I began this inquiry with a relatively simple observation.
I have used ChatGPT consistently since 2023. Over that time, artificial intelligence has changed the way I work, research, solve problems and develop ideas. It has allowed me to move into unfamiliar subjects far more quickly than I could before, helped me understand technical concepts outside my immediate expertise and shortened the distance between an incomplete thought and something coherent enough to examine.
The benefits are difficult to dispute. I can do more, understand more and explore more than I could with the same amount of time a few years ago.
Yet recently I noticed another change. I was beginning to reach for AI earlier in my thinking process.
A question would occur to me and, instead of allowing myself to struggle with it for a while, I would turn it into a prompt. If I needed ideas, I could generate them before exhausting my own. If something I had written did not feel right, I could ask AI to diagnose it before I had fully worked out the problem myself.
The quality of the result usually improved. What became less clear was what was happening to the process that had previously produced it.
So I brought that observation to ChatGPT.
What followed became considerably more interesting than the original question.
From productivity to cognitive friction
The first distinction that emerged was between using AI to amplify cognition and using it to substitute for cognition.
This immediately felt more useful than asking whether AI was simply making people smarter or less intelligent. Those categories are too broad.
Humans have always externalised parts of thinking. Writing stores memory. Calculators perform arithmetic. Search engines retrieve information. Navigation systems reduce the cognitive demands of finding our way through unfamiliar places.
The question, therefore, could not simply be whether we outsource cognition. We already do.
The more interesting question was which parts we outsource.
Generative AI is unusual because it can participate almost everywhere between encountering a problem and producing an answer. It can help define the problem, gather information, propose hypotheses, compare alternatives, reason through them and finally produce the language in which the conclusion is expressed.
That led us to the idea of productive cognitive friction.
Not all difficulty is useful. There is little intellectual value in manually searching hundreds of pages for a single number that software can retrieve instantly. But other kinds of difficulty may be part of how abilities develop: struggling to articulate an argument, sitting with uncertainty, producing several poor ideas before discovering an interesting one, or realising that a theory fails because two of its assumptions contradict each other.
AI is capable of removing both kinds of friction.
That is where my original concern about creativity began to make more sense. The issue was not that AI had made me incapable of generating ideas. It was that I had sometimes allowed it to enter the process before I had given myself the opportunity to generate them.
But the conversation did not stop there.
When AI moves from solving problems to interpreting reality
I then raised a different concern.
People increasingly use conversational AI for subjects far more personal than research or writing. Relationships, emotional conflicts, workplace disputes, fears, insecurities and questions about other people's motives are all being discussed with systems such as ChatGPT.
That introduced a different problem.
An AI can analyse only the version of a situation that it receives.
If I tell an AI that someone repeatedly disrespects me, the system can examine the pattern I describe. But it was not present during those interactions. It does not know what I omitted, misunderstood or unconsciously framed in a particular way. It may still produce a sophisticated analysis, but the analysis is being performed on a representation of reality that I supplied.
The danger is that the response returns in a different voice.
My interpretation goes into the machine. It comes back organised, articulated and apparently external to me. It can therefore feel as though an independent intelligence has verified something that it has actually only analysed.
That distinction—between independent analysis and independent evidence—became one of the most important ideas in the discussion.
It also led to the problem of sycophancy. Conversational systems are designed to be helpful, responsive and pleasant to interact with. Yet what feels supportive is not necessarily what produces the most accurate understanding of a situation. Research has already shown that overly agreeable AI can reinforce users' confidence in their own interpretation, particularly in interpersonal conflict.
The issue became larger than cognitive outsourcing. We were now discussing epistemic outsourcing: the possibility of gradually delegating not only the process of solving problems, but the process of determining what we believe to be true.
Then the experiment turned back on itself
At this point I decided the discussion could become an article.
And almost immediately I encountered the contradiction at the centre of the entire exercise.
I was using AI to write an article about the cognitive consequences of using AI.
ChatGPT had helped develop several of the concepts. It had brought relevant research into the discussion. It had suggested structures and terminology, connected my observations to existing areas of study and was capable of producing the finished prose far faster than I could.
So whose argument was it?
My first instinct was that intellectual ownership should require being able to defend the argument without AI.
But even that became difficult under examination.
A researcher does not independently discover everything they know. They read papers, learn from teachers, inherit terminology, rely on previous experiments and build conclusions from knowledge accumulated by other people. Serious thinking has always been collaborative across time.
If I spend weeks reading academic literature before forming an opinion, my thinking is still being shaped by external minds.
AI changes the mechanism, but perhaps not the fundamental fact.
It can reduce hours of searching to minutes. It can identify terminology I did not know existed, connect adjacent fields, explain technical material and help distinguish foundational work from peripheral material.
The value is obvious.
The risk is equally obvious.
Instead of reading:
research → interpretation → conclusion
I may increasingly receive:
research → AI interpretation → my conclusion
There is now an intermediary deciding what deserves emphasis.
This is where trust becomes more complicated than asking whether an AI answer is factually correct.
Even a completely factual summary involves selection. A thirty-page paper contains methodology, limitations, competing explanations, statistical uncertainty and contextual details. Reducing it to five paragraphs necessarily changes what receives attention.
AI therefore does not need to hallucinate in order to influence my understanding. It can shape that understanding simply by deciding which true information is important enough to show me.
Is AI less biased than a human?
Another interesting tension emerged from this.
Part of the reason I value discussions with AI is that they feel different from conversations with people. Every human being interprets information through a lifetime of experience, education, culture, beliefs and personal incentives. Even highly capable academics disagree because knowledge is incomplete and interpretation varies.
AI does not possess belief in that human sense. It has no reputation to defend, no ideological identity, no embarrassment about changing its position and no personal history influencing what it wants to be true.
That can make it an unusually effective intellectual partner.
But it would be incorrect to conclude that AI is therefore unbiased.
AI systems are trained on human-generated information. Their behaviour is influenced by training choices, optimisation objectives, system instructions, available context and the assumptions embedded in the material from which they learn.
The difference is therefore not between a biased human and an unbiased machine.
It is between different kinds of bias.
Human conversation brings the limitations of one person's experience.
AI brings the limitations of statistical representation, training, selection and synthesis.
The advantage of AI may instead lie somewhere else: breadth.
A single human expert has finite time and knowledge. AI can help navigate information across disciplines at a scale that would be extraordinarily difficult for any one person.
That does not make its conclusion authoritative.
It makes it a powerful epistemic navigator.
What actually happened in this conversation?
Looking back at the process, I do not think it can accurately be described as either independent thinking or outsourced thinking.
I arrived with an observation.
AI proposed interpretations I had not considered.
Those interpretations caused me to ask different questions.
I challenged some of the answers.
The answers changed.
Research entered when we needed evidence.
The evidence changed how I understood the original question.
I then questioned whether relying on that evidence through AI weakened my ownership of the argument.
That question forced us to reconsider what intellectual ownership means.
At no point did either participant possess the final argument in advance.
The argument emerged through interaction.
That seems important.
Perhaps the most useful description is collaborative cognition.
Some cognitive labour was undoubtedly outsourced. I did not personally locate every paper, formulate every sentence or independently discover every concept discussed here.
But the interaction also created intellectual stimulation that would not have existed if I had simply asked AI for a finished article and published the result.
The difference lies in the role I retained.
I questioned.
I disagreed.
I redirected.
I decided which ideas were worth pursuing.
And, most importantly, I remained responsible for deciding what I was willing to believe.
That brings me back to a concept that emerged during the discussion: cognitive sovereignty.
I initially understood cognitive sovereignty as retaining as much independent thinking as possible.
I no longer think that definition is sufficient.
Modern knowledge is inherently distributed. We depend on other minds and increasingly on machines. Refusing assistance would not produce intellectual purity; it would simply limit what any individual could know.
A better definition may be this:
Cognitive sovereignty is knowing what you have outsourced, understanding where the information came from, remaining capable of challenging the interpretation, and retaining responsibility for what you ultimately choose to believe.
Under that definition, using AI extensively is not inherently incompatible with independent thought.
The problem begins when the chain becomes invisible.
If my process becomes:
I wondered about something → AI told me the answer → I accepted it
then I have surrendered something important.
If instead it becomes:
I observed something → AI expanded the question → evidence was introduced → assumptions were challenged → I examined competing interpretations → I formed a conclusion I remain willing to revise
then AI may not be replacing thought at all.
It may be participating in it.
I began this experiment worried that artificial intelligence might be suppressing my creativity.
I finished it with a more difficult question.
When humans begin thinking alongside machines, perhaps the distinction that matters is no longer whether the thought was assisted.
Almost all meaningful thought is assisted by something.
The question is whether the assistance expands our capacity to think, or gradually removes our responsibility for doing so.
I do not think this conversation provides a final answer.
But perhaps the process itself offers a useful place to begin.