What it will take for us to trust AI
Like humans, computers need to behave as we would expect
The early days of artificial intelligence (AI) have been met with some very public hand wringing. Well-respected technologists and business leaders have voiced their concerns over the (responsible) development of AI. And Hollywood’s appetite for dystopian AI narratives appears to be bottomless. This is not unusual, nor is it unreasonable. Change, technological or otherwise, always excites the imagination. And it often makes us a little uncomfortable.
But in my opinion, we have never known a technology with more potential to benefit society than artificial intelligence. We now have AI systems that learn from vast amounts of complex, unstructured information and turn it into actionable insight. It is not unreasonable to expect that within this growing body of digital data — 2.5 exabytes every day — lie the secrets to defeating cancer, reversing climate change, or managing the complexity of the global economy.
We also expect AI systems to pervasively support the decisions we make in our professional and personal lives in just a few years. In fact, this is already happening in many industries and governments.
However, if we are ever to reap the full spectrum of societal and industrial benefits from artificial intelligence, we will first need to trust it.
Trust of AI systems will be earned over time, just as in any personal relationship. Put simply, we trust things that behave as we expect them to. But that does not mean that time alone will solve the problem of trust in AI. AI systems must be built from the get-go to operate in trust-based partnerships with people.
The most urgent work is to recognize and minimize bias. Bias could be introduced into an AI system through the training data or the algorithms. The curated data that is used to train the system could have inherent biases, e.g., towards a specific demographic, either because the data itself is skewed, or because the human curators displayed bias in their choices. The algorithms that process that information could also have biases in the code, introduced by a developer, intentionally or not. The developer community is just starting to grapple with this topic in earnest. But most experts believe that by thoroughly testing these systems, we can detect and mitigate bias before the system is deployed.
Managing bias is an element of the larger issue of algorithmic accountability. That is to say, AI systems must be able to explain how and why they arrived at a particular conclusion so that a human can evaluate the system’s rationale. Many professions, such as medicine, finance, and law, already require evidence-based audit ability as a normal practice for providing transparency of decision-making and managing liability. In many cases, AI systems may need to explain rationale through a conversational interaction (rather than a report), so that a person can dig into as much detail as necessary.
In addition, AI systems can and should have mechanisms to insert a variety of ethical values appropriate to the context, such as the task, the individual, the profession, or the culture. This is not as difficult as it sounds. Ethical systems are built around rules, just like computer algorithms. These rules can be inserted during development, deployment, or use. And because these are learning systems, researchers believe that AI systems can, over time, observe human behavior to fill in some of the gaps.
It is incumbent upon the developers of AI systems to answer these questions in a way that satisfies both the industry and the general public. This is already well understood throughout the technology industry, which is why IBM is working together with some of its fiercest competitors — including Google, Microsoft, Amazon and Facebook — on the “Partnership on AI,” a unique and open collaboration designed to guide the ethical development of artificial intelligence.
Business leaders considering artificial intelligence solutions should include trust and accountability as part of their criteria for adoption. They should be thoughtful about how and where this technology is introduced throughout the organization. And they should work with their technology vendors to identify any unwanted behaviors and correct them if necessary.
But delaying the implementation of artificial intelligence is not an option. We pay a significant price every day for not knowing what can be known: not knowing what’s wrong with a patient, not knowing where to find a critical natural resource, or not knowing the hidden risks in the global economy. We believe that many of these ambiguities and inefficiencies can be eliminated with artificial intelligence.
Artificial intelligence is an undeniably powerful technology. And as with any powerful technology, great care must be taken in its development and deployment. Just as it is our obligation to apply this technology to complex, societal problems, it is our obligation to develop it in a way that engenders trust and safeguards humanity. In other words, building trust is essential to the adoption of artificial intelligence.
And we believe that its adoption is essential to humanity.
This story first appeared in the Harvard Business Review .