Promise & Peril, Revisited: Who Really Benefits from the AI Revolution?
Nearly three years after writing about AI’s weaknesses, I’m less concerned with whether the technology will succeed and more curious about what happens if it does.
I recently revisited an article I wrote in January 2024 examining the weaknesses and uncertainties surrounding artificial intelligence. It’s an interesting exercise, going back to something you wrote and seeing how your thinking holds up after nearly three years of technological change. Some of my concerns have aged reasonably well. Others deserve reconsideration.
What surprised me most wasn’t how much AI had changed, but how much my questions about it had changed.
Back then, I was interested in the technical limitations, reliability, security, and economic promises surrounding AI. Could these systems deliver what their developers envisioned? Were organizations prepared to use them responsibly? Would the investment eventually justify the enthusiasm?
Today, I use AI regularly for research and personal learning. I’ve been impressed by its capabilities, amused by some of its mistakes, and occasionally astonished by how confidently it can explain something that turns out to be completely wrong. I suppose that’s one way to keep a security professional entertained.
I’m more optimistic about AI’s capabilities than I was in 2024. I’m also increasingly concerned about the economic, environmental, and institutional consequences of developing those capabilities. I don’t see a contradiction there. Perhaps we should have expected both developments to advance together.
Technological progress and human progress aren’t necessarily the same thing. They can reinforce one another, but history gives us plenty of reasons not to assume they always will.
An Extraordinary Technology. But Whose Prosperity?
The financial investment in artificial intelligence is extraordinary. Venture capital firms, technology companies, and major enterprises are committing enormous resources to developing models, acquiring computing infrastructure, building data centers, and competing for a position in what they expect will be a transformed economy.
Some of those investments will fail. Others may generate tremendous wealth. But increased productivity and broadly shared prosperity are two different things.
Consider a company that uses AI to increase output while reducing its workforce. Operating costs decline, margins improve, and shareholders benefit. By conventional business measures, that’s a success. But what happens to the employees whose positions disappear? What about the young people who might once have entered the profession through those jobs?
I’ve spent years working in security, where experience and judgment develop through exposure to problems, mistakes, investigations, and occasionally some very long days. You don’t become an experienced security analyst simply by reading the right documentation. You develop expertise by doing the work.
If AI eliminates some of the routine tasks that traditionally helped people develop those skills, what replaces that learning experience? I’m not suggesting that we preserve inefficient work merely to give people something to do. That would be a rather strange approach to progress. But I do think we should consider what happens to the pathways through which people become competent professionals.
There’s another side to this, of course. AI could give individuals and small businesses capabilities that previously required substantial resources or entire teams. As someone building a small consulting business, I find that possibility genuinely exciting. I can explore ideas, research unfamiliar markets, develop technical solutions, and accomplish things that would have required considerably more time or outside assistance just a few years ago.
That’s a meaningful advantage. But here’s where my curiosity takes me next: If AI makes it easier for small businesses to compete while simultaneously making them dependent on a handful of technology providers, who ultimately captures the economic value?
We could be democratizing access to expertise while concentrating ownership of the infrastructure that provides it. Both can be true.
A 2025 International Monetary Fund working paper examined this tension. Its economic modeling suggests AI could reduce some wage inequality while increasing wealth inequality through higher returns to capital. These are modeled outcomes, not predictions we should treat as settled fact, but the distinction between earning a living through work and accumulating wealth through ownership is worth examining.
Perhaps the economic question isn’t simply whether AI will replace jobs. Perhaps it’s also who will own the machines doing the work.
The Physical World Behind the Digital One
There’s something almost magical about interacting with modern AI. You type a question into a screen and receive a sophisticated response seconds later. It feels remarkably effortless, until you remember what’s happening behind the screen: data centers, processors, cooling systems, transmission lines, power generation, and an enormous amount of capital and physical infrastructure.
According to the International Energy Agency’s 2025 Energy and AI report, data centers consumed approximately 415 terawatt-hours of electricity globally in 2024. Its base-case projection puts that figure around 945 terawatt-hours by 2030. That’s more than double in six years. Not all of that demand is attributable to AI, and projections are subject to uncertainty, but AI is a major contributor to the expected growth.
I don’t automatically regard increased energy consumption as a bad thing. Electricity supports economic development, and technological progress has always required infrastructure. AI may even help us develop better energy systems, improve grid management, and discover more efficient technologies.
Still, I have questions. Who pays for the additional electrical infrastructure? How are those costs distributed among technology companies, utilities, businesses, and households? What happens when communities must weigh promised economic development against demands on local resources?
Some communities may benefit considerably from these investments. Others may discover that the costs and benefits aren’t distributed quite as evenly as the promotional materials suggested.
And there’s something amusing about asking an AI assistant to help me investigate the environmental consequences of AI. Somewhere, a collection of processors is consuming electricity to help me contemplate whether those processors are consuming too much electricity. I appreciate the irony. I’m less certain what to do about it.
But perhaps that’s the point. I don’t need to reject the technology to question its costs, nor do I need to ignore those costs to appreciate its benefits. I’d simply like us to examine the entire equation.
Who Is Watching the Watchmen?
My background is in security, so I’m accustomed to thinking about risk, accountability, and what happens when systems (and people) don’t behave as expected. Much of the public discussion about AI safety focuses on hallucinations, bias, deepfakes, and malicious uses of the technology. Those concerns are legitimate, but governance involves more than preventing a model from producing an inappropriate answer.
I’m increasingly interested in the institutions responsible for overseeing these technologies. The companies developing advanced AI possess tremendous technical expertise, financial resources, and access to information about their systems. The people responsible for evaluating and regulating those systems don’t necessarily have comparable resources.
How do we establish meaningful independent oversight under those circumstances? And how do we do it without creating regulatory barriers that inadvertently protect established technology companies from smaller competitors?
I don’t find either extreme particularly convincing. More regulation isn’t automatically better governance. Less regulation doesn’t necessarily produce greater innovation or better outcomes. I’d rather know which risks we’re trying to address, what evidence would demonstrate that a proposed control works, and how we intend to measure unintended consequences.
These are familiar questions in security. I’ve seen enough security programs to know that having a policy and having an effective control are not the same thing. A beautifully documented process can still fail spectacularly when confronted with reality.
The same principle applies to AI governance. We should be interested in whether organizations have adopted responsible AI principles, and whether those principles influence actual decisions, survive commercial pressure, and produce measurable results. And when something goes wrong, can we determine who was responsible?
That’s not a particularly glamorous question, but it’s one I’d like answered before we hand increasingly consequential decisions to automated systems.
Are We Becoming More Capable, or Just More Efficient?
Of all the questions surrounding AI, this may be the one I find most fascinating.
I’ve always enjoyed learning, especially when it involves subjects outside my professional experience. AI has made exploring unfamiliar ideas considerably easier. It can introduce competing perspectives, explain difficult concepts, and help me examine assumptions I might otherwise overlook. One question can lead to another, and suddenly I’ve wandered from security into economics, philosophy, or the history of some obscure invention. I rather enjoy those detours.
Used thoughtfully, AI can be an extraordinary companion for intellectual exploration. But there’s a potential contradiction here: The easier it becomes to obtain an answer, the less incentive we may have to develop the knowledge necessary to evaluate that answer.
Imagine a junior security analyst using AI to investigate an unfamiliar event and produce a very respectable-looking report. Perhaps the conclusions are even correct. But could that analyst explain why they’re correct? Could they recognize when the same system reaches the wrong conclusion tomorrow?
I don’t know that we spend enough time distinguishing between performing a task and understanding it.
Consider education. A student might use AI to produce an excellent essay about a historical event. The writing is polished, the arguments are organized, and the assignment receives a good grade. But did the student develop a better understanding of history? Perhaps. Or perhaps the student simply became better at producing the appearance of understanding.
The distinction isn’t whether AI was involved. It’s whether the person using it exercised and developed their own judgment.
There’s an interesting historical parallel here. Socrates, as portrayed in Plato’s Phaedrus, worried that writing could weaken memory by allowing people to rely on external records rather than their own recollection.
It’s a fascinating concern to revisit more than two thousand years later, particularly since we know about it because someone wrote it down.
Writing didn’t destroy human thought. It transformed how knowledge could be preserved, shared, and developed. Perhaps AI will produce a similarly profound transformation, but that doesn’t mean every concern about its effects on human learning is misplaced.
I suspect the challenge will be learning which cognitive tasks we can comfortably delegate and which we should continue practicing ourselves. After all, a calculator is enormously useful, but I’d still prefer that the person designing a bridge understands mathematics.
We may be entering an era in which access to answers becomes abundant while the ability to evaluate those answers becomes increasingly important, and perhaps increasingly difficult to cultivate. That’s a curious predicament.
What Would Success Actually Look Like?
I’m not pessimistic about artificial intelligence. Quite the opposite, in many respects. The possibilities for scientific discovery, medicine, education, accessibility, and entrepreneurship are remarkable. I’m excited to see what develops over the next several years.
But I don’t think technological achievement alone tells us whether a transformation has been good for society. We can measure model performance, productivity, investment, and corporate valuations. Those are useful measures, but I’d also like to know whether ordinary people are becoming more economically secure, whether smaller businesses are gaining independence or trading one form of dependency for another, and whether communities supporting AI infrastructure are sharing in its benefits.
I’d also like to know whether we’re becoming more capable thinkers or simply more efficient consumers of answers. Perhaps these aren’t the questions that generate the most exciting headlines, but they’re still worth asking.
When I wrote about AI’s weaknesses in 2024, I was largely concerned with the obstacles that might prevent the technology from fulfilling its promise.
Today, I’m increasingly interested in the consequences if it does.
I don’t know exactly what an AI-driven economy will look like or how its benefits and costs will ultimately be distributed. I suspect anyone claiming to know with certainty deserves a few follow-up questions.
What I do know is that technological progress doesn’t relieve us of the responsibility to think about its consequences. We have choices to make about how these technologies are developed, deployed, governed, and incorporated into our lives, and we should be willing to reconsider those choices as we learn more.
What has changed most in your thinking about AI over the past few years?
Comments