The Mirror Doesn’t Accuse: On AI, Value, and the Things We Leave BehindReveals About Us

Introduction

Artificial intelligence models rely on massive datasets, but those datasets are the result of choices about what we consider valuable. It is tempting to see AI outputs as neutral reflections of reality, but in truth, they mirror the priorities and judgments we have already made. When we assess what a model does well, we are really seeing which skills and knowledge we have chosen to formalize and automate.

We rarely ask ourselves which skills and qualities we have chosen to prioritize, and which we have left behind. Judgment, careful attention, and thoughtful debate are often overlooked in the data we gather. As a result, these essential aspects of human work are absent from what AI can reproduce. The technology only reflects what we have decided to capture and value.

**A Premise That Should Be Obvious—and Isn’t **

We need to recognize that AI models are trained on data chosen by people. Selecting what goes into a dataset is an act of curation, shaped by values and priorities. The knowledge and perspectives we include determine what the model learns and what is preserved for the future.

What Gets Included

Most training data comes from sources such as the internet, digitized books, public code repositories, and licensed forums. These sources reflect what people have chosen to document and preserve. Some fields are well represented, while others are missing because their knowledge was never recorded or made digital. In the end, AI models can only learn from what is available, which already reveals our collective decisions about what matters.

What Gets Compressed—and What Gets Lost

AI models can mimic the style of expert writing, but they miss the uncertainty and the process of wrestling with complex ideas. These subtle elements are rarely documented, so they do not appear in the training data. As a result, models reflect only what has been recorded, leaving out the deeper, less visible dimensions of human thinking.

What the Popular Benchmarks Actually Measure

We measure AI models using standardized tests and benchmarks—bar exams, math competitions, code generation tasks. These are domains where we have already set clear criteria and repeatable formats. Models excel here because they are built to fit the structure of the data. In the end, these benchmarks tell us more about our ability to formalize skills than about the full range of human intelligence.

The Harder Reading

We learn the most from the places where AI models struggle or fail. These gaps often point to skills and qualities that are missing from our data—careful reasoning, ethical judgment, the ability to weigh different perspectives. The model’s limits reflect the limits of our data, and ultimately, the choices we have made as a society about what to value and preserve.

The real work ahead is deciding which knowledge and skills we want to preserve and which we want to develop. AI models can only reflect what we give them. Our challenge is to make sure we capture and value the full range of human abilities, especially those that are hardest to formalize or measure.

• Thomas Aldric