Artificial intelligence has reached an inflection point. Just a few years ago, AI was defined by possibility, demonstrating that machines, large language models, algorithms, and software could together process enormous amounts of information, generate insightful outputs, and predictions, and perform tasks once thought to rely largely on human intelligence. Though advanced, they were prone to fabrication, erroneous self-misdirection, reasoning degradation, and outright hallucination. To the extent that their outputs could be "trusted," human verification was essential.
The next phase brought more sophistication: deeper and larger models with increasingly specific reasoning capabilities for deep research and complex tasks, followed by agentic capabilities, better models, and more hooks into our systems and everyday productivity tools. Tasks that once required considerable effort can now be completed with just a few clicks and some basic direction.
Despite all these advancements, these incredibly powerful solutions came with the same caveat: Their results could not always be trusted. “Check my work, I make mistakes” remains stamped at the bottom of every commercial-grade AI interface. A somewhat underwhelming caution, to say the least. The industry was, and remains, captivated by the prospect of increasingly sophisticated, even sentient, AI. All the while, the industry is ignoring the needs of the customer.
As the groundswell of adoption took hold, a feeding frenzy erupted. Astronomical investment dollars and eye-watering valuations dominated the headlines as the rush to ride the AI wave penetrated almost every aspect of our existence. Companies began announcing AI-driven layoffs and large-scale workforce reductions, claiming that AI would deliver massive cost savings.
They rushed to implement prototype enterprise-grade AI across every layer and sector of their businesses, only to discover that their dreams of AI-powered profitability engines could not simply be trusted. Workers began to be called back. Promising implementations turned out to be incredibly time-consuming, requiring unforeseen levels of customization. Human-led trust mechanisms had to be built around these systems to prevent the unthinkable: critical reliance on fallible outputs.
The promise of AI today is still marred by the sneaky hallucination, the errant runaway “thinking process,” and the nagging uncertainty over whether an AI system is actually on the rails. Is the problem serious or merely cosmetic? And if it is serious, how deeply does it affect otherwise compliant operations and their ability to meet established standards and reasonable expectations?
These are not concerns that inspire confidence in a CEO, board of directors, teacher, engineer, small business owner, military commander, or world leader. They are concerns shared by a growing base of end users who have experienced just how difficult it can be to trust AI without human intervention and transparent guardrails.
As organizations rush to adopt AI across industries, we must recognize a fundamental distinction: An AI system producing an answer is not necessarily a system producing reliable intelligence. We also need to ask a more consequential question: What is the cost of being wrong? This is not just economics, but the cost to society, safety, security, and life itself.
Yet the frothy AI gold rush continues. Today’s consumer and enterprise-grade offerings are surrounded by influencers touting their astounding capabilities, with comparatively little discussion about whether those capabilities can be trusted. Perhaps the next phase of AI will be shaped less by promises of what these systems could become and more by the realities of implementing them today. That shift could bring something far more important to the center of the conversation: Trust.
AI needs to be fit for purpose
The rapid commercialization of large-scale AI models has created extraordinary innovation. However, it has also created new questions that must be addressed, particularly in industries where decisions carry significant consequences. How was the model trained? What information influenced its output? What are the boundaries of its knowledge? How does it determine confidence? Can its conclusions be independently evaluated? These are not theoretical questions; they are essential questions that are fundamental to our very safety. In traditional product development, answers to these questions are tested against user requirements and product specifications. Did we decide to throw all of that out the window when it came to AI?
In many applications today, AI systems are increasingly treated as authoritative sources despite limited visibility into how they have arrived at their conclusions. An interface that appears sophisticated can create a perception of certainty, even when the underlying reasoning remains opaque.
A system designed to write a document, summarize information, or generate creative content operates under very different requirements from an AI system supporting decisions in healthcare, drug development, or life sciences research. In these environments, accuracy alone is not enough. The system must provide transparency, context, validation, and a clear understanding of its limitations.
Some of the risks associated with AI in these industries are mitigated by established validation requirements, which mean that predictions and outputs must ultimately be tested in the lab or clinic. But that does not eliminate the risk of error. It makes understanding where, why, and how an AI system can fail all the more important.
The future of AI in highly regulated and lives-at-stake industries will not be built simply by applying larger models to larger datasets. It will require purpose-built approaches that combine AI with scientific methodology. OpenAI’s Sam Altman was recently quoted as saying that one of his biggest fears is that a small number of companies will control all AI. I would take that concern one step further: My biggest fear is an industry in which companies assume a single AI solution can be applied effectively to every problem, industry, and use case.
The right tool for the right problem has always been a fundamental principle of good decision-making. The same should apply to AI. And when it comes to the data that powers these systems, quality should matter at least as much as quantity.
Building AI people can trust
Companies that position their AI solutions as one-size-fits-all answers across industries, with little visibility into how or why those systems reach their decisions, risk losing the trust of the people they were designed to assist. The goal is not to create AI that appears intelligent. The goal is to create AI that can be trusted.
The companies and institutions that successfully navigate this next era will be those that recognize a simple truth: Intelligence without transparency creates uncertainty. Solving that problem does not require simply building a better foundation model. It requires intentional design, including human-in-the-loop strategies that give people the ability to understand, evaluate, and challenge AI outputs.
AI has tremendous potential to accelerate scientific discovery, improve human health, and so much more. But realizing that potential requires moving beyond the question of what AI can do and focusing on what AI can responsibly support.
The future of AI will not be determined solely by the size of a model or the volume of data behind it. It will be determined by whether people can understand, evaluate, and trust the intelligence those systems provide.
That is the standard we should be working to advance.
