Machines That Think: How Artificial Intelligence Works and What It Means for Us is available for purchase from Rheinwerk Publishing/SAP Press. ASUG members can enjoy 15% off any SAP Press titles with the discount code 15ASUG. 

When Inga Strümke, Associate Professor at Norwegian University of Science and Technology (NTNU), wrote Machines That Think: How Artificial Intelligence Works and What It Means for Us, a Rheinwerk Publishing/SAP Press publication, she wanted the book to be accessible to readers with all levels of familiarity with AI technology. Although her editor was initially skeptical, she stuck to this idea and has seen it pay off. 

“The reason is simply that AI affects all of society, and society consists of everybody,” Strümke said. She has received messages from readers saying that they found the book was useful for the board members at their company, but also for their grandmother, “which I think is the greatest compliment you can give to an author of a technical subject.”

In the book, Strümke walks readers through the early days of computing, providing a historical foundation for AI advancements and missteps. She poses philosophical questions around bias and regulation, while also sharing context-rich information around symbolic versus non-symbolic AI, neural networks, and what we can expect next. 

This interview has been edited for brevity and clarity.

ASUG: What do you think is the biggest misconception about AI today?

Strümke: In all honesty, [the biggest misconception is] that current AI systems have a stable, unified “understanding” of the world. Many people implicitly grant language models systems intentionality (beliefs, desires, goals) because they have human-like language capabilities. But coherent language production is neither necessary nor sufficient for understanding.

How does that misconception affect people in the business world who might be leaning on AI?

I think this leads to people simultaneously over-trusting AI outputs (assuming understanding backs them) and under-investing in alignment and evaluation infrastructure. AI systems are still powerful but fundamentally brittle in ways unlike human cognition. This might change very soon or in a very long time, but what we see going on with rogue AI agents is just disconcerting and pretty embarrassing for the companies behind.

While researching for this book, what surprised you the most about how AI came to be, or what it has been capable of? 

To be very frank, I have never found engineered systems anywhere near as surprising as natural phenomena. ​​They are, in a sense, epistemically transparent. Now, don’t get me wrong, I know that we cannot interpret neural networks (that’s my research focus; explaining AI systems). Still, I think the “surprise ceiling” is bound by our design capabilities and innovative imagination capacity. I think this explains why much of the popular AI discourse can feel somewhat hollow at times. The “no one expected it could write poetry!” reaction treats AI systems as more alien than they are. 

In your first chapter, “A Carefully Selected History,” you talk about various people who contributed to early computing and eventually AI. If you could talk to any one of them, who would it be and what would you ask? 

Easily: Turing. Not because it’s the obvious answer, but because he struggled with the question we’re actually asking today: can machines think? The popular reading of the Turing Test is behaviorist (can the machine fool a human?). But Turing only replaced the actual question (“can it think?”) with an operational proxy he knew was imperfect, as a way of forcing the conversation past metaphysical hand-waving. I think it was a philosophical move disguised as pragmatism. 

So I would ask him: You replaced “does it think?” with “can it imitate?” as a pragmatic maneuver. But we now have systems that pass your imitation game in many contexts while perhaps lacking something grounded in understanding. Did you anticipate that the imitation criterion would be satisfied before the underlying question was resolved? And if so, what test would you propose now, to find out whether machines can think?

When you wrote about the games that computers have won over human opponents, you said “machines defeat us because we ask them to. It’s not the machines, but their creators—researchers like myself—who work hard to build machines that can surpass humans in any domain we can think of.” Just how far do you foresee AI going to surpass humans?

I think the safe thing to do when planning our future is assuming that machines will be able to outperform us at every task. Then, we can get to the important questions of “should they?” and “what will happen—to humans, culture, the environment—if we put machines to these tasks?”

How do you suggest that organizations best avoid bias in their data, training of AI, or execution of machine learning? 
Oh, I have bad news for you. Bias elimination is a category error: bias is irreducible because every dataset is a finite sample from a socially and historically situated process, and every objective encodes value judgments about which errors matter more. The actionable goal is governance, meaning making your inductive choices explicit and auditable. Concretely: treat your data pipeline the way you’d treat a chain of custody in legal proceedings. Documentation, demographic stratification analysis, adversarial evaluation before, during, and after deployment, etc. The organization must make explicit decisions about which criteria reflect their context, document that choice, and defend it as a policy decision rather than hiding behind “the model is objective” or whatever. 

There’s a section of your book called “The Simple Answer Is Often Wrong,” in which an AI model incorrectly identified a husky as a wolf, due to there being snow in the background of most photos of wolves. How can organizations parse AI findings that might be wrong, like the wolf example?

Excellent question. From my biased position, I will say that I think understanding the models better is part of the answer—and also part of the reason I decided to dedicate my career to XAI. You might assume that organizations can test their way to safety. I wish that were so. It would align nicely with engineering practices we’ve spent decades developing, but I think it’s literally impossible. Check the discussion on the curse of dimensionality in the book for details.

Machines That Think: How Artificial Intelligence Works and What It Means for Us is available for purchase from Rheinwerk Publishing/SAP Press. ASUG members can enjoy 15% off any SAP Press titles with the discount code 15ASUG.