When is it OK for AI to lie?

Artificial intelligence has captured the public imagination by besting chess grandmasters and one-upping game show contestants. On top of that is the Silicon Valley hype, not to mention the doomsday science-fiction scenarios of machines taking over humanity like Skynet in "The Terminator" and or the increasingly sentient robots of "Westworld."

But for Arizona State University computer scientist Subbarao Kambhampati, another big milestone of AI’s development will be in its ability to lie — even the little white ones. He sees it as a natural progression as algorithms and neural networks grow ever more sophisticated.

“We always tell each other white lies,” said Kambhampati, a professor in the School of Computing, Informatics, and Decision Systems Engineering within the Ira A. Fulton Schools of Engineering. “It’s part of the societal lubricant. For example, if someone fixes you some food to eat, you are not supposed to say, ‘That sucks!’ Otherwise, the society falls apart. So, the question then, is, under what conditions are people willing to let AI systems tell them white lies?”

At this week’s conference on Artificial Intelligence, Ethics and Society, Kambhampati, along with his former graduate student Tathagata Chakraborti, presented different scenarios exploring when — and if — it would be permissible for AI to lie, and how to keep humans in the loop if they do.

“First of all, this comes from our research, which says that once I have a mental model of you, I can manipulate it and tell you lies," Kambhampati said. "In fact, I joke in many of these talks that I was ecstatic when my kid could tell his first lies, since that is a sure sign of intelligence.

“Telling lies is a way to get you to believe in an alternate reality which can lead to personal gain or greater good. Since we are making these AI systems, we can control when they can and cannot fabricate, or essentially tell lies.”

The researchers designed a thought experiment to explore both human-human and human-AI interactions in an urban search-and-rescue scenario: searching all locations on a floor of an earthquake-damaged building. They enlisted 147 people through crowdsourcing on Amazon’s Mechanical Turk to survey how human reactions change between dealing with humans or AI.

Results of their thought experiments and surveys indicate that public perception is positive toward AI lying for the greater good. Fabrication, falsification and obfuscation of information can be used by an AI agent to achieve teaming performance that would otherwise not be possible.

Subbarao Kambhampati

“In this paper, we attempted to take the first steps towards understanding the state of the public consciousness on this topic,” Kambhampati said. “We got a sense of when people are willing to be told white lies.”

But it poses several unresolved ethical and moral questions with regards to the design of autonomy in AI, which the research group will continue to explore.

At the end of the presentation, they discussed scenarios where white lies are considered acceptable, as in certain circumstances in the doctor-patient relationship. AI is already rapidly coming to the forefront in pathology and imaging diagnostics, such as interpretations of X-rays and CT scans for cancer.

But what will be the role of AI in the future of medicine? After all, the Hippocratic Decorum states – “Perform your medical duties calmly and adroitly, concealing most things from the patient while you are attending to him.”

Medical lies are usually done to give as much truth as is good for the patient, especially in the delivery of bad news. This is obviously done for the good of patient, to help maintain a positive attitude even if given a late stage cancer diagnosis.

“The rationale, here, being that such information can demoralize the patient and impede their recovery,” Kambhampati wrote. “As we saw in the study, participants were open to deception or manipulation for greater good, especially for a robotic teammate.”

But there are also known deceptions like the placebo effect. And what if it is a medical AI that is concealing information from a patient, or the results of cancer medical imaging tests from a doctor?

The future of AI in medicine will also involve trust. And perhaps, there will be times when it makes sense for an AI to lie to a person.

“The doctor-patient relationship, and the intriguing roles of deception in it, does provide an invaluable starting point for conversation on the topic of greater good in human-AI interactions,” Kambhampati said.

In addition to the ASU talk, the AIES conference provided a platform for research and discussions from the perspectives of several disciplines to address the challenges of AI ethics within a societal context, featuring participation from experts in computing, ethics, philosophy, economics, psychology, law and politics.

“AI is evolving rapidly, and as a society we’re still trying to understand its impact — both in terms of its benefits and its unintended consequences,” said conference co-chair Vincent Conitzer, of Duke University. “The AIES conference was designed to include participation from different disciplines and corners of society, in order to offer a unique and informative look at where we stand with the development and the use of artificial intelligence.”

The AIES conference was chaired by a multidisciplinary program committee to ensure a diversity of topics. Conference sessions addressed algorithmic fairness, measurement and justice, autonomy and lethality, human-machine interaction and AI for social good, among other focuses.

AIES presenters and about 300 attendees include representatives from major technology and nontechnology companies, academic researchers, ethicists, philosophers and members of think tanks and the legal profession.

View a complete list of research papers and posters presented at the AIES Conference. The proceedings of the conference will be published in the AAAI and ACM Digital Libraries.

The conference was sponsored by the Berkeley Existential Risk Initiative, DeepMind Ethics and Society, Google, National Science Foundation, IBM Research, Facebook, Amazon, PWC, Future of Life Institute and the Partnership on AI.