Guide to Navigating AI and Its Ethical Challenges
Lesson 2 of 3
Hybrid Technology: Artificial Intelligence and Human Input
Free. Watch without an account. Sign in to save progress and take the exam.
Up next
Ethical Concerns of AI UseTranscriptThe player shows YouTube’s automatic captions, which are machine captions we do not review. The full transcript is below.
00:26:54And so with that in mind, I'm going to just run us through sort of, you know, so I think we covered
00:27:00a lot of this information here. So I'm just going to run through it. And if you guys want to weigh
00:27:03in with any particular point, but I mean, logical rules-based approach kind of already discussed
00:27:08that. I think Moira alluded to it earlier, you know, dog, cat, mouse, how do you know the
00:27:14difference? And then sort of human experts ontologies and how they all kind of mesh
00:27:19together with the AI we've kind of covered. And then, you know, these expert systems,
00:27:25I think you're going to see machine learning will be deployed in a more nuanced or focused
00:27:30fashion. We're already seeing it in pathology and tax and whatever. It's just, again, a function of,
00:27:37you know, application of creating specific models. And then, yeah, I think the cars and the vehicles,
00:27:44I think that, you know, most successful AI systems today, I think everybody sort of alluded to work
00:27:51with human intelligence. I always question whether to put the word intelligence or input. But
00:27:56the point being is that with some sort of degree of human, you know, oversight or engagement,
00:28:02we don't really have machines learning and teaching machine model. I mean, in the labs,
00:28:06they're working on these sorts of things, but not in any sort of stream of commerce widely as
00:28:11today. You will probably see that most likely in vehicles, but where they talk to each other and
00:28:18then share data and so on and so forth. But that's sort of ran through those slides. And we'll
00:28:23jump into the ethical issues of it. But I guess maybe if there's any particular point that
00:28:31the panelists want to touch on, maybe Karen, Mayra, Jenny, on the slides we just covered,
00:28:36we're happy to touch on them. We'll jump into sort of the ethical issues that concern regarding the
00:28:41actual use of these technologies that lawyers need to be thinking about and discussing with
00:28:46their clients. Karen? Yeah, so the only thing I would want to circle back on is that as we've
00:28:52now layer these different kinds of AI and kind of applications for AI, whether it's the original
00:28:58kind of more narrow AI or these generative AI tools, they're all, just I want to remind you,
00:29:03they're all still doing the same thing. They're all predicting outputs based on vast amounts of
00:29:07data and specifically selected logic, right? And so they have a lot of, they share a lot
00:29:15of characteristics, both strengths and weaknesses. And as we talk about ethical concerns,
00:29:21they may manifest in different ways and the legal problems may manifest in different ways,
00:29:25but they're still basically have the problem that Jenny alluded to, which is they don't have
00:29:30or care about ground truth, right? Like, so with carrots, a health food or not a health food or
00:29:35junk food or not a junk food, it doesn't really know what a carrot is. It just has a lot of
00:29:40references to carrots, right? And so that's the one thing I would say. The other thing is,
00:29:46as we're talking about ethical concerns and sort of systems into which we put these
00:29:50AI tools, humans are very important, but there's a distinction that gets conflated a lot,
00:29:58particularly when you start looking at some of the proposed regulations and some of the
00:30:02self-governance rules. The distinction is that I, right there too, that I focus on. One is,
00:30:09where are we talking about responsible use and sort of the human category of questions and where
00:30:14are we talking about trustworthy technology and what are we, you know, the characteristics of
00:30:18the machine itself. And that interaction is really important, but the responsible
00:30:23human use of trustworthy technology needs to be sort of kept in mind because both the problems
00:30:31and the solutions are, I think, clarified when we know whether we're talking about which part of
00:30:39that equation. And then the second thing that gets conflated a lot is automated decision-making
00:30:45versus sort of decision support tools. And Daniel, as you were alluding to, I think most of the AI
00:30:51that we're using today that we're aware of or that we're conscious of or that like the generative AI
00:30:57stuff, those are really decision support tools, right? They're helping humans make better choices
00:31:02or do better in a job that is still a human job. Automated decision-making, there is some
00:31:07of that, but that's sort of a different category and raises different kinds of ethical concerns.
00:31:12And so I think it's really important when we're talking about some of these issues and
00:31:18opportunities that we know whether we're talking about a truly automated decision-making system or
00:31:23whether we're talking about a decision support system. So responsible use of trustworthy
00:31:28technology, is it a decision-making system or is it a decision support system? I think that'll
00:31:33help us go forward in this conversation with a little more clarity. I think to add to that,
00:31:40when you're talking either to a vendor or to a client, to get back to the definitional issues
00:31:50when we're talking about, everybody calls all of this stuff AI. And it's important to find out,
00:31:59are we talking about an expert system that has a bunch of rules in it? Are we talking about deep
00:32:04learning? Because they work very differently. The risks are very different. An expert system
00:32:10is going to be a lot more explainable than a deep learning system. And use cases is critical,
00:32:18as Karen said. Not all of these things are the best answer for every single use. Some of them
00:32:27are better for certain use cases. So you really want to be clear with the person,
00:32:33if it's a vendor, what it is they're offering, what kind of AI it is, how it works. And you have to
00:32:42be clear about what kind of problem you're trying to solve and whether that technology is the right
00:32:49technology to solve that problem. You can have the greatest technology on the planet, but if it's
00:32:54not the right technology to solve the problem you're trying to solve, it's not going to be a
00:32:59good hope. Yeah, I think the only thing that I would add here is that when we're talking about
00:33:10human interaction with AI and people have been, I think, rightly concerned about the potential
00:33:21job loss from the use of AI systems. But one thing that we're actually seeing is that
00:33:27something coming out of the current crop of AI models is the concept of a prompt engineer.
00:33:35The idea being that the better your questioning of the AI model, the better your response will be.
00:33:40So there are actually organizations out there that are hiring people to work with AI models,
00:33:45with the idea that the output from those models can be curated. So in a way, it's almost taking
00:33:51that automated model and fine-tuning it through the use of humans. This all presents interesting
Automatically generated; it may contain errors.
Lesson notes
The role humans play in building AI systems.
Study materials
Notes, flashcards and practice questions for this lesson, reviewed before publication.
Study notes
From rules-based expert systems to hybrid AI
AI has developed along several lines: logical, rules-based approaches; expert systems built on human experts' knowledge and ontologies; and machine learning that learns from examples, such as telling a dog from a cat from a mouse. Machine learning is increasingly deployed in focused, specific models, for example in pathology and tax.
The most successful AI systems today work together with human intelligence or human input. Some degree of human oversight or engagement remains the norm. Machines teaching other machines exists mainly in research labs rather than widely in commerce, though vehicles that talk to each other and share data are a likely early example.
Whatever the type, from narrow AI to generative tools, all of these systems do the same basic thing: they predict outputs based on vast amounts of data and specifically selected logic. They therefore share strengths and weaknesses. They do not have or care about ground truth. A model does not know what a carrot is; it only has a great many references to carrots.
Responsible use versus trustworthy technology
Two distinctions often get conflated, especially in proposed regulations and self-governance rules. Keeping them separate clarifies both problems and solutions.
- Responsible use versus trustworthy technology. Responsible use concerns the human questions: how people choose to deploy and rely on a tool. Trustworthy technology concerns the characteristics of the machine itself. The goal is responsible human use of trustworthy technology, and it helps to know which side of that equation a given issue falls on.
- Automated decision-making versus decision support. Most AI in use today, including generative AI, consists of decision support tools that help humans make better choices in jobs that remain human jobs. Truly automated decision-making systems exist but form a different category that raises different ethical concerns.
Questions to ask a vendor or client about an AI system
Because everything gets called AI, counsel should find out what kind of system is actually involved:
- Is it an expert system built on rules, or a deep learning system? They work very differently and carry different risks. An expert system is far more explainable than a deep learning system.
- What is the use case? Not every technology is the best answer for every use. Some are better suited to certain tasks.
- What problem are you trying to solve, and is this the right technology for it? The greatest technology on the planet will not help if it does not fit the problem.
When dealing with a vendor, be clear about what it is offering, what kind of AI it is and how it works.
Prompt engineering and the human role in AI output
Concern about job losses from AI is understandable, but the current generation of models has also created a new role: the prompt engineer. The better the questions put to a model, the better its response. Organizations are hiring people to work with AI models so that the output can be curated, in effect fine-tuning an automated model through human skill. This reinforces the broader point that today's AI works best as a hybrid of machine output and human input.
Flashcards
10 cards. Test your recall, then flip to check.
Select the card, or press Space, to flip it. Left and right arrow keys move between cards.
1 of 10
Practice quiz
Practice — not graded, not the final exam