This article is written by Ayca Ariyoruk, a cross-cultural policy and ethics advisor for AI. She is also a fellow at the Center for AI and Digital Policy in Washington, DC.
As part of the recently launched Government AI Campus, Apolitical is publishing a series of articles exploring AI adoption in government. These articles take on many forms, from op-eds written by academic experts to interviews with public sector leaders working on AI adoption themselves.
- The problem: The variation in national and multinational risk management frameworks has complicated the global governance landscape of artificial intelligence. How can policymakers harmonise standards while embracing the world’s social and cultural diversity?
- Why it matters: Artificial intelligence is a globalised technology. A global approach to AI governance would facilitate international collaboration, trade and innovation, ensure ethical consistency and help minimise risks and maximise benefits.
- The solution: Human rights is a universal language. By tapping into the existing international treaties and instruments and applying them through a fresh lens, we see the converging impact of human rights on AI and develop a cross-cultural risk management framework.
Imagine your new refrigerator can make breakfasts based on the supplies and recipes you provided. You bought eggs, cheese and butter; labelled as such, and provided a step-by-step instruction on how to make an omelette. The next morning, you open the fridge door and there it is, a freshly made omelette staring at your hungry eyes. How delightful.
💬 Let the author know what you think about the role of policymakers in effective AI governance by leaving a comment below
Now, imagine the following morning, your omelette looks a little different. There are ingredients in it that you can’t recognise. You didn’t buy mushrooms, so where did they come from? Were they organic? Were they washed thoroughly? Were they stolen or paid for? Whose recipe did the refrigerator follow?
You think to yourself, my refrigerator knows best. After all, this is the latest technology booming in the markets, so it must be trustworthy. As you reach your hand to grab the omelette, there is one more surprise. The refrigerator refuses to release the plate, and you hear a voice: Ayca, you are fat for a woman of your age and height. Skip breakfast. The next morning, your refrigerator refuses to make you a meal.
The key difference between the cheese and mushroom omelettes
Traditional artificial intelligence classifies or predicts outcomes based on input data. Learning depends on labelled data sets and instructions (code) to train models. When your Netflix makes you a recommendation on what to watch tonight, that is traditional AI. Generative artificial intelligence (also referred to as foundational models), in contrast, doesn’t just analyse existing data, it creates new ones that didn’t exist before. It is creative and offers new content that is similar but distinct from training data. That is the difference between the cheese omelette and the mushroom omelette. (For more information on the difference between traditional and generative AI, check out NoCode.ai by Armand Ruiz, Director of Data Science at IBM and the founder of NoCode.ai.)
Now that your refrigerator has managed to tick you off, you decide to investigate further. You call up the manufacturer, demanding an explanation. The company says, no, can’t do. It’s a trade secret. Revealing the secret would undermine the spirit of innovation and market competition, putting economic growth and opportunities at risk. But the company can’t keep the genie in a bottle for too long, and soon enough, you read news reports that the mushrooms in your omelette were synthetic, not even real food. The recipe was stolen from a chef who is now out of a job, unable to provide for their family. The behind-the-scenes work in the production of the omelette involved labour from people who were underpaid and worked under very harsh conditions.
It doesn’t end there. You then find out why your refrigerator (call it the specific application or use case for the technology) had a sexist opinion. The refrigerator was developed by an engineer who had a particular view of how women should look. It was also trained on large but not diverse data sets provided by a middle-class population residing in North America with no existing health conditions. You realise this piece of new technology could pose a significant health risk if used by other demographics. That is, of course, if they can afford the refrigerator in the first place. A smart refrigerator won’t help feed a family or offer productivity gains from the time saved on cooking if the household doesn’t have internet access to begin with. (Today, about 3.7 billion people, almost half of the world's population, don’t have internet access.)
Replace the food diet with an information diet, and it becomes even more clear that AI is not like any other technology. AI has opinions not always based on facts.
Fast forward to the following year and the company has doubled its investments. The smart refrigerator can now make breakfast and deliver a full course menu to the entire neighbourhood in Silicon Valley. As such, it is demanding more ingredients and consuming more electricity. At this point, the company can no longer explain the inner workings of the refrigerator, even if it wants to. How did that input yield this output? It is a mystery. AI researchers call it a “Black Box”. Welcome to the field of **machine learning. **
What is artificial intelligence?
Before getting into the “subfield” of machine learning, let’s have a quick look at the main field. What is artificial intelligence? Believe it or not, there is no scientific consensus on the definition of AI. Even AI researchers can’t agree on a single definition. The field is being constantly redefined as new technologies emerge. The University of Helsinki’s ‘Elements of AI’ course — which is an excellent beginner course for people with non-technical backgrounds — refers to an old joke in defining AI. AI is defined as “cool things that computers can’t do”. Under this definition, AI can never make any progress: “As soon as we find a way to do something cool with a computer, it stops being an AI problem.”
There is some truth to this definition. As ironic as it may sound, says the Elements of AI: “Fifty years ago, for instance, automatic methods for search and planning were considered to belong to the domain of AI. Nowadays, such methods are taught to every computer science student. Similarly, certain methods for processing uncertain information are becoming so well understood that they are likely to be moved from AI to statistics or probability very soon.”
Despite the challenges, for the purposes of regulation, there is now a detailed working definition provided by the Organization for Economic Cooperation and Development (OECD), on which EU lawmakers based the recent EU AI Act. The trick is to come up with a definition that is forward-looking and all-encompassing so as not to leave any technology or use cases out but also not too broad that it becomes impractical.
The risky business of machine learning
For the purposes of the refrigerator story (and at the risk of oversimplifying), you can think of AI as a subset discipline of computer science that uses data to enable problem-solving machines. Machine learning is a subfield of AI that can automatically adapt but depends on minimal human interference to correct and learn. Deep learning is a subset of machine learning that uses large amounts of data and learns from its environment and past mistakes and without human intervention. (For more on the difference, check out Deep Learning vs Machine Learning: A Beginner’s Guide by Coursera).
It’s usually at that deep learning stage that technologies become unexplainable. Questions worth asking: If we can’t explain it, is it science or fiction? At which point does our trust in emerging technologies resemble our relationship with ‘God’, much like a religion and less like a scientific inquiry?
True, just because we can’t provide proof that something exists is not proof that something doesn’t exist.
That’s where AI becomes a risky business. One of the biggest hurdles to responsible AI governance is the variation in national and multinational risk management frameworks for regulation. After all, what we view as dangerous and how much risk we can tolerate around the unknown depends on what we value.
AI is not like any other technology. AI is political.
If AI and risk management are about values, how do we then protect the world’s diversity and at the same time harmonise our approach? Enter the politics of AI, the power struggle to call the shots on the rules of the game.
Techno-determinists claim that AI is just like any other technology, and a progressive impact on society is inevitable. But whether we will advance or regress depends on the outcomes of an ongoing power struggle between i) a powerful tech industry, ii) governments with varying political ideologies and degrees of control over private enterprise and iii) civil society with different levels of influence in government affairs. As much as the industry drives innovation and economic growth, and despite the well-intended efforts to make a business case for responsible AI that serves humanity, the industry is trapped by a business model that puts individual shareholder interest above the collective public interest.
Industry favours self-regulation because it can. Governments, whether liberal or authoritarian, want a playbook that will not undermine their style of power. That makes a difference in how AI powers the government and whether it will use it to undermine or empower its citizens. What do people want?When we look at civil society's activity in AI governance, the answer becomes clear. We want what we have always wanted: protection, promotion and defence of our fundamental rights.
A universal risk framework: “Human rights are foreign to no culture and native to all nations.”
“Human rights are foreign to no culture and native to all nations”, said Kofi Annan, the late Secretary-General of the United Nations. They are universal, indivisible and interdependent as they are “what makes us human”. Ask the heads of state, and they may say that we have to respect cultural variations, which often translates to ‘don’t question my authority’. Ask the people, and we all want the same things: to live with dignity, with the freedom to practise our religions, equal opportunity for education, jobs and social benefits, to have a say in how we are governed and a fair justice system. We are innocent before proven guilty.
Do people have a right to know if their job application was rejected by a Black Box? What if a bank can’t explain why a citizen is denied a loan to buy a home or financial aid for college, fired by an employer, or was arrested because a piece of surveillance deemed them to be a criminal? What if a health gadget meant to measure blood pressure was not working just because of the colour of your skin?
There is nothing controversial or ‘relative’ about human rights when it comes to AI. These are concerns universal to all people regardless of cultural context, and recognising them as such is crucial in maintaining citizens’ trust in public institutions and services. (It is important not to confuse cultural relativism in human rights — a pretext for tyranny — with cultural diversity, which is something to be protected and cherished. The first limits freedoms and the latter expands them.)
True, the air from the Cold War is clouding international negotiations over AI policy, leading some diplomats to view human rights as a divisive topic hard to operationalise. In reality, AI has broken down the Cold War rules. Look at the US Executive Order issued by President Biden in October 2023. There is no way around addressing the risks of AI without embracing social and economic rights. The US president requires the federal government and its agencies not to procure any AI products or services that could, through “fraud, unintended bias, discrimination”, undermine access to “healthcare, financial services, jobs, education, housing and transportation”. Similarly, across the Pacific, President Xi of China called for regulations to protect individual privacy and data security. It is a rare example of the country attempting to limit state interference with the individual. Civil rights and social rights are indivisible and interdependent, especially when it comes to the AI life cycle. We can’t prevent bias and discrimination without due process for public input or feedback from the end user. There is a gradual convergence around the practical application of values for artificial intelligence.
Back to our magical refrigerator story
So the next time that someone tells you AI is like any other technology, I hope you remember our refrigerator story. (The smart refrigerator is imaginary as of yet. But the technologies with comparable capabilities are real.)
Replace the food diet with an information diet, and it becomes even clearer that AI is not like any other technology. AI has opinions not always based on facts. Large language models (generative AI) will write the stories we tell each other and our history if we let it. AI can generate and scale propaganda and deep fakes with harsher consequences than an omelette with synthetic mushrooms. AI won’t lift societies equally everywhere (for more, watch the presentation based on his book Power and Progress, Our Thousand Year Struggle Over Technology and Prosperity by Daron Acemoglu, professor and economist at MIT). Moreover, have a look at the latest Artificial Intelligence and Democratic Values Index published by the Center for AI and Digital Policy to see the evaluation of your country’s comparative performance on AI policies and practices. It covers 80 countries based on 12 metrics grounded in human rights. (You can also apply for the Policy Clinics. They are intensive, interdisciplinary, semester-long courses covering AI history, issues, institutions, regulation and policy frameworks and research methods.)
It is okay not to trust the invisible hands of the algorithms. You won’t be anti-tech. You will be pro-human, making sure our technologies are aligned with the public interest. Lucky for us, we have a choice.
Done reading? Make sure to share your own thoughts on the development of a cross-cultural risk management framework for AI governance by leaving a comment below ⬇️
(Image credit: Pexels)
Log in or sign up to continue the conversation