ASIA
19 August 2026
AI – beyond the hype
Separating the wheat from the chaff for AI.
The world is awash with commentary and analysis about Artificial Intelligence (AI), and its impact on the economy, society, politics and international relations.
But much of this would be speculative hype, according to Arvind Narayanan and Sayash Kapoor in their book, “AI Snake Oil” which examines what AI can do, what it can't, and how to tell the difference. AI snake oil is AI that does not and cannot work. While AI often works to some extent, it can be accompanied by exaggerated claims by the companies selling it.
The authors seek to clarify and demystify the discussion about AI. They argue that in reality AI is an umbrella term for a set of loosely related technologies. For example, ChatGPT has little in common with software that banks use to evaluate loan applicants, and yet both are referred to as AI. In fact, there is no consensus about what is and isn’t AI, and its usefulness varies greatly depending on the application, task, and context.
They also note that some automation – like autopilot in planes, autocomplete on phones, handwriting and speech recognition, spam filtering, spell-check – was initially considered to be cutting edge examples of AI, but that was no longer the case when the product became “normal”. The authors predict optimistically much of what we call AI today will fade into the background and be seen as normal technology. They note a humorous AI definition: “AI is whatever hasn’t been done yet.”
When ChatGPT was released, it was widely seen as a dramatic leap in AI. In reality, the authors argue that the underlying technology has an eighty-year history and had been advancing gradually – and ChatGPT was only a minor improvement over its predecessor, GPT-3.
The authors argue that predictive AI is a fundamentally error prone technology. They do not argue against the use of predictive AI, but it must be used with great caution. They believe that it is too hard to predict the future, and people can be harmed by ill-founded automated decisions.
They also question the ethics of some AI companies which may misrepresent their “AI” products to induce investors and sell their products. Some AI products are simply immature, unreliable, and prone to misuse. Most AI snake oil is concentrated in the area of predictive AI. They also argue that “Content moderation” by social media companies using AI is ineffective.
The authors are cautiously optimistic about “Generative AI”, which they use in their work. Generative AI generates text, images, or other media based on a description. They believe that it is useful to all knowledge workers. Chat GBT is a well-known example. There have been a lot of advances in the technology behind this and people are discovering many uses for it, but there are also many pitfalls.
One of the most problematic applications of AI has been face recognition. Over the last decade or so, face recognition has become so accurate that it can be used by governments for social surveillance. This is not a case of AI not working, but it working so well that it can be abused. Another issue is AI applications that can take a photo of someone and create a nude image. This has affected hundreds of thousands of women especially since AI companies and policy makers have been slow to recognise the problem. Another issue has been the numerous AI-generated books that can be found on Amazon.
The authors insist that AI is no threat to education, any more than the introduction of the calculator was. With the right oversight, it can be a valuable learning tool. But to get there, teachers will have to overhaul their curricula, their teaching strategies, and their exams. They note the concerns about students using AI to do their homework, with some teachers turning to cheating detection software. But such software doesn’t work and has led to a spate of false accusations of academic dishonesty. In fact, the authors encourage their students to use generative AI, and to learn its pros and cons.
A lot of the superintelligence and catastrophic risks have been greatly exaggerated. We are already in a good place to address some of those risks. A major impact of AI will be on the future of work. Governments must urgently figure out how to strengthen existing safety nets and develop new ones to better absorb the shocks caused by AI and reap its benefits.
Despite its catchy title, this book offers a balanced and thorough analysis of the issues. However, the book is mainly based on the American experience of AI, with little reference to China, the other important player in the AI space. Nor does it explore the role of AI in the current geopolitical great power rivalry.
But much of this would be speculative hype, according to Arvind Narayanan and Sayash Kapoor in their book, “AI Snake Oil” which examines what AI can do, what it can't, and how to tell the difference. AI snake oil is AI that does not and cannot work. While AI often works to some extent, it can be accompanied by exaggerated claims by the companies selling it.
What is AI?
What is AI? One definition is that it is a field of computer science focused on building smart machines that can do tasks normally needing human intelligence. Instead of following hard rules, AI systems learn from data to reason, solve problems, and make decisions.The authors seek to clarify and demystify the discussion about AI. They argue that in reality AI is an umbrella term for a set of loosely related technologies. For example, ChatGPT has little in common with software that banks use to evaluate loan applicants, and yet both are referred to as AI. In fact, there is no consensus about what is and isn’t AI, and its usefulness varies greatly depending on the application, task, and context.
They also note that some automation – like autopilot in planes, autocomplete on phones, handwriting and speech recognition, spam filtering, spell-check – was initially considered to be cutting edge examples of AI, but that was no longer the case when the product became “normal”. The authors predict optimistically much of what we call AI today will fade into the background and be seen as normal technology. They note a humorous AI definition: “AI is whatever hasn’t been done yet.”
When ChatGPT was released, it was widely seen as a dramatic leap in AI. In reality, the authors argue that the underlying technology has an eighty-year history and had been advancing gradually – and ChatGPT was only a minor improvement over its predecessor, GPT-3.
Different categories of AI
Three broad categories of AI are dissected by the authors. Many others could have been selected such as robotics and self-driving cars, but the authors were selective. “Predictive AI”, uses machine learning based on historic data, to predict behaviour for a wide array of issues like the best candidate to hire, the likelihood of criminals committing crimes again, a potential borrower’s credit risk, and which patients should be given priority for an organ transplant. Indeed, it is more like a traditional statistical model that has been rebranded into AIThe authors argue that predictive AI is a fundamentally error prone technology. They do not argue against the use of predictive AI, but it must be used with great caution. They believe that it is too hard to predict the future, and people can be harmed by ill-founded automated decisions.
They also question the ethics of some AI companies which may misrepresent their “AI” products to induce investors and sell their products. Some AI products are simply immature, unreliable, and prone to misuse. Most AI snake oil is concentrated in the area of predictive AI. They also argue that “Content moderation” by social media companies using AI is ineffective.
The authors are cautiously optimistic about “Generative AI”, which they use in their work. Generative AI generates text, images, or other media based on a description. They believe that it is useful to all knowledge workers. Chat GBT is a well-known example. There have been a lot of advances in the technology behind this and people are discovering many uses for it, but there are also many pitfalls.
Challenges for generative AI
It is necessary to incorporate guardrails into generative AI applications. For example, there are issues like hallucinations that plague generative AI systems today. Hallucinations refer to a generative AI system that makes up stuff. In the coming years, research may shift from building ever bigger AI models to reducing the problem of hallucinations.One of the most problematic applications of AI has been face recognition. Over the last decade or so, face recognition has become so accurate that it can be used by governments for social surveillance. This is not a case of AI not working, but it working so well that it can be abused. Another issue is AI applications that can take a photo of someone and create a nude image. This has affected hundreds of thousands of women especially since AI companies and policy makers have been slow to recognise the problem. Another issue has been the numerous AI-generated books that can be found on Amazon.
The authors insist that AI is no threat to education, any more than the introduction of the calculator was. With the right oversight, it can be a valuable learning tool. But to get there, teachers will have to overhaul their curricula, their teaching strategies, and their exams. They note the concerns about students using AI to do their homework, with some teachers turning to cheating detection software. But such software doesn’t work and has led to a spate of false accusations of academic dishonesty. In fact, the authors encourage their students to use generative AI, and to learn its pros and cons.
What future do the authors see for AI?
They note that there are three common narratives: (i) a superintelligence which ushers in a new utopia; (ii) a superintelligence which will doom us; (iii) scepticism about AI which can be seen as an overhyped fad which will pass very soon. The authors steer clear of these scenarios and believe that AI will have transformative effects – like the Industrial revolution, electricity, and the Internet – but they will unfold over a period of many decades, rather than suddenly, with both good and bad effects.A lot of the superintelligence and catastrophic risks have been greatly exaggerated. We are already in a good place to address some of those risks. A major impact of AI will be on the future of work. Governments must urgently figure out how to strengthen existing safety nets and develop new ones to better absorb the shocks caused by AI and reap its benefits.
Despite its catchy title, this book offers a balanced and thorough analysis of the issues. However, the book is mainly based on the American experience of AI, with little reference to China, the other important player in the AI space. Nor does it explore the role of AI in the current geopolitical great power rivalry.