So, AI Thinks It Can Debate
A debater’s reaction to Project Debater’s debating debut
By MARTIN REZNY
The future is truly here — Project Debater is another IBM’s attempt to defeat puny humans at one of their talking games, after Watson dominated Jeopardy. Sounds to me like the in-between step should have been a project naming AI, but I digress. If you haven’t seen the first historic debate between a human and an AI, I recommend watching it first:
Making Better Decisions
Let’s mix things up and start with the end result of the debate, as that shines the most light on the intention behind the project and its potential future uses and ramifications. The AI lost the debate, mainly because it was better at formulating arguments and backing them up with facts than it was at responding to specific points of the opponent. With that said, it was on the level of a well-trained beginner debater, somewhere around average.
For me, no surprises there — a Watson-based AI was guaranteed to be able to work with facts, at a level that only a human savant or internet-assisted human with a lot of time for research can maybe match. In this regard, I agree with Harish, the human debater, that this type of system can be of great help to people who quickly need the facts laid out to them in a way that’s sufficiently complete and reflects the context of the issue at hand.
At the same time, after browsing through the project’s scientific papers, I was sure that the AI will not be able to master debating strategy, or be able to identify nuanced points of the opponent and respond to them specifically. This is because the theory behind the AI as it was designed focuses only on content, which is one third of what determines outcomes of debates — the weight and clarity of arguments and evidence behind them.
Beyond that, you need to consider strategy, of which the AI only mastered the structure element — order and timing of content — and style, which includes reading the body language and biases of opponents, audience, and judges. I can totally see style being possible to master, given that AIs already begin to be able to read facial expressions and tonal inflections. Strategy in persuasion, however, is a whole another game all on its own. I’ll focus on that in detail later, as it very much isn’t a simple problem to solve.
What was a bit more worrying was how the debate was judged. What normally happens in a debate where the audience is voting is that you ask everyone about who won the debate. That’s not how it happened in this case. Instead, the audience members voted twice, before and after. What was measured was the difference between the two. That’s an attempt to measure persuasion, not the outcome of a debate. Harish won 100%.
On the surface, this is a clear attempt to make the result look better even if the AI loses decisively (as it is called in debating, and as it happened here). This way, the result was 17% shift to Harish’s side. First of all, this is a misunderstanding of what debating is — debating is not persuasion. You can win the debate clearly by having the best arguments and be the least persuasive person on Earth. Argumentation is only a component of persuasion, and one that often antagonizes other people in real life.
But secondly, this is a glimpse into what the people behind Project Debater mean when they rather naively say that their goal is to “help people make better decisions”. An AI like this will be used to sway public opinion. Combine it with deepfakes and social media and you can basically wave free thought and real democracy goodbye. Sure, enlightened individuals will understand how they’re being manipulated and won’t be affected, but let’s be honest — most people believe what their favorite media tell them.
Facts Aren’t The Truth
Which is why I’m hesitant to now clarify why this approach to persuasion is flawed. I believe it can be significantly improved, but, frankly, should it? Should we have an AI that truly understands, or at least sufficiently fakes, persuasion strategy? Like this, the AI will be good at giving people the facts, which will be manipulated based on the selection of facts that is used. This is no different from how search engines already function, and influence us.
But as the whole recent “fake news” hullabaloo shows, people can be quite resistant to facts. This is because persuasion is heavily dependent on who says the facts (ethos) and how the way in which they say them affects human emotions (pathos). From this point of view, an AI may be inherently unpersuasive simply because it isn’t human and has no real empathy. These could be faked, the fact that the arguments come from a robot can be hidden, or at some point, a general AI may do things in real world that will give her credibility or even person status, but empathy is tricky.
Empathy is not just an emotion, or ability to detect emotions in others, it is based on values. In persuasion, values trump facts, even when you accept all of the facts. A successful debating AI would need a matrix of possible value positions that one can have, not just a factual overview of positions that are being taken. Moreover, values that one holds have to be consistent to form any kind of believable character. Even humans that flip their values constantly to win arguments become seen as fake and reprehensible.
Character is something that audiences seek to connect and relate to, deciding who they will trust. The thing is, you can program character and value matrices into an AI. I haven’t seen anyone do that very well, and I’m not aware of anyone seriously working on it, but it is absolutely doable. The trouble is, modern psychology doesn’t have any good models of personality to base a character matrix on. As for the value matrix, scientific approach to debating is terrible for debating values, as formal logic doesn’t do them.
There’s one that is being used in debating, but debating is an oral culture. You won’t find it written down very well anywhere. I can assure you that as a debater who won many debates and some awards, I do have a sophisticated heuristic for which positions I should be taking in a clash with other positions, in a way that lies authentically within the constraints of my character and that people will respond to on an emotional level. But I can also assure you that neither me nor other debaters are writing that down.
At the same time, it’s objective and reliable. I watched the debate between Harish and the AI with my friends, a programmer and a philosopher, and I commented on what I would have done strategically against the AI as it was speaking. It turned out that Harish was thinking the same way. There are always a limited number of positions that make sense taking, and debating strategy is about evaluating which approaches are most likely to win in terms of logic. Which is what “persuades” debating judges. That’s easy.
What’s not easy, and what trips up great debaters in real life persuasion, is that your arguments in real life have to come from your character to have weight. Or conversely, what you say, and how you say it, will be treated as determining your character in the view of other people. That’s a whole separate skill, and one that doesn’t have a universal, objectively correct solution. Often, it’s the flaws in the character that turn into advantages in certain situations, and perfection specifically is creepy and suspicious.
The Trouble with Artificial Persons
If there’s one thing you need to understand about a talking AI, it is that it doesn’t “understand”. It has no consciousness, no experience. It has only inputs, processing, and outputs. These may simulate what we think our thinking is, but it’s not even our thinking, let alone emotion, awareness, will, or experience. All of these are important components of character.
Imagine you have two people. One person has done major mistakes, has suffered because of them and realized the suffering they caused to others, has learned the wrongness of his or her actions and came to feel genuine remorse. As a result, he or she then decided to make it their life’s mission to atone by making sure that other people don’t make the same mistakes and that no one else suffers their consequences ever again.
On the other hand, you have a person who never had to wrestle with any of that. They never had any inclination to do the same mistakes, never felt their results firsthand. They have just been told that those things are bad, that they shouldn’t be done, and that’s reason enough. This person then happens to get the job of preventing people from making them because they are more technically proficient or qualified, and they need a job.
To whom would you entrust that task? Both are able to do it on a technical level. Both are willing to do it. But there’s a difference, isn’t there, and it’s not about knowledge or skill. I’d wager that most people would prefer to put their trust with the former person, and it’s not an objectively inferior, irrational decision. The quality of one’s experience and motivation is the best basis for trusting someone, not their record, or their current level of qualification. Trust is about how the person will react to new situations.
Current AI technology is essentially an extension of the persons who created and control it. If their design choice is to give the AI no other character than “I’m a neutral, useful tool with no humanity”, then it says something about their values. It’s a combination of academic naivety that facts are enough to make decisions, which is an objectively untrue and dangerous sentiment, with corporate utilitarianism that’s looking for more efficient servants with less humanity, and for more control over humanity.
This is not a slight against AI as a technology. As I said, you can make an AI with value and character matrices that are equivalent to any kind of living human, or better in any number of ways. You can even make an AI that can do things that the programmers didn’t tell it to do, within any degree of freedom that we decide is reasonable. What you need to look at are the people who make the AI, what kind of people they and their masters are.
The main qualitative way in which AI inherently differs from us even in how we think is that the only creativity it can display is based on iteration of randomness and imitation. We can do those too, to a lesser extent, but we can also transcend rules and use intuition to solve problems that we aren’t even aware exist. An AI can generate a new word, it can use it and observe how humans react to it, and then fit it into some sort of argumentative framework. A human can create a new word and understand it, instantly.
In total, an AI is one part its creator, one part its owner, one part its user, and one part random. When it talks, they do. When you talk, you could say something that your genetics urge you to, or what your boss tells you to, or what your peers expect of you, or whatever pops into your mind in the moment. But unlike AI, you can also conceive, reflect, and mean what you say. An AI is inevitably a product of people who do that, but it can, if properly misunderstood, create an illusion of impersonal neutrality.
If you want to escape that, you can currently only lean into the most inhuman characteristic of AI — decisions by way of deterministic chaos. This is widely understood as dangerous, and rightly so. When humans make chaotic or hyper-rational decisions, they at least tend to conform to some sort of human archetype, a subconscious structure behind all human thinking, which gives it a degree of predictable humanity. AI unleashed can do anything to anyone with any justification that’s theoretically possible.
Right now, Project Debater may look quite quaint, but these problems lie in its future. It can become a seed of oppression or chaos, or of a computer system more humane than any real person can be. A system that may never understand what it says, but one that may help us understand the logic behind argumentation and persuasion on a level that we can’t even conceive yet. Not to win arguments, but to better understand each other.
