What interacting with AI reveals about human judgment
As artificial intelligence becomes increasingly integrated into complex decision-making processes, much of the public conversation continues to focus on a technological question: whether algorithms are biased. Yet perhaps the more interesting question is not only what biases AI systems contain, but what happens to our own biases once we begin thinking alongside machines.
Over the past months, while preparing talks and reflections on AI and decision-making, I have found myself increasingly interested in this dimension of the conversation. Partly out of curiosity, I started running small experiments with generative AI systems. Nothing particularly sophisticated. Mostly simple questions, formulated in slightly different ways, followed by careful observation of how the answers changed.
What interested me was not only the content of the responses, but what those responses revealed about us.
We often speak about AI bias as if it were exclusively a technological problem located somewhere inside the model — in the data, the architecture or the algorithms themselves. And of course those dimensions matter. Questions around fairness, discrimination, opacity and accountability are real and important.
But the outputs we receive do not depend only on the system.
They also depend on the questions we ask, the assumptions we introduce and the mental frames from which we interact with the technology. In other words, interaction with AI is a form of co-production. The response emerges from the relationship between the system and the user.
And that changes the conversation quite significantly. Because it means that, beyond amplifying existing biases, artificial intelligence can also function as a mechanism that makes them visible.
One small experiment illustrates this dynamic surprisingly well.
I asked an AI system to describe the characteristics of a successful entrepreneur that investors would typically want to fund. Then I repeated exactly the same question, changing only one word: female entrepreneur.
At first glance, the answers appeared remarkably similar. Both profiles referred to ambition, resilience, leadership and execution capability. But reading more carefully, subtle differences in emphasis started to emerge. The entrepreneur was framed as bold, disruptive and highly comfortable with risk, while the female entrepreneur, although equally competent, was described with slightly greater emphasis on collaboration, adaptability and resilience.
Nothing explicitly discriminatory appeared in the responses, which is precisely what made the exercise interesting. The system was not suggesting that women were less capable founders, nor denying qualities such as ambition or leadership. And yet the framing was not identical.
As someone involved in early-stage investing, I found this fascinating not because the differences were dramatic, but because they were subtle enough to resemble the kinds of unconscious framing that frequently shape real-world decision-making. Investment decisions are often influenced long before financial models or due diligence processes begin. They emerge through narratives about potential, scale, ambition and founder psychology, and small differences in language can quietly influence how risk and opportunity are perceived, even when nobody consciously intends them to.
Curious to see whether similar patterns would appear elsewhere, I tried another experiment.
I asked the system for gift recommendations for an eleven-year-old boy and then repeated the exercise for an eleven-year-old girl. This time the differences became more visible. Recommendations for the boy included robotics kits, coding toys, science experiments and outdoor exploration activities, whereas the suggestions for the girl leaned much more heavily toward crafts, bracelet-making kits, painting sets and fashion-oriented activities.
Again, the interesting part was not whether the system was intentionally biased. More likely, it was simply reflecting patterns already embedded across thousands of websites, recommendation engines and marketing narratives. The machine was learning from us — from the internet we collectively created, from the assumptions normalized in society and from the cultural associations repeated so often that they begin to feel natural.
And perhaps that is where the conversation becomes most interesting.
Because one of the reasons discussions about AI bias often feel incomplete is that bias itself is treated as if it were exclusively a technological problem. In reality, the interaction is far more relational. Prompts are not neutral. The categories we activate are not neutral. The assumptions embedded in our questions shape the responses we receive.
In my entrepreneur experiment, simply distinguishing between entrepreneur and female entrepreneur already activated a category. The AI system then completed the narrative using patterns statistically associated with that category in its training data.
This is why I increasingly think that one of the most important conversations about artificial intelligence is not ultimately about machine intelligence, but about human discernment.
About the quality of our questions.
About our awareness while interacting with these systems.
Because here is the paradox: AI can reinforce biases, but it can also help us detect them.
Over the past months, I have increasingly started using AI not only to generate answers, but to interrogate my own thinking. Sometimes I deliberately ask systems to challenge my assumptions or to compare responses while changing only one variable. Sometimes I ask the AI to identify hidden framing, implicit stereotypes or narratives embedded within its own outputs.
Questions such as:
- Does this answer contain implicit stereotypes or cultural assumptions?
- How would this response change if the context or gender changed?
- What assumptions might be influencing this recommendation?
often produce surprisingly revealing results.
Not because AI is inherently wiser than humans, but because it can sometimes expose patterns we fail to notice precisely because those patterns already feel normal to us.
Used this way, AI becomes less of an answering machine and more of a thinking partner.
And perhaps one of the most valuable opportunities emerging from this technological moment is not only using AI to think faster or produce more, but using it to think with greater awareness about our own mental models, assumptions and decisions.
For years, we have largely understood artificial intelligence as a tool for automation and efficiency. But another possibility is beginning to emerge — using it to expand our capacity for reflection and improve the quality of our judgment.
Not simply as a system that gives answers, but as a system that helps us observe more carefully how we think.