Monday 7 June 2021


Image: @cvbe

Picture yourself driving on a narrow road in the near future when suddenly another car emerges from a bend ahead. It is a self-driving car with no passengers inside. Will you push forth and assert your right of way, or give way to let it pass? At present, most of us behave kindly in such situations involving other humans. Will we show that same kindness towards autonomous vehicles?

Do we compromise for an AI?

Using methods from behavioural game theory, an international team of researchers at the Ludwig Maximilian University of Munich (LMU) and the Institute of Philosophy at the School of Advanced Study, University of London, have conducted large-scale online studies to see whether people would behave as cooperatively with artificial intelligence (AI) systems as they do with fellow humans.

Cooperation holds a society together. It often requires us to compromise with others and to accept the risk that they let us down. Traffic is a good example. We lose a bit of time when we let other people pass in front of us and are outraged when others fail to reciprocate our kindness. Will we do the same with machines?

Researchers and the public tend to think about human interactions with AI as competition.  Take the rivalry between chess grandmaster Gary Kasparov and the IBM computer Deep Blue, or Lee Sedol and Google’s AlphaGo program, or, for a more hostile version, the chase between Sarah Connor and the Terminator. Cooperation remains the dark matter: can humans and AI work together for mutual gains, and not just see who wins?

Exploiting the machine without guilt

The study, ‘Algorithm exploitation: humans are keen to exploit benevolent AI’, which was published on 3 June in the journal iScience, found that, upon first encounter, people have the same level of trust toward an AI as for human: most expect to meet someone who is ready to cooperate.

The difference comes afterwards. People are much less ready to reciprocate with an AI, and instead exploit its benevolence to their own benefit. Going back to the traffic example, a human driver would give way to another human but not to a driverless car.

The study identifies this unwillingness to compromise with machines as a new challenge to the future of human-AI interactions.

‘We put people in the shoes of someone who interacts with an artificial agent for the first time, as it could happen on the road,’ explains Dr Jurgis Karpus, a behavioural game theorist and a philosopher at LMU and the first author of the study. ‘We modelled different types of social encounters and found a consistent pattern. People expected artificial agents to be as cooperative as fellow humans. However, they did not return their benevolence as much and exploited the AI more than humans.’

With perspectives from game theory, cognitive science, and philosophy, the researchers found that ‘algorithm exploitation’ is a robust phenomenon. They replicated their findings across nine experiments with nearly 2,000 human participants.

Each experiment examines different kinds of social interactions and allows the human to decide whether to compromise and cooperate or act selfishly. Expectations of the other players were also measured. In a well-known game, the Prisoner’s Dilemma, people must trust that the other characters will not let them down. They embraced risk with humans and AI alike, but betrayed the trust of the AI much more often, to gain more money.

‘Cooperation is sustained by a mutual bet: I trust you will be kind to me, and you trust I will be kind to you. The biggest worry in our field is that people will not trust the machines. But we show that they do!’ notes Professor Bahador Bahrami, a social neuroscientist at the LMU, and one of the senior researchers in the study. ‚They are fine with letting the machine down, though, and that is the big difference. People even do not report much guilt when they do,’ he adds.

Benevolent AI can backfire

Biased and unethical AI has made many headlines – from the 2020 exams fiasco to justice systems – but this new research brings up a novel caution. The industry and legislators strive to ensure that artificial intelligence is benevolent. But benevolence may backfire.

If people think that AI is programmed to be benevolent towards them, they will be less tempted to cooperate. Some of the accidents involving self-driving cars may already show real-life examples: drivers recognise an autonomous vehicle on the road and expect it to give way. The driverless vehicle meanwhile expects for normal compromises between drivers to hold.

‘Algorithm exploitation has further consequences down the line. If humans are reluctant to let a polite self-driving car join from a side road, should the self-driving car be less polite and more aggressive in order to be useful?’ asks Dr Karpus.

‘Benevolent and trustworthy AI is a buzzword that everyone is excited about. But fixing the AI is not the whole story. If we realise that the robot in front of us will be cooperative no matter what, we will use it to our selfish interest,' says Professor Ophelia Deroy, philosopher at the LMU and University of London, and senior author on the study.

‘Compromises are the oil that make society work. For each of us, it looks only like a small act of self-interest,’ continues Professor Deroy, who also works with Norway’s Peace Research Institute Oslo on the ethical implications of integrating autonomous robot soldiers along with human soldiers. ‘For society, it could have much bigger repercussions.’

One thing is for sure, our competing drives to treat machines as partners, and as tools, will continue to run and run, and Algorithm Exploitation is a fascinating look at where we are on that journey.

For further information please contact Professor Ophelia Deroy (