In a study conducted by researchers from Stanford and the University of Washington, 70 different major large language models (LLMs) were asked various open-ended questions from philosophy, business, and literature. Each of these models varied in structure and training data, and was developed by different corporations. Surprisingly, however, most of these models produced similar, if not identical, responses. It seemed as if the models were approaching a common asymptote of thought. As Andover makes efforts to incorporate AI models like Khanmigo in the learning experience, we must be aware of these weaknesses of AI and identify what areas we can let them contribute to our learning.
LLMs are a type of AI that is most familiar to the public; ChatGPT, Gemini, Claude, and Grok are all examples of LLMs. These LLMs are inspired by the human brain and the state of connectivity of biological neurons. After this structure is developed, it is trained on data, similar to how young humans learn through textbooks or experiences. Unlike humans, however, LLMs can learn much more and faster due to their superior computing speed and amount of data. The key here is the sheer volume of data. The more a model learns, the more it adjusts towards giving the most probable response.
It’s also possible to interpret this inversely. Just like how the massive amount of data leads to homogenization and loss of uniqueness, it stands to reason that uniqueness and diversity come from uneven and partial data. Think of it like this: there are billions of dots on a scatter plot. How will the estimated mean differ based on how many sample points are given? Statistically, more points given lead to less noise, leading to a more accurately predicted mean. However, fewer points given translates to noise having more voice in the mean, which produces a more variable mean. We, as humans, are that low-sample graph. Each of us is shaped by a singular, irreproducible set of experiences. That limitation is not a flaw. It can be the source of what is original in us.
This matter of LLMs is becoming increasingly important as Andover is beginning to incorporate AI into learning. Khanmigo, an LLM developed by Khan Academy, is being introduced into selected classes starting this spring. Khanmigo, much like other learning tools, has much to offer. It can provide personalized help that may not be accessible due to larger class sizes or a lack of time. It is also efficient at providing fast, real-time responses without having to email or schedule a meeting with an instructor.
Because of these clear pros and cons of AI-integrated learning, the crucial question to ask is not whether to use or ban AI, but when to use it. LLMs have their clear disadvantages, including a tendency towards uniformity. Khanmigo, though a learning tool, is nevertheless an LLM and may display similar trends seen in the study mentioned earlier. Furthermore, learning often doesn’t come from simple answers to questions. When I think back to moments that spurred growth and realization, it was through discussions and the sharing of diverse views that allowed me to see the full picture. These methods are obviously not as efficient as AI tools can be — discussions take time and patience, and may not be as accessible as a website. On the other hand, Khanmigo and other AIs can be outstanding in their ability to help clarify concepts and guide practice. But for responding to questions with no determined answers that require discussions, the diversity of perspectives matters more than efficiency.
Thus, to adapt and to adopt AI in our lives as students is not a simple yes or no question. It is to recognize AI’s strengths and weaknesses, and use them accordingly. Ultimately, the goal of education at Andover is not to reach the most “likely” or common answer, but to find and cultivate the unique “noise” that only you can produce. If we rely on LLMs to navigate the gray areas of learning, we risk converging our thinking into a single, predictable asymptote. By preserving the productive inefficiency of human discussion, we ensure that our growth stays human and beautifully unpredictable.
