The future of human-AI interactions is set to get fraught. With the push to incorporate ethics into artificial intelligence systems, one basic idea must be recognised: we need to make machines that can say 鈥渘o鈥� to us.
Not just in the sense of not responding to an unrecognised command, but also as the ability to recognise, in context, that an otherwise proper and usable directive from a human must be refused. That won鈥檛 be easy to achieve and may be hard for some to swallow.
As quickly as artificial intelligence has been spreading, we are still only at the beginning of the revolution. Far from making autonomous decisions in isolation, many sophisticated artificial intelligence systems will be working alongside, or under the direction of, people for some time yet.
Advertisement
AI that rather than usurps us is increasingly talked about as the preferred trajectory for this technology. But, crucially, that means if AI is to be ethical, it will have to be able to refuse to carry out illegal or unethical commands from a person.
As systems become more powerful and more prevalent, the question of whether or not they will behave ethically has become more salient. Fortunately, research on this abounds in and . DeepMind, the London AI research group acquired by Google in 2014, is just the latest to launch an ethics project, bringing in a broad assortment of advisers to help examine the dilemmas arising from .
But as we increasingly explore the underlying decision-making process for self-driving cars, reflect on whether military robots should be permitted to make kill/no kill decisions and try to unmask the presence of embedded racism and sexism in machine learning algorithms, we need to step back and recognise the context and complexity of what we are trying to do.
Resistance is useful
In nearly every situation, the decision made by a machine can be traced back to a request made by a human. For a machine to behave ethically, it must be able to recognise if a request is itself ethical 鈥� perhaps it must even resist attempts to override its decision.
Humans are imperfect. Consequently, we have begun to see arguments in favour of the idea that military robots would do a better job than people at deciding whether or not to destroy a target. Such arguments recognise the role of cognitive errors in human behaviour.
Similarly, on their own, combat AI would not engage in atrocities driven by rage or hatred. But what if an AI-controlled drone is ordered to attack a village? US soldiers are taught to recognise and refuse illegal orders. Military AI should operate with the same rules.
The same should happen beyond the military sphere, whether we are talking about a financial AI approving or rejecting loans, a self-driving vehicle attempting to avoid dangerous road conditions or a medical system dispensing painkillers. An authorised human user could direct the AI to do something that would be considered problematic or even illegal if done by a person, such as make it harder for minorities to get a loan, take a closed or hazardous route, or overprescribe opioids.
Unless that AI could recognise and refuse such commands, we risk building a world of 鈥渋ntelligent鈥� machines that function as just another vector for human flaws, but one obscured by a sheen of computerised objectivity.
A refusal may sometimes offend, but an ethical AI future will mean getting used to machines that know when to say 鈥渘o鈥�.
Read more: Moral dilemma: should we reshape society because we can?; 7 mind slips that cause catastrophe 鈥� and how we can avoid them
