Let me start by saying something that might sound contradictory given the argument I am about to make: I am a fan of AI.
I use it regularly. It has become part of my workflow in the same way search engines, smartphones, and cloud computing became essential tools over the past several decades. I use it to brainstorm ideas, accelerate research, write and review code, organize thoughts, and explore solutions to problems that would have taken significantly longer without it. I believe artificial intelligence will fundamentally change the way we work, create, and solve problems. I don’t think it is a fad, and I don’t think we are going to look back five years from now and say, “Remember when everyone thought AI was going to change everything?”
AI will change everything.
The question is not whether the technology is important. The question is whether we are mature enough to responsibly deploy something this powerful.
That distinction has been bothering me more and more.
Recently, I was rebuilding my personal website and, like many people, I used AI to help accelerate the process. The experience was fascinating because it captured both sides of the current AI conversation. On one hand, it was genuinely impressive. AI could generate code in seconds, explain unfamiliar frameworks, suggest design improvements, and help me solve technical problems that would have taken me hours to research on my own.
But there was another side to the experience.
AI could solve individual problems very effectively, but it struggled with the broader context of the project. It could improve a page but lose sight of the overall design direction. It could fix one issue while accidentally creating another. It could produce something that looked correct in isolation but did not always fit with the larger vision of what I was trying to build.
Interestingly, this article itself has been a small example of the challenge I am describing. Using AI to help draft and refine writing is incredibly useful, but there is a difference between generating text and understanding the intent behind the text. At various points, AI could produce something that was technically polished but drifted away from the voice and argument I was trying to make. It had the ability to write a compelling paragraph, but it struggled with maintaining the larger context of the entire piece.
That distinction is important.
A good writer is not simply someone who can produce sentences. A good writer understands why those sentences exist, how they connect to the broader argument, and what the reader should take away from them. AI can accelerate the process, but it still benefits from a human editor who understands the purpose behind the work.
And that is a relatively low-risk example. The consequences of a misplaced paragraph are minor. The consequences become much different when the same limitations appear in healthcare, transportation, finance, or other areas where decisions affect people’s lives.
That experience did not make me less excited about AI. It actually made me more realistic about it.
The technology is incredibly powerful, but power and judgment are not the same thing.
A human designer, developer, or product manager does not simply make decisions based on whether something technically works. They consider the broader context. They understand the audience, the purpose, the tradeoffs, and the consequences. They ask whether a solution makes sense, not just whether it can be implemented.
That distinction is becoming increasingly important as AI moves from helping people complete tasks to making decisions on their behalf.
The more I think about the current AI landscape, the more I am reminded of the dot-com boom of the late 1990s. That comparison is not because I believe AI is another internet bubble waiting to collapse. In fact, I believe the opposite. The internet changed the world, and AI will likely become one of the defining technologies of our generation.
The lesson from the dot-com era was never that the internet was overhyped. The lesson was that technological capability does not automatically translate into sustainable value.
During the late 1990s, investors poured money into almost anything associated with the internet. Companies launched because they had an idea involving the web, not necessarily because they had solved a meaningful customer problem. There were incredible successes, but there were also spectacular failures. Pets.com became the symbol of the era, not because selling products online was a bad idea, but because enthusiasm for the technology overshadowed basic questions about economics, customer value, and long-term viability.
The internet was transformative; every company built around the internet was not.
I keep thinking about that distinction because it feels increasingly relevant to AI. The technology itself is real, and important, and likely to reshape a lot of what we do. But the fact that a technology matters does not mean every company using it is doing something meaningful with it. I was reminded of that recently at a trade show where AI seemed to be everywhere. Every booth had a story. Every demo had a chatbot. Every pitch deck had the same familiar glow of inevitability. After a while, the pattern became hard to ignore.
Some companies were using the word AI to describe things that were not really AI at all, at least not in the way the term is usually understood by researchers and practitioners. They were talking about automation, recommendation engines, routing logic, predictive models, and rules-based systems that have been around for years. None of that is trivial. In fact, a lot of it is genuinely useful and in some cases deeply sophisticated. A routing algorithm that saves a fleet hours of driving time is valuable. A rules engine that processes transactions accurately and at scale is valuable. Predictive analytics that surface patterns humans would miss are valuable. But when every piece of software that makes a decision faster than a person gets folded into the AI bucket, the term starts to lose meaning. Customers stop being able to tell what is actually happening under the hood, what problem is being solved, and whether there is any real innovation behind the label.
Then there were the companies that did have AI capabilities, but the capabilities felt bolted on rather than thought through. Chat was the most obvious and glaring example. Suddenly everything had an AI chatbot. Sometimes that made sense. A conversational interface can be a genuinely better way to help someone find information or complete a task. But just as often it felt like the company had started with the technology and then worked backward to invent a reason for it to exist. This is, unfortunately, a trend that repeats itself often, and we have seen versions of it before. There was a stretch when every company needed a mobile app, even if the app simply recreated a clumsy experience on a smaller screen. There was another stretch when companies bought analytics platforms before they had decided what decisions they actually wanted to improve. Technology became the strategy instead of the tool. We are currently making the same mistake with AI. The question should not be where can we add AI. The question should be what problem does AI solve better, faster or cheaper than the alternatives.
Technology became the strategy instead of the tool. We are currently making the same mistake with AI. The question should not be where can we add AI. The question should be what problem does AI solve better, faster or cheaper than the alternatives.
The companies that actually caught my attention were the ones where AI was not the headline but the mechanism that progressed the product forward or solved a problem. I would leave those booths thinking, quietly, that something interesting was happening there. Not because they had the flashiest demo or the most fashionable language, but because the technology was doing work that had previously been too difficult, too expensive, or too slow to do well. In those cases, AI was not the product. The value created by AI was the product. That distinction matters more than people seem willing to admit, because it is what separates a durable business from a temporary wave-rider.
That distinction matters even more now that AI is moving into more serious territory. For a lot of people, AI first showed up as a productivity tool. It writes emails, summarizes meetings, generates images, answers questions, and helps with repetitive administrative work. Those are useful applications, but they are also relatively low stakes. If the model writes a mediocre social post or produces a strange image, the consequences are mostly annoyance. But AI is no longer staying in that lane. It is moving into places where context matters and mistakes have consequences. Healthcare organizations are using it to assist with diagnosis and patient care. Financial institutions are exploring it for lending, fraud detection, and investment recommendations. Schools are experimenting with tutoring and personalized learning. Manufacturers are using it for quality control and predictive maintenance. Governments are looking at it for public services and infrastructure. The conversation changes when the cost of failure is no longer measured in inconvenience.
One of the reasons autonomous driving has become such an interesting case study for AI is that it exposes the difference between capability and judgment. Driving seems like a perfect problem for automation. Humans make mistakes. Humans get distracted. Humans get tired. A machine that can process thousands of data points every second should, in theory, be safer.
And yet the hardest problems in autonomous driving have not necessarily been the obvious ones. The challenge has been the edge cases. The situations that require context, judgment, and an understanding of the world beyond the immediate data in front of the system.
The industry has made incredible progress, but it has also learned some difficult lessons. In 2018, an autonomous Uber test vehicle struck and killed a pedestrian in Tempe, Arizona. Investigators found that the system detected the pedestrian but failed to properly classify and respond to the situation. In 2026 a passenger was stuck in a Waymo that was driving through a construction zone while being trailed by San Francisco police. The incident was a reminder that building a system capable of navigating thousands of normal scenarios does not necessarily mean it understands the rare scenarios where judgment matters most.
That is the challenge with many AI systems. They can be extraordinarily capable within the environments they have been trained for, but the real world is full of exceptions. A human driver does not simply recognize objects. They understand context. They know that a person standing near a construction zone may behave unpredictably. They understand that a stopped emergency vehicle, unusual road markings, or a person waving their arms may require interpretation beyond what is immediately visible.
The goal of autonomous driving should not be abandoned because of these challenges. Quite the opposite. The technology has enormous potential. But it is an example of why deploying AI in the real world requires more than proving that something works most of the time. It requires understanding how it fails, when it fails, and what safeguards exist when it does.
An AI that helps someone organize notes is not the same thing as an AI that helps interpret a scan. An AI that drafts marketing copy is not the same thing as an AI that evaluates creditworthiness. An AI that schedules a meeting is not the same thing as an autonomous system that can move money, alter records, or make decisions that affect someone’s livelihood. Context matters, and too much of the current conversation treats AI as if it were one single category of technology with one single acceptable level of risk. It is not. The amount of testing, oversight, transparency, and regulation should depend on what the system is doing and what happens when it fails.
That is why the question we should be asking is not simply whether we can build something. It is whether we should.
That question has been strangely absent from a lot of the AI conversation, as if asking it makes you anti-innovation. It does not. In many cases the answer will absolutely be yes. But asking the question is part of being responsible, and responsibility matters more when the technology is powerful enough to be useful in places where people are tempted to trust it too quickly. We have already seen what happens when humans stop applying judgment and assume the machine has done the thinking for them. In 2023, attorneys were sanctioned after submitting a legal brief that included citations generated by AI to court cases that did not exist. The model produced something that looked convincing, and the people using it failed to verify it. That incident should not be read as proof that AI is useless. It should be read as proof that expertise and accountability do not disappear just because a new tool enters the workflow.
A calculator does not eliminate the need to understand mathematics. A spreadsheet does not eliminate the need to understand financial analysis. AI does not eliminate the need for human judgment. If anything, the more powerful the tool becomes, the more important judgment becomes. That is why healthy skepticism is not a barrier to progress. It is part of how serious technologies mature. Somewhere along the way we started treating questions about AI adoption as if they were attacks on innovation. They are not. Some of the most important advances in history happened because people were willing to challenge assumptions and demand evidence. Aviation became safer because engineers investigated failures and changed procedures. Medicine advanced because researchers insisted on proof before accepting new treatments. Cybersecurity evolved because professionals stopped assuming systems would be secure by default and started designing for the reality that they would be attacked. Those fields did not mature because people stopped believing in the technology. They matured because people took the technology seriously.
That is what I mean when I say we need adults in the room.
We need people who understand that building something impressive is not the same as building something responsible. We need engineers thinking about reliability and security. We need domain experts who understand the consequences of failure. We need executives who can look past the excitement of the moment and ask whether the thing being built is actually useful. We need product leaders who keep asking whether we are solving a real problem or just adding technology because we can. And yes, we need thoughtful regulation, not because technology is scary and not because innovation should be slowed for its own sake, but because once a technology becomes important enough, society needs standards, accountability, and protections. Aviation has safety standards. Financial markets have oversight. Healthcare has approval processes. AI should not be exempt simply because it is new and exciting.
I remain optimistic about AI because I believe in what it can become. I believe it can help doctors identify disease earlier, help teachers give students more personalized support, accelerate scientific discovery, and remove a huge amount of repetitive work from people’s lives. But optimism is not the same thing as blind acceptance. The future of AI will not be determined only by what we are capable of building. It will be determined by whether we have the judgment to decide what should be built, where it should be deployed, and what safeguards need to exist around it.
AI is powerful.
That is exactly why we need to be careful with it.
The next phase of AI should not be defined only by asking what machines can do. It should be defined by asking what humans should allow them to do.