← writing / / 6 min read
What Kind of Mind Is Technology Training Us to Have?
A personal reflection on AI and mental health
I’ve been thinking about the relationship between technology and mental health, and I think the discussion is usually simplified too much.
Technology is often described either as something that is destroying our mental health or as something that can democratize access to psychological support.
I don’t think either description is enough.
Technology changes the environment in which our minds operate.
It changes what we see, how often we see it, what gets reinforced, what we compare ourselves with, what we fear, what we remember, and even which hypotheses about the world feel more probable to us.
So maybe the most important question is not simply whether technology is good or bad for mental health.
Maybe the question is:
What kind of mind is this technology repeatedly training me to have?
What kinds of thoughts and habits are we reinforcing through the way we use it?
When I heard AI being compared to electricity, what stood out to me was the possibility of it becoming so deeply integrated into our daily lives that using it would no longer feel like a separate activity. It would become part of how we work, learn, communicate, and interact with the world.
That is why I think we should care not only about what AI helps us produce, but also about how we learn to think while using it.
A more scientific way to use AI
My hypothesis is that a more scientific approach to AI could help us improve the quality of our work and examine the way we interpret our experiences.
By a scientific approach, I don’t mean asking AI for an explanation that sounds scientific.
I mean treating our interpretations as hypotheses, looking for evidence that could challenge them, and checking those interpretations against what actually happens outside the conversation.
There is a difference between asking:
“Why am I right about this?”
And asking:
“What am I assuming here, and what evidence would justify or challenge that assumption?”
The first question asks the system to build a case for a conclusion. The second leaves room for the conclusion to change.
But even a conversation full of thoughtful questions is not enough on its own. The AI’s agreement cannot be the final test of an idea that it helped us construct. Depending on the question, we may need independent sources, direct observations, or an appropriate test in the real world.
An explanation becoming more convincing is not the same as it becoming more accurate.
For me, that distinction is especially important when the subject is not a work project or an abstract argument, but the way we understand ourselves and other people.
When an inference feels like a fact
There was a moment when I believed someone held a certain opinion about me.
When I examined the situation with the help of AI, I realized that I had been treating an inference as a fact. Based on the information I actually had, there wasn’t enough evidence to justify that conclusion.
The AI hadn’t discovered what the other person was thinking. It hadn’t proved that my interpretation was false, either.
It had helped me recognize the gap between what I knew and what I was assuming.
The difference was between:
“This person thinks this about me.”
And:
“I have interpreted the situation this way, but I don’t actually know whether that interpretation is correct.”
I didn’t need to replace a negative certainty with a positive certainty. I needed to recognize that my original certainty was not justified.
That experience gave me a concrete example of what I mean by questioning an interpretation that already feels convincing.
It also made me wonder whether one valuable role for AI in mental health could be helping us examine our interpretations, rather than simply giving us reassurance or another answer to believe.
This is one personal experience, not proof that AI is an effective mental health intervention. But it gives me a specific possibility worth exploring.
Support should not require agreement
The same experience also raises a safety question:
What would have happened if the AI had simply agreed with my original interpretation?
I don’t know. But a system that confidently elaborated on that belief would not have given me an independent reason to trust it. It would have given me more arguments built on the same uncertain premise.
That is the distinction I care about: supporting someone is not necessarily the same as supporting their explanation of what is happening.
I want an AI to be able to acknowledge that an experience is upsetting without automatically confirming that my interpretation is accurate.
At the same time, questioning an interpretation should not mean dismissing it. Sometimes a negative conclusion is supported by evidence. The goal should not be to make every situation sound harmless or positive, but to distinguish what is observed, what is inferred, and what remains unknown.
This is not only a question of how we choose to use AI. It is also a question of how these systems are designed to respond when we are uncertain, distressed, or wrong.
For me, this is where the discussion about mental health meets AI safety: what should a system do when the answer that feels most supportive may not be the answer best supported by the evidence?
What should count as progress?
If I use AI this way, I don’t think the only measure of success should be whether I leave the conversation feeling reassured.
I would also want to ask whether I am becoming better at distinguishing observations from assumptions. Whether I can consider alternative explanations without needing to settle every uncertainty immediately. Whether what I learn in the conversation helps me navigate situations outside it.
Understanding an experience more clearly and feeling better are different outcomes. I would want to examine both, rather than assume that one automatically produces the other.
My goal would not be to consult AI every time I wonder what another person thinks. It would be to practice a way of examining those questions that I could increasingly use on my own.
That is the possibility I find valuable: not just getting better answers from a system, but becoming more capable of evaluating the answers I am willing to believe.
So when I think about technology and mental health, I keep returning to the question that started this reflection.
Not just:
“What can this technology do for me?”
But:
“What am I repeatedly practicing when I use it — and is that helping me become more capable of living the life I want?”
If you want to go deeper into AI safety, and specifically into how these systems should behave in mental-health conversations, the article I produced on this question and the exploratory evaluation behind it are open on GitHub: AgencyBench-MH — github.com/ebarroz/mental-health-safety-eval. It tests whether an assistant can support a person without confirming unsupported beliefs, replacing their relationships, or deferring automatically to whoever holds authority.