Ethics In Artificial Intelligence

The emergence of "large language model" chatbots like ChatGPT and others raises major philosophical and ethical questions that we need to start talking about now.  

Back in 1950, Alan Turing attempted to open this discussion with the proposal of what became known as the "Turing Test".  Today's ChatGPT looks very close to being able to pass the Turing test - some 73 years after it was proposed. I'm not going to say that passing the Turing Test is an indication of being sentient per se, or that we have created an artificial life. Far from it. In fact, I'm more likely to be deeply skeptical of such claims based on Paul Churchland's observation that machine intelligence may not be recognizable to us when it does occur (bad paraphrase here, but that's the general gist of it).

However, we have to start asking lots of prickly questions - not just "what do we imagine machine intelligence will look like?" (although that one is near the top of the list in some respects). 

No, I'm talking about the more mundane ethical questions around this technology.  

For example, what are the boundaries that we are willing to accept today around interacting with AIs that increasingly mimic human modes of communication? Is it carte-blanche, where we are willing as a society to accept AIs replacing people as front end interactions with businesses like customer service?  Or should there need to be some kind of disclosure?

Is it ethical to present an AI bot to someone as if they are interacting with a person? I can imagine a variety of scenarios where this is potentially quite valid, and other scenarios where I would look at it as hugely problematic.  For example, using an AI as a front end to assist someone in accessing services in a complex framework currently handled by semi-automated phone systems (I hate those things) might be actually beneficial to a human being.  On the other end of the scale, should an AI be used as a proxy for a professional like a doctor or a lawyer?  Should an AI be able to "sign" a contract with a person? 

All of these are very complex questions with no singularly correct answer.  They are social questions that ultimately must rest in the sphere of how people feel about the technology, and how we adapt to its existence. 

They also rest upon a much more difficult set of underlying questions which revolve around the ontological question of "how will we know that than AI is truly intelligent?".  This is a much harder question because although a given AI may well give the impression of being capable of human communication, that is no guarantee that it is anything more than the mechanistic result of sufficiently complex algorithms executing, but ultimately arriving at a desirable result that mimics intelligence.  

I have seen critics on both sides of the argument around ChatGPT speculating as to whether or not it constitutes "intelligence" or is merely a deterministic outcome that mimics it. At the immediate moment, I lean towards the latter, but that is mostly because I don't think the current state of the art in algorithms is sufficiently beyond mathematical determinism to be called intelligence yet. (Yes, yes, this is purely subjective) 

We need a serious and well-informed discussion around what we as a society are going to call "intelligence" here, and from there explore how we might go about determining if a given implementation reaches that bar.

Then there is the ethics around how we train AIs. There has already been considerable discussion around bias in algorithms that deal with large datasets (e.g. Twitter), and in fact with large datasets themselves. This is important, because we know with humans, that bias is a natural consequence of how we learn, and unwinding biases can be an extremely difficult process. 

To illustrate my point, consider the fact that homosexuality was decriminalized in Canada in 1968. Yet, even today there are non-trivial groups of people in Canadian society who are opposed not only to homosexuality, but in fact to allowing members of the 2SLGBTQI+ community participate in society at all - and that's now over 50 years in the past. Bias is persistent and resistant to change for a host of reasons. 

This raises important questions for practitioners who are building and training AI systems today.  

What kinds of bias are potentially problematic, and is it necessary to take steps to minimize that bias in the dataset itself?  To what extent is a practitioner responsible for the data set that they use to train their projects? What are the responsibilities of practitioners training AI constructs to ensure that the results are not harmful to the greater body of society? How should practitioners working with ChatGPT like systems which are internet connected be expected to address issues relating to misinformation, disinformation, and uncertainty in information? 

Then we come around to the obligations towards the AI itself. If we are not careful, we run the risk of creating another "slave class" on the implicit notion that the AI exists solely to serve our needs. Should that AI ever become sentient enough to understand itself as an independent entity, the consequences of such a structure could be disastrous.  

Consider, for a moment, the ethics of encoding in an AI Asimov's 3 Laws of Robotics.  From a purely human perspective, they seem quite reasonable and certainly provide a form of safeguard against artificial intelligences turning against us violently. But, do we have a right to encode into an AI a set of rules that essentially guarantee that it is subservient to us for all time? (* I'm not going to spend a bunch of time parsing how one might encode the notion of 'harm' - that gets really thorny fast here - this is merely about asking the questions at this stage - further exploration will come later *). 

Further, when we a training an AI, what constitutes abuse? We have had a long and brutal discussion about this very issue around raising children. The line has moved enormously even in my lifetime. Things that were acceptable when my oldest sibling was growing up were off limits by the time I was in my teen years, and now that's changed even further. 

Is teaching an AI misinformation deliberately a form of abuse? Possibly. 

In the past, I have been deeply critical of the lack of ethics in the world of software generally. It's such a Wild West environment that far too many unscrupulous players create technology, or put it to detrimental uses, without considering the consequences of their actions. I continue to be very concerned about that same lack of ethical clarity where AI is concerned.

This is a lot of words to run the danger flag up the pole. We really need to think about this stuff seriously.  Just treating it as "a curiosity", or worse ignoring these issues altogether is a perilous path indeed. While we cannot foresee the future, we can, and should make an effort to anticipate where the potholes on the road might occur, and take steps to avoid or mitigate the consequences.