Mind Readings: AI Bill of Rights, Part 1: Safe and Effective Systems

In this five part series, we're going to take a look at the proposed AI Bill of Rights and see what's in the box. See what uh commentary we can offer about this document. Now, you probably want to get a copy of this to read along. I'll put a link in the in the notes below. Um, but this is an important document for anybody who's working with artificial intelligence, uh, machine learning, data science systems.

In particular in marketing, if you are using any kind of marketing automation software or CRM software that has machine learning and artificial intelligence built in from automated lead scoring to anything else, you should know what's in this document. Uh, this was released by the White House uh just a couple months ago, and while it as a whole does not carry the force of law yet, there are many provisions with it that are already covered by existing laws. So that's one of the reasons why we want to uh take a look at this document. Uh also, full disclosure and disclaimer. I'm not a lawyer.

If you want a lawyer on AI, go talk to Ruth Carter. Go to GeekLawfirm.com for that. I'm not a lawyer, this is not legal advice, but we are going to talk about some of the implications that could happen uh if your business is not doing things uh as ethically uh or as uh aligned with the law as you should be. So there could be legal penalties for not using artificial intelligence in the right way. Five part series, because there are five fundamental AI rights that are in this document.

And I think it's important to point out this is not a document about machines becoming sentient and having their own rights. This is about the rights of individuals, human beings, uh, when subject to the outputs of AI systems. So that important clarification. So let's dig in. Right number one safe and effective systems.

You should be protected from unsafe or ineffective systems. Automated systems should be developed with consultation from diverse communities, stakeholders, and domain experts to identify concerns, risks, and potential impacts of the system. Now, one of the things that I really like about this document is that it's not just theoretical. In each of the sections that we're going to talk through, there are examples of the kinds of things that each right is supposed to mitigate or prevent. So in this one for safe and effective systems, the first counterexample is a proprietary model was developed to predict the likelihood of sepsis in hospitalized patients and was implemented at hundreds of hospitals around the country.

An independent study showed that the model predictions underperformed relative to the designers' claims while also causing alert fatigue by falsely alerting likelihood of sepsis. Example two, on social media, black people who quote and criticize racist messages have had their own speech silenced when a platform's automated moderation system failed to distinguish counter speech or other critique in journalism from the original hateful messages to which such speech responded. I'm pretty sure that one's Facebook. That would be Apple's AirTags. That was a relatively recent thing.

Number four, an algorithm used to deploy police was found to repeatedly send police to neighborhoods they regularly visit, even if those neighborhoods were not the ones with the highest crime rates. These incorrect crime predictions were the result of a feedback loop generated from the reuse of data from previous arrests and algorithm predictions. So this first one, safe and effective systems, is a good idea. I mean, all these are good ideas. Some of them are going to be harder to implement than others.

Safe and effective systems is a pretty as slam dunks go for AI. This one seems to be a pretty safe one. Your system should do what it says it does, right? And the outcomes. One of the reasons why this right is necessary to even be discussed is because nobody who's building AI systems in these examples, in these examples, um, is asking the very simple, straightforward question: what could go wrong?

Right? You make a tracking device and don't anticipate that someone could misuse it, right? How could somebody turn the system against its intended use? How could somebody use it off label? What are the things that you could use that you wouldn't want your system to be used for?

Say maybe you make a social network and it's being used to undermine democracies around the world. That seems like an unintended use. Now, here's the challenge. And this is something that the guide only peripherally discusses. It talks a lot about things like consultation, independent audits, uh, evaluation, regular reporting, and things like that.

But there needs to be legislation in place to create penalties for violating these rights. Because right now there isn't. Right now, there's there's no penalty for Facebook undermining democracy. There's no penalty for Apple making air tags that can be used by stalkers. There's no punishment for bad actors.

And bad actors, most of the time, when it comes to AI technology, are the technology manufacturers. There are obviously individuals who misuse the technology, right? They they intend to use it in ways that are not authorized, but there also needs to be some level of liability for the manufacturer of the technology, or this is all just nice thoughts, right? This is the AI equivalent of thoughts and prayers. It doesn't do anything.

And somebody highlights, hey, you could use this to stalk somebody, and then you don't take steps to uh mitigate that, you should absolutely be held liable for it. You should absolutely be held liable for creating something that someone highlighted this is could be a potential problem and a realistically potential problem, right? Putting a tracking tag in someone else's luggage, that's not like sci-fi, right? That's not uh some crazy James Bond uh uh thriller, which is totally unrealistic. No, that's very realistic.

That's very easy. Taking a uh a tracking tag and taping it to somebody's car bumper. That's not only realistic, that's been in like every spy movie since the 1960s. So when we're talking about artificial intelligence systems, we're talking about how are the ways that it could be misused now. And the big question that we have to ask with all these systems is how can they create unsafe outcomes, right?

What are the outcomes that would be inappropriate? What are the outcomes that the ways you can misuse these systems? For example, uh deep fakes has been a differential use of technology for a long time. I mean, the initial purpose was adult entertainment, but it has since been used to simulate uh you know world leaders saying things, literally putting words in their mouths they didn't say, um, as a propaganda tool. The systems and their manufacturers that enable that to some degree have to be have a part of the accountability in it in order for these these regulations to have teeth.

Now, the catches for some of this is going to be open source systems. Open source systems by definition do not have a level of accountability, right? You release some of your software to the public, you say, here it is public, do with it what you will. Um, we expressly disclaim any liability, and we provide no support or help, right? It's it's just here's the thing.

With AI, that's becoming a challenge, right? Uh services and systems like Hugging Faces Transformers, the T5 transformers, uh the GPT Neo X models, uh, stable diffusion, these are all open source products. They're given away freely. Anyone can take them and reuse them, and like any other tool, some people are going to misuse them. So there is no provision right now in any of this document for dealing with the question of open source.

Because what you don't want to do is you don't want to stifle open source development either. It is responsible for a substantial amount of the progress in the space, academia and open source. That's been the case for decades. So there has to be that part as well. But overall, the safe and effective system seems pretty logical.

And our takeaway as people who operate these systems is what could go wrong? How could we be misusing these systems or using them from uh in ways that are unanticipated, or what harm could we do with a system and not realize it? For example, I met with um an insurance company a few years at a conference uh and they were talking about how proud they were that they've developed a system that would only market to specific segments of the audience because they wanted to get only the best customers, and unfortunately they reinvented redlining and the process, which is the act of discriminating against certain um demographics within a city. Again, no one was there to ask, hey, how could this thing be misused? And clearly in this case it was.

So that's uh part one. Stay tuned. Uh next up will be part two on algorithmic discrimination. If you like this video, go ahead and hit that subscribe button.