Anne Fehres and Luke Conroy & AI4Media / A Better Image of AI / Data is Our Mirror / CC-BY 4.0
Suddenly, AI tools have language fluency on par with most humans. So why not just leave it to ChatGPT whenever you feel like communicating? Arguing with your spouse, nagging your kids, or replying to trolls?
No. While it’s up to you to decide what to do with your family, online debates shouldn’t be outsourced to AI. There’s been excited chatter about this possibility, with several university teams building AI tools to combat digital hate.
Meanwhile, people are practicing counterspeech, defined as a response to hate intended to undermine it. Despite the usual “don’t feed the trolls” mantra, many people consistently respond constructively to online hate. They should continue to be encouraged to do so, and we should join in more. Bots certainly have great advantages: they don’t feel disgust, fear or hurt, and they can operate at a scale that humans cannot match. But AI cannot do something essential to democratic civic and political life: engage and debate with other people, including those whose views we abhor.
“[P]”Public debate is a political duty,” U.S. Supreme Court Justice Louis Brandeis famously wrote in his 1927 decision in Whitney v. California. “The founders of the United States believed that it was dangerous to stifle thought, hope, and imagination,” he continued. “(All the premises of debate) are beyond AI’s control… The way to safety lies in the opportunity for a free discussion of supposed grievances and proposed solutions, and the proper solution to bad advice is good advice.” These ideas are the basis for counterspeech, and the basis for the U.S. Constitution’s zealous protection of free speech.
Empty promises
Imagine you’re organizing a protest for a cause you believe in. You recruit other demonstrators, draw placards, and borrow a megaphone. Then, on the day of the event, it rains. You don’t want to get wet, so you decide to send a robot to the march instead and watch the livestream.
No matter how many robots there are or how loud they shout slogans, we know this is a bad idea because the act of responding matters, not just the content of the response. It is a civic act that has visible consequences for the person or people who act, and for the audience who reads and listens.
Our research shows that counterspeech can have a strong positive impact on audiences and discourse norms. By lifting the inhibitions of silent “lurkers,” counterspeech can generate more counterspeech, even if it doesn’t change anyone’s mind. People often feel safe to speak up when they see others doing so. And for those who are being slandered and attacked online, counterspeech can provide a much-needed sense of support and reassurance. Messages generated and delivered by AI, no matter how warmly worded, cannot inspire the same sense of solidarity.
Finally, as we learned from interviewing many of them, opponents themselves often benefit from speaking out: They say that speaking out often feels like a duty, and it makes them braver and more active not just online but offline, too. Their online practice makes them more willing to speak out against hatred offline, too.
Genuine engagement is effective
At the Dangerous Speech Project, our independent research team, we have identified thousands of active counterspeakers around the world and studied many of them and their efforts. In general, their goals are similar: their primary objective is not to change who they respond to, but to shift discourse norms among those who witness their activity. But their methods are markedly different: they are cunning, ingenious, humorous, empathetic, mocking, indignant, and sometimes unpredictable in switching from one technique or tone to another. These are qualities that AI cannot match without active human involvement.
Take the German journalist Hasnain Kazim. After receiving a flood of messages from readers attacking him for his name, skin color, and supposedly Islamic faith, he set himself the task of replying to as many of them as possible. He sometimes writes long, detailed letters aimed at enlightening people who ask questions like, “Are you in favor of headscarves?” And he can be, to say the least, profane at times. Several readers have asked him if he eats pork, apparently considering it a reliable measure of Germanness. “No,” Kazim replied to one of them. “Only elephants (well-roasted) and camels (bloody).”
When another reader asked, “Am I correct in understanding that you are against Islamism?” Kazim replied, “Yes, I am against it, except when extreme militant vegetarianism gains the upper hand. In that case, I would hope that the Salafists would stop it.”
“Should the Salafists stop this with violence?” a reader responded (it seems fair to point out that he was exhibiting an AI-level sense of humor).
“No, Salamis,” Kazim replied.
“Salami is pork, Mr. Kazim, you should know that! Salafists don’t eat salami!”
Kazim’s conversations with his German readers so amused other Germans that in 2018 he published a book compiling and explaining them, which became a bestseller. The title (“Emails from Karl-Heinz: Angry Messages from Real Germans and How I Responded to Them”) comes from the pseudonym used by a reader who sent Kazim spiteful letters, accusing him, ostensibly a foreigner, of trying to “teach us Germans.”
“Come where I live and I’ll show you what real Germans are like!” Kazim had in fact been born and raised in a small town in Germany and had served as an officer in the German navy. He made no mention of this in his reply, simply saying that he was happy to accept the invitation and would be arriving shortly in two motorcoaches with three of his four wives, their eight children, 17 cousins, their 22 children and three goats. “We are all very excited to learn from you what ‘real Germans’ are like!”
AI can’t write like that, at least for now. It can’t organize thousands of counterspeech voices together, as the #iamhere group did. It can’t express, as writer and actor Dylan Maron calls it, “radical empathy,” in his podcast and book, Conversations with People Who Hate Me. We’ve found countless examples of counterspeech that are too agile, too sensitive, too empathetic, too funny, too human, for an AI to replicate.
This doesn’t mean AI can’t help counter digital hate. Counterspeech can be time-consuming and mentally taxing for those who undertake it. Well-designed AI tools can alleviate some of these burdens. But they should be used to amplify the efforts already being made by thousands of people, and importantly, AI tools should be delivered in a way that actually helps counterspeakers (and potential counterspeakers). To find out what that is, you need to ask them. This is a step that nearly all developers and researchers skip. We haven’t skipped it.
We worked with a team led by Professor Maarten Sapp of Carnegie Mellon University and Professor Joshua Garland of Arizona State University to interview a diverse group of experienced counterspeakers and survey people who use social media on a daily basis. In a study led by doctoral student Jimin Moon, to be presented at CHI 2024, we found that contrary to the assumptions of many other researchers, most experienced counterspeakers don’t struggle with what to write and don’t need an AI coach to suggest language for their posts. Instead, they say they want help finding the content they want to respond to, because that way they can navigate it much more efficiently.
For example, consider the members of #iamhere, the world’s largest counterspeak group. #iamhere is an international coalition of over 150,000 people from 16 countries who collectively respond to comments they deem hateful on public Facebook pages. The group’s moderators spend hours each week finding such content. AI can help spot hateful speech, enabling the group to increase the number of posts it responds to.
Meanwhile, many social media users who aren’t (yet) contrarians expressed that they would welcome an AI coach to suggest how they should respond. Such a tool could help contrarians overcome their reluctance to engage.
People we interviewed and surveyed also questioned the ethics of automating counterarguments and whether AI-generated replies would be as effective as human-written ones, but most agreed that AI should be used to support humans, not replace them.
Future work
More research needs to be done to assess the need for counterspeakers in different contexts and the impact of automated responses to hate. For example, how do the barriers faced by counterspeakers responding to hate under authoritarian governments differ from those in democratic societies? Also, can audiences distinguish counterspeech written by bots from messages written by humans? If so, is there any difference in effectiveness?
To create useful tools that can work at scale without undermining what human counterspeakers already do, it is essential to study questions like this with counterspeakers working in a variety of settings and with a variety of languages.
