A Framework for Responsible Artificial Intelligence

In my last column, I started a conversation on the latest buzz in the technology industry: generative artificial intelligence (AI) based on large language models, or the transformer model pioneered by Google. I talked about a couple of areas in which generative AI technology, marketed by Open AI as ChatGPT or as Bard by Google, could be utilized in a sensitive and regulated domain such as healthcare. Only a short two months have passed since then, but for how fast generative AI is moving, it seems many years have passed. The buzz is louder and many companies are jumping on the bandwagon every day; it is becoming almost impossible to keep track of where true innovation is happening and where a me-too copycat is being introduced. This is a common phenomenon observed when a new wave of technology hits the shore of humanity. Everyone wants to ride the wave for fear of missing out. It becomes difficult to separate noise from real improvements.

Meanwhile, another concept is becoming loud and clear. A group of people – which includes the father of AI, Geoffrey Hinton, as well as CEOs and country leaders – are warning us about the potential dangers of this unbridled race to develop artificial intelligence that could someday, not that far out, surpass the intelligence of human brains. In a recent survey conducted during a Yale CEO Summit, 42% of CEOs expressed their opinion that AI could potentially destroy humanity in the next 5-10 years. In a joint press conference, President Biden and U.K. Prime Minister Rishi Sunak mentioned that both countries are committed to “seize the extraordinary possibilities of this new technological age and do so with confidence.” To that end, they are also convening a global summit on AI safety later this year. With such bold pronouncements, it is obvious that the AI age is upon us like never before, and we cannot walk away from the unfolding of this global phenomenon. In this article, I will dive a bit deeper into the concept of responsible AI, touching on the framework for the core of AI safety that so many leaders are worried about.

Responsible AI: An Intersection of Technology, Psychology, Morality and Ethics

Let’s face it – this is a complex topic. The idea that it could be possible to explain in a matter of a few paragraphs is perhaps overblown. But let me try. What is meant by responsible AI? Techopedia.com defines responsible AI as the “development and use of artificial intelligence in a way that is ethically and socially trustworthy.” Builtin.com makes a separation between responsible AI and ethical AI by telling us not to use those terms interchangeably. In their definition, responsible AI has an impact not just on ethics but also on “individuals, communities, and societies as a whole” in terms of “fairness, transparency, and accountability” so that all harms to humanity could be minimized. Ethical AI, on the other hand, is focused more on ethical considerations such as discrimination and bias, which could impact human rights. I subscribe to the view of Builtin.com’s definition, but I do feel that responsible AI and ethical AI have overlaps, which I hope to describe here.

Not violating human rights in specific use cases is also part of being responsible, and hence, the development of AI needs to follow that principle as well. AI engines, which are rooted in deep learning and transformer models, are immensely complex and trained with billions of data sets scraped from the internet. Many times, consideration to data privacy and adequate attribution are not part of the output. So, when a generative AI model such as ChatGPT or Google Bard produces output based on user prompts, there is no way to know which and whose data it has used to create its answer. This is an obvious violation of data privacy rules. Similarly, inherent biases in the data that skew results in a certain way could also mean certain populations, such as women or people of color, are excluded from decisions made by the AI. If this continues, such practices could cause immense harm and inequitable distribution of opportunities, both within organizations and externally within the broader society.

Mechanics of Implementing Responsible AI

Many organizations are internally developing a responsible AI framework, which is typically a systemic approach that begins with education for everyone within the organization. From the C-suite to engineering and human resources (HR) departments, this education needs to be pervasive. Everyone needs to understand the basics of AI, the use cases for their organization, how their company is using it and the risks that come with it. There needs to be a clear guide rooted as much in the values and principles of the organization as in the policies. Ethical, moral and privacy concerns are to be adequately addressed along with legal and regulatory accountabilities. This is where I see a tremendous cross-pollination opportunity of technology with liberal art concepts of morality, ethics and psychology. No longer will simply understanding computer science or data science be adequate to navigate this complex world of responsible AI development. Technologists will need considerable help to ensure what is being developed meets the standards of humanity.

Healthcare and Responsible AI

The need for responsible AI is necessary in the domain of healthcare, perhaps more so than in any other domain. For a long time, digital health innovators have tried to bring the power of AI to healthcare. As I have previously written, machine learning algorithms have been deployed for reading radiology images, cancer detection, ophthalmology in general and any kind of image processing work where the human eye could miss nuances and patterns embedded in complex images. The next-generation AI – which generative AI, transformer model AI or large language model AI systems are capable of – could do much more than that. In a recent podcast interview with Harvard Medical School assistant professor of epidemiology Andrew Beam and New England Journal of Medicine AI journal co-deputy editor Raj Manrai, Dr. Peter Lee, Microsoft’s corporate vice president of research, said he believes that in the next five years, generative AI will become part of every doctor’s tool set. For example, after-visit summaries will be written by such AI tools. He opined that medicine would allow AI in specific use cases without being completely protective to utilize its potential but while putting up guardrails to make the AI more responsible. If the AI “hallucinates” and generates content that digresses from the truth, Dr. Lee predicts that a second AI model could be deployed to correct such hallucinations. In other words, AI could be used to make AI more responsible! Andrew Beam mentioned that inherent biases exist in the data that the AI is trained on, which could make the predictive output of AI also biased and potentially harmful medically. So, what could be done by us, humans, to erect adequate guardrails of responsibility to keep AI on track?

Framework for Responsible AI

1. Start with a diverse team of humans with varied perspectives, backgrounds and experiences. Eradicating bias from the AI requires inclusiveness of perspectives in the responsibility framework for AI development. The organizational framework will need to require teams with backgrounds in ethics, social sciences, human psychology and morality to come together and collaboratively work with AI engineers to create products that would be responsible to society. It might sound a bit vague now, but organizations could start with baby steps in this direction and tackle certain specific use cases and address those with such varied collaboration.
2. Build accountability as the core of the framework. Any organization that is involved in creating today’s AI systems should take responsibility for the actions of their system. If they have blindly used AI products from another organization without understanding the consequences, accountability is also needed. The responsible AI framework needs to embed accountability so that it can become the guiding principle for all. It is not OK to rush to this AI revolution just for the fun of creating something new. Violation of data privacy laws already has accountability measures in California’s Consumer Privacy Act and the EU’s General Data Protection Regulation. But accountability goes beyond that, especially when we are talking about healthcare, where mistakes could be extremely costly. Even one mistake could prevent AI’s further use in this domain for many years.
3. Improve fairness in the data used to train the AI. Current generative AI technology learns from the data it is fed from the internet. There are many biases and discriminations prevalent in this data. Current AI systems are unable to determine what is biased. This could lead to perpetuating discrimination in all AI-generated recommendations or actions. A hospital HR system using AI might discriminate against hiring certain demographic groups for nurses or add extra bias toward others. It is not an easy task for AI developers to ensure that data used to train such AI systems becomes representative of the real-world population, but this needs to be done to build responsible AI systems. I can foresee a new job description for people with a social science background who would be AI trainers. This job does not properly exist today but hopefully will soon.
4. Build models to demonstrate transparency. Last but not least, the current generative AI system is not yet fully transparent. AI chatbots do not tell us how they derived their output and with what content. Some could argue that it is not important to know this information, but it would be immensely important in medicine. AI needs to explain itself to become transparent. As Microsoft’s Peter Lee mentioned in the podcast, I have also alluded that generative AI models will become part of the medical toolkit in the future, and transparency is a must-have to make that tool stick and become widely adopted with trust. This should be part of the responsible AI framework for AI engineers.

Overall, it is an exciting time. Like Dr. Lee, I am very excited to be part of the AI revolution in both my life and career and feel fortunate to see this rapid progress. I am optimistic but feel trepidation. Our exuberance with the new technology should not overpower our rational cognizance. I am glad to see that the U.S. and U.K. are not treating this lightly and are seriously considering meaningful first steps to establish guardrails. I hope other countries will follow suit and not treat this technology as a means to get ahead of each other. If they do, then all of us would be losers at the end.