麻豆传媒

IN DOUGLAS ADAMS鈥橲 novel Dirk Gently鈥檚 Holistic Detective Agency, a computer program called Reason can retroactively justify any decision, providing an incontrovertible argument that whatever was decided was the right thing to do. The software proves so successful that the Pentagon buys it lock, stock and barrel, shortly before a dramatic increase in public approval of military spending.

We鈥檙e not quite there yet. Machines may have beaten us at remorselessly logical games like chess and Go, and they are increasingly giving us a run for our money at games of bluff and chance like poker. But no computer has ever come close to beating humans where it counts: in an argument.

Were one ever able to do so, it isn鈥檛 just the ears of the military that would prick up. The first wave of artificial intelligence, able to crunch huge amounts of information and spot interconnections ever more efficiently, gave us search engines such as Google. A machine capable of formulating an argument 鈥� not just searching information, but also synthesising it into more or less reasoned conclusions 鈥� would take the search engine to the next level. Such a 鈥渞esearch engine鈥� could aid decision-making in arenas from law to medicine to politics. And with an array of ongoing projects looking to build an argumentative streak into AI, it seems only a matter of time before we鈥檒l be testing our mettle against silicon here, too.

Arguing is something humans are peculiarly good at. From polite disagreements over the dinner table to vein-popping run-ins over a parking space or presidential politics, exchanging contrary views is what we do. 鈥淭here are few conversations in which not a single argument is exchanged,鈥� says , a cognitive psychologist at the University of Neuchatel in Switzerland. 鈥淚t鈥檚 a behaviour that鈥檚 omnipresent.鈥�

That is unlikely to be a mere cognitive sideshow. As the world our ancestors lived in grew in complexity, individuals who questioned the truth of each other鈥檚 claims would have had a powerful evolutionary advantage. Follow that argument, and argumentativeness could be the fount of all rational thought: our ability to ponder a situation鈥檚 pros and cons may have originated in rehearsals for these showdowns. 鈥淵ou are anticipating an argument you might have with someone else,鈥� says Mercier.

And that鈥檚 a fact

The social roots of human argumentation make it tough for an artificial intelligence not programmed to think like we do. Even IBM鈥檚 Watson, the supercomputer that, in 2011, wiped the floor with two human champions of the long-running US quiz show Jeopardy!, was only demonstrating an unimaginative ability to answer factual questions. Primed with over 200 million pages of content drawn from books, film scripts and encyclopedias, Watson was trained to analyse what a questioner wanted to find out and locate the answers in its databanks.

In the messy real world, such techniques can only get you so far. 鈥淎 lot of the questions we encounter in life are not factual,鈥� says at the IBM Haifa Research Lab in Israel. 鈥淨uestions where there is no one clear answer.鈥�

Since the Jeopardy! success, Slonim has been collaborating with the Watson team to see whether a machine could graduate from facts to arguments. Ask it, for example, 鈥淪hould violent video games be sold to children?鈥�, and instead of presenting you with links to other people鈥檚 opinions, it would synthesise facts into arguments for and against the idea.

鈥淚n an argument, it is rarely facts alone that win people over鈥�

Users would still have to decide which arguments to trust, of course, just as we decide which links to trust. But in a world where we are often swamped with information, an argument engine could save lawyers, for example, the hassle of trawling through vast archives in search of legal precedents, when a simple press of a button would produce an ironclad summary. Doctors could plug in symptoms and get robust recommendations from case histories on file. Companies might use machines to create arguments for buying their wares. Politicians could secretly test the strength of their manifestos. We might even consider consulting an argument engine before we vote.

All of that means Slonim, once ploughing a solitary furrow, is no longer working alone. 鈥淚 started by saying, let鈥檚 build this machine that can generate arguments,鈥� says Slonim. 鈥淚n 2011, it was only me.鈥� Now he has a team of more than 40 people, and research groups are springing up around the globe.

The first question Slonim鈥檚 team had to tackle was what an argument is, logically speaking. A rough answer might be that it is a claim backed up by evidence. But then the word claim itself could do with a definition. Producing a foolproof spotter鈥檚 guide for an AI is surprisingly tough.

By way of illustration, let鈥檚 return to the argument for or against selling violent video games to children. The statement 鈥淰iolent video games increase children鈥檚 aggression鈥� is a relevant claim, and can possibly be backed by evidence. The statement 鈥淰iolent video games should not be sold to children鈥�, however, is a statement of opinion, and just reiterates one side of the argument. Something like 鈥淕overnments should not restrict the activities of free citizens鈥�, meanwhile, might be a relevant claim, but its connection to the topic is not immediately obvious.

To train Watson in such distinctions, Slonim and his team turned to Wikipedia, surmising that the online encyclopedia鈥檚 entries would be a rich source of claim and counterclaim. It turned out to be a gargantuan task 鈥� less like looking for a needle in a haystack, says Slonim, and more like searching for specific pieces of hay. 鈥淚n Wikipedia you have something like 500 million sentences,鈥� says Slonim. 鈥淎nd a claim is not a sentence. A claim is usually hiding within a single sentence.鈥�

The work has begun to identify key features that set claims apart from generic statements. Claims, for example, are more likely to mention specific times and places, and to include sentiment words such as 鈥渆xceptional鈥� or 鈥渟trong鈥�. Later, the team hopes to shift their focus to flagging up evidence that supports the claim, as well as teaching the system to distinguish between anecdotal data and expert testimony 鈥� and learn how much weight to give different forms of evidence.

That鈥檚 all very well when we want a logical, dispassionate assessment of the facts 鈥� but Slonim is the first to admit that it鈥檚 rarely facts alone that win people over. 鈥淲hen you are debating with another human being, there is a lot of emotion that somehow influences the discussion,鈥� he says.

Once more with feeling

In his 4th-century BC , the ancient Greek philosopher Aristotle distinguished arguments rooted in facts and figures (which he called logos) from appeals that rely on the speaker鈥檚 credibility and expertise (ethos) and statements playing on an audience鈥檚 emotions (pathos). All three strands are readily discernible in public debate today. The successful campaign for the UK to vote to leave the European Union was arguably a triumph of pathos over logos; when Donald Trump punctuates his speeches with the refrain 鈥渂elieve me鈥�, he is employing ethos, urging listeners to respect his authority.

Any artificial intelligence that aspires to rise beyond a mere fact-driven research machine to become a fully fledged 鈥渁rgument machine鈥� 鈥� one that doesn鈥檛 just argue, but argues with human guile 鈥� must master these elements of argument, too. But why would we want one?

鈥淎rgumentation is what makes us capable of resolving conflicts,鈥� , an AI researcher at Imperial College London, said in a recent lecture on the subject. 鈥淎nd machines capable of this could help us evaluate conflicts easier, better, avoiding mistakes.鈥� Chris Reed, an AI researcher at the University of Dundee, UK, thinks that鈥檚 a little utopic, but that argument machines could help raise the level of public discussion. 鈥淲e should be building a technology to facilitate and encourage good quality argumentation and debate,鈥� he says.

In part that would be to counter the existing effects of technology. Easier access to information perversely means we tend to retreat into echo chambers of our own making, where our search histories and social media feeds dictate the information we consume. As a result, opinions and prejudices can become more entrenched. 鈥淭he growth of social media has radically awakened our individual expressive capacity,鈥� says of the London-based think tank Demos. 鈥淏ut it hasn鈥檛 allowed us to compromise any better.鈥� Reed agrees. 鈥淭his is a deep structural problem. It鈥檚 really hard 鈥� even if you鈥檙e very motivated 鈥� to build up a coherent picture of the arguments pro and con on a particular debate.鈥�

Over the past few years, he and his team have been on a mission to seek out good arguments, then dissect and rework them into a form that can be used to train an AI to argue like we do. It鈥檚 a quest that has led them to some unexpected places. The boisterous debates in the UK parliament, for example, are not good reference material: too much performative swagger, and too many procedural interjections and references to previous debates. 鈥淭here is much less quality argument there than you would either expect or hope,鈥� says Reed. Some online forums, on the other hand, contain surprisingly well-structured arguments, despite users being inclined to wear their heart on their sleeve.

Reed鈥檚 favourite source is the BBC radio show Moral Maze, in which panellists debate the ethics of an issue of the day. Its quasi-legal cut and thrust, laced with pathos and ethos, is just the thing from which to build a general framework for the essence of human argument. By analysing and providing a classification of the sorts of arguments we use, and how they relate to one another, Reed and his team aim to produce a tool that can then be used to train an AI.

Back in July 2012, they performed their first real-time argument analysis, of a Moral Maze episode on the ins and outs of British colonial rule in Kenya. Claim and counterclaim, and the connections between them, were represented on a giant touchscreen in a form ready to be fed into an AI. It was an oddly satisfying moment, says Reed. 鈥淚n the same way that a director can appreciate a nicely configured shot, I can appreciate a nicely configured argument,鈥� he says.

鈥淥ut-and-out negation of opposing views is a rarity in human argument鈥�

His team has since repeated the exercise many times, dissecting episodes of Moral Maze and other broadcast and print sources, plus some online forum postings, and turning them into a public databank of argument maps, accessible at . Analysis is ongoing and the results are not yet published, says Reed, but plenty of insights about the way we tend to argue are emerging. When putting forward a hypothesis that we hold particularly dear, for example, we tend to phrase it using questions rather than statements so as to save face if our argument is rejected. Likewise, out-and-out negation of opposing views tends to be avoided in favour of less confrontational challenges. We rarely ask the bluntest question we can 鈥� 鈥渨hy?鈥� 鈥� to elicit justification.

Reed and his team have also started collaborating with IBM in an attempt to build Watson鈥檚 familiarity with webs of human reasoning. Meanwhile, a project started by and at the Technical University of Darmstadt in Germany aims to go a step further, analysing not just what sort of arguments we use, but which ones are the most effective. Earlier this year, they asked nearly 4000 people to say which of two arguments, each making the same case in different ways, they found the more convincing 鈥� and to explain why. Over 80,000 responses later, they now have a database that can be used to teach computational systems to rank the arguments they process, and so argue more convincingly. 鈥淭o me the goal is to change somebody鈥檚 mind, to persuade them,鈥� says Habernal.

Will we swallow it? A fully blown argument machine sounds as implausible as Douglas Adams鈥檚 tongue-in-cheek Reason software. It seems hard to believe that anyone would trust a machine to tell them, say, how to vote, or suggest what they should think about a certain issue. Then again, if anyone had said two decades or so ago that we would trust an artificial intelligence to serve up and rank sources of information for us, few people would have believed them.

One objection to the idea is that any technology is invariably shaped by the prejudices, witting or unwitting, of its designers 鈥� as witnessed, for example, by the recent furore over online ad pickers that displayed higher-paying jobs to male users. An argument machine trained on material with a skew towards liberal newspapers, say, will probably itself develop a liberal slant in the way it makes a case. 鈥淭he bias in the data we are going to explore will probably be reflected in the output,鈥� says Slonim.

So what would be so different between the output of an argument machine and a mainstream edited media source, with its often obvious bias? Undoubtedly a research engine鈥檚 ability to synthesise large bodies of information as easily as we can perform a web search will find its uses. But as for a fully blown argument machine browbeating us with its superior reason, that would require overcoming that most cussed trait of human nature: we never believe a word anyone says anyway.

This article appeared in print under the headline 鈥淭he debater bots鈥�