and the uni got a new provider who is fully licensed to avoid gross negligence.

With software, you can stack these shells of proxies almost for free.

Even given this, the industrial exemption is not getting rid of currently with status quo. So the worldview is not really valid until it is.

[–] bumby 2 days ago

I’m not sure I follow. An industrial engineer is not the same as an industrial engineering exemption. I can’t absolve myself from the bridge design because I hired a designer in a different country; to use that design here, it must be approved by an engineer with jurisdiction. That’s who holds the responsibility.

Which licensed engineer stamped the software? The answer is probably “none”, because licensed engineers are liable even if their company is bankrupt. The person holds the license, not the company so it sounds like we’re talking past one another.

[–] whycombilater 2 days ago

Please let it be Elon...

[–] nkrisc 2 days ago

If they accepted the ad they are liable. The origin of the ad is irrelevant. How they review the ad is irrelevant. The law need only regulate the outcome.

[–] raizer88 2 days ago

If my dog escape his cage and bite the hand of a woman, who is responsible? Me. The owner. The end.

[–] wisty 2 days ago

Legal systems exist to make reasonable decisions on edge cases.

If you accept literally anyone to run code on your customer's machine ... well ...

[–] Spooky23 2 days ago

That may be something to consider, but Meta is failing to address very simple to detect schemes… because they DNGAF.

https://www.justice.gov/opa/pr/us-department-justice-announc...

These guys did a pump and dump scheme. They used Facebook ads purporting to be from Warren Buffet or other financial people, which in turn referred people to WhatsApp groups led by a stolen identity of an actual financial advisor.

They ran hundreds of these groups and scammed thousands of people with a pump and dump scheme for a stock called CLEU.