The Airplane Doesn’t Flap Its Wings: Maharaj Mukherjee on Redefining Intelligence for the Age of AI

Q1. You and Dr. Monideepa Roy argue for a mechanism-agnostic definition of intelligence — one based not on how a system is built, but on what it demonstrably does: perceive, predict, learn, adapt, collaborate, and act under uncertainty. That is a deliberately provocative stance in a field where debates about symbolic reasoning versus neural networks versus emerging quantum paradigms often get treated as debates about which architecture is “truly” intelligent. What convinced you that the question of mechanism should be set aside entirely, and what is the strongest objection to this framing that you had to work through before you were satisfied it holds up?

First, thank you for asking this question, because it really goes to the heart of one of the central premises of our book.

I would make one distinction at the outset. We don’t argue that mechanism is unimportant. We argue that mechanisms should not be the gatekeeper for deciding whether something is intelligent. We actually spend considerable time exploring mechanisms. But we distinguish between how intelligence is implemented and what intelligence demonstrably does.

What convinced us was actually a very simple observation: nature itself does not seem committed to one architecture for intelligence. Human intelligence is largely centralized in the brain. But an octopus has a remarkably distributed nervous system, with much of its neural processing occurring in its arms. Bees and ants present yet another model, where relatively simple individuals can collectively produce sophisticated behavior. Nature has found very different ways of producing what we recognize as intelligent behavior. 

So, if nature can produce intelligent behavior through such radically different mechanisms, why should we insist that an artificial system must use one particular mechanism before we are willing to call its behavior intelligent?

I like the analogy of flight. Birds fly by flapping their wings; airplanes don’t. If we defined flight by the mechanism—flapping wings—we would have to conclude that an airplane doesn’t really fly. Instead, we recognize flight by what it accomplishes. The mechanism remains enormously important for understanding how it flies, but not for determining whether it flies.

We take a similar approach to intelligence. Rather than getting trapped in the centuries-old question “What is intelligence?”, we ask: What does intelligence do? Can a system perceive, learn, predict, adapt, act purposefully under uncertainty, and learn from or collaborate with others? These give us observable capabilities that we can investigate and ultimately try to build. 

There is also a deeper difficulty: defining intelligence is somewhat self-referential—we are using human intelligence to define human intelligence. We should therefore be cautious about assuming that because human intelligence is the only intelligence we understand from the inside, its mechanism must be the blueprint for every possible intelligence. 

The strongest objection is: are we confusing intelligent behavior with genuine understanding? Searle’s critique is important here. A machine might produce appropriate responses without possessing meaning, consciousness, or understanding in the human sense. We don’t dismiss that objection. But we think it addresses a somewhat different question. Whether a machine is conscious and whether it can demonstrate intelligent capabilities are not necessarily the same question. 

So, I would summarize our position this way: we are mechanism-agnostic, but not mechanism-indifferent. Mechanisms matter enormously because we want to understand how intelligence works and how to build it. What we resist is the assumption that biological, symbolic, neural, quantum—or some architecture we haven’t yet invented—has exclusive rights to the word intelligence.

Intelligence may be less like a particular machine and more like a capability that very different kinds of machines can realize. And that is one of the starting premises of AI That Thinks.

Q2. Your career spans holding close to 300 US patents and over 150 international patents as an IBM Master Inventor, and now working in banking as Vice President at Bank of America. Those are two very different institutional contexts for thinking about AI — one built around long-horizon research and IP creation, the other built around regulated, risk-averse, real-time financial decision-making. How has moving from IBM Research to the banking sector changed what you believe intelligent systems actually need to demonstrate before they can be trusted with consequential real-world decisions, and did that experience shape any specific argument in the book?

I might frame my experience a little differently. I don’t necessarily see IBM Research and banking as representing a contrast between innovation on one side and risk aversion on the other. I have been extraordinarily fortunate throughout my career to work with some of the sharpest minds in the world. Whatever I have accomplished, whether through research, patents, or applying technology to real-world problems—has come from collaborating with exceptional people. I genuinely feel that I have been standing on the shoulders of giants.

What these different experiences have reinforced for me is the importance of understanding and managing risk. I would distinguish that from being risk averse. Innovation inherently involves risk. A startup must manage risk to survive, and a large corporation must manage risk at scale. We manage risk in our personal lives as well—even maintaining a happy family involves anticipating consequences, balancing competing needs, and making decisions under uncertainty.

That has strongly influenced how I think about intelligent systems. Intelligence is not simply the ability to make a prediction or arrive at the right answer. It is also the ability to recognize uncertainty, learn from feedback, adapt when circumstances change, and understand that actions have consequences. The more consequential the decision, the more important these qualities become.

That idea is a recurring theme throughout AI That Thinks. We repeatedly look at intelligence as a dynamic process: perceive the environment, learn from it, make a prediction or decision, act, observe what happens, and then adapt. Whether we are discussing biological intelligence, computer vision, learning systems, robotics, Edge AI, or more autonomous and collaborative forms of AI, intelligence emerges not from getting everything right the first time, but from operating effectively in an uncertain and changing world.

And I think that leads directly to the question of trust. We should not trust an AI system merely because it is impressive, or even because it is usually right. For consequential decisions, we also need to understand how it behaves when information is incomplete, when conditions change, and when it makes a mistake.

So one lesson that my career has reinforced, and that runs throughout the book, is this: trustworthy intelligence is not intelligence without risk. It is intelligence that can operate responsibly in the presence of risk and uncertainty.

Q3. The book’s central thesis is that the future of AI belongs not to a single paradigm but to hybrid, embodied, distributed systems that integrate learning with reasoning and autonomy with accountability. “Autonomy with accountability” is easy to state as a principle and notoriously difficult to engineer in practice — particularly once you introduce agentic AI systems that act with real independence. Drawing on your own decades of experience taking research prototypes all the way to productization, what is the single hardest engineering trade-off between autonomy and accountability that you have encountered, and how does the book propose resolving it?

I think the hardest trade-off is actually one of speed and control. We want autonomous systems because they can perceive, decide, and act at a speed and scale that humans simply cannot match. But the more autonomy we give them, the less realistic it becomes to assume that a human being can examine and approve every decision.

This becomes particularly important with agentic AI. Imagine millions of agents interacting, negotiating, making decisions, and initiating actions at machine speed. We cannot put a human checkpoint in front of every action. By the time a human intervenes, thousands—or millions—of subsequent actions may already have occurred.

So my view is that accountability cannot be an afterthought. It has to be built into the framework and architecture itself. An autonomous system must have mechanisms to verify its actions against its objectives, constraints, permissions, and observed consequences. When appropriate, it should be able to stop, reconsider, escalate, or correct itself rather than blindly continue.

There is a useful analogy in the biological systems we discuss in AI That Thinks. When you catch a ball, your brain does not issue a command to your hand and simply hope it works. It continuously compares the intended movement with sensory feedback and corrects the motion in real time. In other words, action and verification are part of the same loop. I believe accountable autonomous AI needs to work on a similar principle.

This also connects with the book’s broader discussion of distributed intelligence. Intelligence does not necessarily reside in one central controller; cognition can be distributed across components, people, tools, and environments. As AI becomes increasingly distributed, governance may also need to become distributed and operate at machine speed.

But there is an important distinction here: self-governance is not the absence of human governance. Humans must establish the objectives, boundaries, permissions, ethical and regulatory constraints, and conditions under which the system must stop or escalate. What the AI needs to do autonomously is continuously enforce and verify those constraints at a speed at which human supervision alone is no longer practical.

And self-verification does not mean self-reference. A system simply saying, “I checked myself and I am correct,” is not accountability. Effective self-verification requires independent checks—different mechanisms, constraints, or agents capable of challenging an action rather than simply confirming it. The book itself discusses why self-reference presents fundamental difficulties. 

So I don’t think the solution is choosing between autonomy and accountability. The engineering challenge is to make accountability part of autonomy itself. Humans define the rules and remain ultimately accountable, but verification and enforcement increasingly have to occur within the architecture and at the speed at which the intelligent system operates.

Q4. Your book moves from human cognition, through computer vision and robotics, to agentic AI, edge AI, and finally quantum AI — a genuinely wide arc. Given your own background spans machine learning, blockchain, and IoT rather than quantum computing specifically, what did you learn in researching and writing the quantum AI chapter that most changed how you think about the near-term versus long-term promise of the field — and where do you believe the current hype around quantum AI is most likely to disappoint practitioners in the next five years?

That is a great question, because Quantum AI was probably the chapter for which I personally had to learn the most before I could write it. My background is not in quantum mechanics or quantum computing, and I still would not call myself an expert in either. I had to go back and start with the fundamentals of quantum mechanics before I could even think seriously about Quantum AI.

Interestingly, my biggest struggle was not understanding the mathematics—it was figuring out how to explain the basic premise to someone who has never studied quantum mechanics. I was not satisfied with many of the popular analogies that describe electrons as little balls or their behavior through familiar physical or planetary metaphors. The quantum world is fundamentally unlike anything we experience directly with our senses. 

Eventually I realized that there is something even a child understands intuitively: probability. A child playing hide-and-seek understands that before opening a closet, her friend might be there, or behind the curtain, or under the bed. So the chapter uses probability and hide-and-seek as an intuitive bridge to superposition, uncertainty, and eventually observation. 

But what changed my own thinking most was going back to the original promise of quantum computing. I describe the idealized notion of an oracle: imagine a machine that could somehow consider an enormous space of possibilities and give us the answer to a computationally hard problem. That is an extraordinarily seductive idea, particularly for AI, because so many interesting problems involve huge spaces of possibilities. 

Then comes the reality check that I found most important. Having all the possibilities represented in a quantum state is not the same as being able to read all the answers. When we observe the system, measurement gives us an outcome; quantum algorithms therefore have to use interference very cleverly so that the desired answer has a high probability of being the one we observe. The quantum computer is not the magical oracle we might initially imagine. 

And that is where I think some of the Quantum AI hype will disappoint people over the next five years. The expectation that quantum computers will simply take today’s AI algorithms, make them exponentially faster, and suddenly solve arbitrary NP-hard problems is, in my view, the wrong expectation. The chapter traces how that early optimism has already become much more pragmatic: quantum advantage is likely to be highly problem-specific, while today’s systems still face noise, decoherence, error correction, data encoding, measurement, and scaling challenges. 

Writing this chapter actually made me simultaneously more skeptical about the near term and more excited about the long term. In the near term, I expect hybrid classical-quantum approaches and specialized applications to be much more realistic than some revolutionary replacement of classical AI. But in the long term, if we learn to exploit superposition, interference, and entanglement effectively, Quantum AI could open genuinely new ways of searching, optimizing, and reasoning over extraordinarily complex spaces. That possibility is what makes the field fascinating to me.

Q5. As an IBM Master Inventor for Life with close to 300 US patents, you have spent decades identifying genuinely novel ideas worth protecting — a discipline that requires distinguishing real innovation from incremental variation or hype. Applying that same inventor’s discipline to the current AI landscape, what is the one claim being made today about “thinking machines” or artificial general intelligence that you believe would not survive the kind of rigorous scrutiny a patent examiner or a serious inventor would apply to it?

That is an interesting question because an inventor learns to be skeptical of very broad claims. In a patent, it is not enough to say that something should work. You have to explain what is genuinely novel, how the pieces work together, and why the claim is supported.

If I applied that discipline to some of today’s discussion about AGI, the claim I would scrutinize most closely is the idea that if we simply keep making today’s AI models larger, give them more data and more computing power, general intelligence will somehow emerge.

I don’t say in AI That Thinks that AGI is inevitable—or even that it is possible. What we ask instead is: if something resembling AGI is possible, what might it actually require? And the journey through the book leads us away from the idea of one gigantic, all-knowing machine.

Nature gives us an interesting clue. Intelligence itself is remarkably distributed. Even human cognition extends beyond an isolated brain through our bodies, other people, tools, and our environment. Social insects demonstrate another form of intelligence in which relatively simple individuals collectively produce sophisticated behavior. So perhaps the assumption that AGI must reside inside one enormous model deserves much more scrutiny.

My conjecture is that, if AGI ever emerges, it may look less like one giant brain and more like an ecosystem of intelligence—distributed across specialized agents, edge devices, sensors, robots, and other computing resources that perceive, learn, reason, predict, collaborate, and act together. That is consistent with the distributed and embodied view of cognition developed throughout the book.

And farther into the future, some of those edge systems might themselves incorporate quantum computing or Quantum AI. In Chapter 8, we explore the possibility of edge, agentic, and swarm AI eventually converging with quantum capabilities into what we call “distributed quantum cognition.” But we deliberately present that as a future possibility, not as an established path to AGI. 

If I put on my inventor’s hat, my question about today’s AGI claims would be very simple: Where is the architecture? Showing that a model can converse, reason about certain problems, generate software, or use tools is impressive—but those individual capabilities do not by themselves demonstrate general intelligence.

A collection of impressive capabilities is not yet a theory of how general intelligence emerges. That, to me, is where the real invention still has to happen.

Q6. You have mentored new inventors throughout your career as an IBM Master Inventor, helping others learn to see genuine innovation where it exists. Now, through this book, you and Dr. Roy are trying to help a much broader audience — engineers, researchers, business leaders, students — see past the current AI hype to understand what building genuinely intelligent systems actually requires. What is the most important thing you learned about how to teach that kind of discernment, either from your years mentoring inventors at IBM or from the process of co-writing this book, and did anything about writing it for a broader audience change how you personally think about intelligence?

The most important thing I learned from mentoring inventors is that innovation should be fun. Creativity should be fun. If someone begins the process of invention by asking, “What reward will I get from this—money, recognition, a patent, a promotion?” they may ultimately be disappointed. Those things may come, but they cannot be the primary motivation.

The real reward is the experience of discovering something that wasn’t there before—or solving a problem that nobody around you knew how to solve. To me, that feeling is a little like discovering a new continent or being the first person to visit a new planet. You suddenly see something that, a moment earlier, you didn’t know existed. That sense of discovery has kept invention exciting for me throughout my career.

That is also what I tried to teach new inventors: don’t start by looking for an invention; start by becoming curious about a problem. Ask why something works the way it does. Ask why it cannot work differently. Question assumptions that everybody else has stopped questioning. Genuine innovation often begins with curiosity rather than with the intention to be innovative.

Co-writing AI That Thinks reinforced that lesson for me. We deliberately did not want the book simply to tell readers what AI is. We wanted readers to travel with us through the questions—what intelligence means, how humans and other organisms learn, how perception and purposeful movement contribute to intelligence, how intelligence can be embodied and distributed, and how those ideas might inform artificial systems. The book itself describes this as assembling smaller “puzzle pieces” rather than trying to answer the entire question of intelligence at once.

Writing for a broader audience also changed my own thinking. When you have to explain an idea without hiding behind mathematics or technical terminology, you discover very quickly whether you really understand it. I experienced that particularly strongly while writing about quantum computing, but it happened throughout the book.

And perhaps that led me to a broader realization about intelligence itself: intelligence may begin with curiosity, the willingness to explore something you don’t yet understand, learn from what you find, and change your understanding as a result.

That is also what invention is. And if readers finish AI That Thinks with more questions than when they started—but with better questions—I would consider that a success.

Resources

AI That Thinks: From algorithms to autonomy and building machines that truly think (English Edition) by Dr. Maharaj Mukherjee (Author), Dr. Monideepa Roy (Author).  BPB Publications, June 20, 2026

…………………………………………………..

Maharaj Mukherjee, Senior Vice President and Senior AI Architect at Bank of America.

Maharaj Mukherjee is a distinguished technology leader, researcher, and prolific innovator and author of the top selling book “AI That Thinks”, with over 25 years of experience spanning academia, industrial research, and enterprise-scale artificial intelligence. He currently serves as Senior Vice President and Senior AI Architect at Bank of America, where he leads strategic initiatives in AI, machine learning, digital twins, and agentic AI systems, driving innovation in enterprise-scale financial technologies.

Prior to this, Dr. Mukherjee spent two decades at IBM Research, where he made foundational contributions to computational geometry, electronic design automation, and AI-driven systems. His work in model-based optical proximity correction for semiconductor lithography enabled high-value technology transfers to global industry leaders. He also led several early applied AI initiatives using IBM Watson, including robotics-based workspace mapping and predictive analytics systems.

Dr. Mukherjee began his professional journey in academia as an Associate Professor at the Indian Institute of Technology (IIT) Kharagpur, where he taught core computer science subjects and led industry-sponsored research in computer vision for manufacturing.

An IBM Master Inventor Emeritus and inductee into the IBM Inventor Hall of Fame, Dr. Mukherjee holds over 290 U.S. patents and more than 160 international patents, making him one of the most prolific inventors in his field. He has been consistently recognized as a top inventor at Bank of America for multiple consecutive years.

His contributions have earned him numerous prestigious honors, including the IEEE Region 1 Industrial Innovation Award (2025), the MachineCon Top 100 Influential Leader Award (2025), and the Bank of America Technology Excellence Award (2023). Earlier in his career, he received the IBM “20 Patents for 20 Years” Award, the Young Scientist Award from the Government of India, and the IEEE Richard E. Merwin Scholarship, among other recognitions.

Dr. Mukherjee earned his Ph.D. in Computer and Systems Engineering from Rensselaer Polytechnic Institute, along with an M.S. from SUNY Stony Brook and a B.Tech (Hons.) from IIT Kharagpur. He remains deeply engaged with the global engineering community through leadership roles within IEEE, including serving as Chair of IEEE Region 1 PACE and contributing to multiple national and regional awards committees