Leadership decisions about implementing AI tools are often framed around speed, feasibility or ROI. But those are not the right starting points. What matters first is responsibility, accountability and stewardship of the trust people place in us. AI should not be the default solution to every problem. Applying it without understanding the underlying implications can introduce more risk than value, especially where AI-generated outcomes carry real consequences for people's futures. In those contexts, leadership is not about saying no to AI nor is it about saying yes without hesitation. It is about defining a clear, governed yes that can hold up at scale.

The pressure on every leader right now is to move fast. Ship the model. Launch the agent. Personalize everything. But the public is paying attention, and they are not on the side of speed no matter the cost. A 2025 study by KPMG and the University of Melbourne, surveying more than 48,000 people across 47 countries, found that while AI use is climbing, trust is the central problem and the public wants stronger governance. A December 2025 YouGov survey put it more bluntly: 68% of Americans say they would not let an AI system act without specific human approval.

And in educational assessment, where ETS operates, a single score on a high-stakes test can shape a learner's entire career. In our space, the cost of getting AI wrong isn't measured in user complaints but in opportunity lost.

It would be easy to conclude that the right move is to slow down, that leadership in this moment is about saying no to AI-driven innovation. I disagree. Our job as leaders is to define where AI adds value, set the conditions for using it responsibly and move faster with confidence, not at the expense of trust.

At ETS, our confidence in saying yes to AI comes from the guardrails we put around it. We have been using AI and natural language processing since the early 2000s, beginning with applications like scoring essays and spoken responses. By the end of the year, almost 90% of test items will be generated by our internal AI engine. Yet every one of those items has a human checkpoint we will never remove. And when an AI score and a human score diverge meaningfully, the human wins and drives further review and improvement to the models for the future.

Before any test question reaches a learner, it goes through a fairness and accessibility review. After delivery, we run a psychometric program audit on every major product to look for systematic differences between subgroups. We use an ensemble of AI models, not a single vendor's tool, so that no one system's blind spots become ours. And we treat our proprietary, internationally diverse data, including TOEFL speech samples from test takers around the world, as a check against the narrowness of off-the-shelf models.

ETS literally wrote the book on fair assessment standards, and the arrival of AI doesn't lower the bar. It raises it significantly. Where we used to audit a major program once every three years, we now check more frequently when AI is in the loop. Where we used to assume that trained human reviewers were applying our standards correctly along the way, we now build explicit checks into the middle of the process rather than only at the end. In some areas, we're writing new standards entirely, because the questions AI raises didn't exist a decade ago.

One advantage of working at a company with deep domain expertise is that we bring it to bear from the start. For decades, we have had to rigorously check our work for bias, copyright, fairness, factual accuracy and more. Our AI capability at ETS is built on 77 years of experience and verified intellectual property. Combined with AI models, that foundation powers our new offerings and underpins the trust we share with customers.

More fundamentally, the biggest misconception about AI in business right now is that the goal is to replace people. It really should be about augmenting humans and helping them do more. The real opportunity is to create new value, free people up to solve harder problems and earn the trust to do more over time. That path is slower, and it's the only one that survives contact with reality.

That is the work. The more powerful AI models become, the more important it is to apply them with the discipline, measurement science and human judgment that high-stakes environments demand. When stronger models are paired with stronger standards for fairness, validity and consistency, we do more than move faster. We expand what is possible, deepen trust in the systems people rely on and exponentially increase the positive impact of our mission around the world.