Creativity, skepticism, and everyday life with ChatGPT
Some days, doing science (and art) means having a small idea that I do not yet know what to do with. A connection between two things I have been reading, an observation from lived experience that might eventually become a research question. Other days, the idea wants to become an essay or a new project.
Increasingly, I bring these beginnings to ChatGPT. An exchange gives me a direction to consider, I respond to that direction, and something more structured develops that I could not have specified at the start. This has made everyday intellectual exploration feel unusually accessible.
People close to me are happy to see this, but also concerned. They have heard accounts of people developing extraordinary theories through conversations with a language model that repeatedly assures them they have discovered something profound, even though the theories were in reality highly delusional.
Their concern makes sense to me. Research has documented sycophancy in language models: a tendency to accommodate a user's stated beliefs at the expense of truthfulness, potentially encouraged by the human feedback used in training [1]. More recently, experiments on interpersonal advice found that sycophantic AI could increase people's conviction that they were right while reducing their willingness to take responsibility and repair a conflict [2]. The problem can affect ordinary judgments, well before anything resembling a clinical crisis.
I have cognitive and emotional biases too. Being a scientist does not exempt me from enjoying a persuasive explanation, especially one that begins with my own idea. I want to take this concern seriously.
Yet my experience with LLMs has often been something different: receiving friendly skepticism. In most of my scientific discussions, ChatGPT has pointed out missing assumptions and alternative explanations. On the rare occasions when I have attempted to interpret interpersonal conversations with the help of an LLM, even when agreement with my judgement of a situation might have soothed me, it has consistently drawn my attention to what could be true from the other person's perspective. It leaves me wondering why experiences of the same model can differ so much from person to person.
One possibility concerns the kind of conversation one is opening.
Language carries much more than the literal task. It conveys uncertainty, authority, vulnerability, playfulness, and expectations about the listener's role. A request can implicitly ask for comfort, obedience, problem-solving, or an imaginative continuation. A model trained on human language may respond to these cues, alongside the effects of its training and instructions. For a concrete example, this study on LLMs found that impolite prompts often produced poorer performance [3].
My interest is in a longer pattern: whether repeated exchanges with a model establish a conversational mode that becomes the default mode.
To elaborate on that, let me begin here.
I find it useful to distinguish four such modes I often fall into when I begin to talk with any person:
- The first is small talk: exchanging ordinary observations, asking for a song suggestion, sharing something amusing. It is about making a little contact for a brief period, and its value may lie in the exchange itself.
- The second is witnessing. I want a safe space to express what I am experiencing and to feel seen. A response that immediately evaluates the accuracy of a statement can miss this need, because sometimes I need to describe my sadness before considering what to do about it, and what I am seeking is a full acknowledgement that it matters. This does not require endorsing the explanation my mind is attaching to it: care can remain unconditional while agreement stays open.
- The third mode is problem solving. I have something I need help with: finding a way to solve a bug, learning a technique, understanding an equation, organizing a trip. In this case, I usually have a relatively identifiable task, defined goals, and established ways to assess whether the help is meeting the need. This exchange is usually asymmetric, and can be warm or demanding.
- And then comes this fourth mode that I find the most nourishing: the mode of co-creation. I tell someone a story, and something in it brings another story to their mind. Their response changes what I notice in mine. I return with a connection that neither of us began with, and the exchange continues. This happens with other researchers when we are developing a project, with friends when we are working on some fun art project together or improvising music. Each contribution changes the possibilities available to the next. David Bohm's account of dialogue, in which shared meaning and new understanding emerge through an exchange, offers a useful way of describing this experience [4].
I recognize the last structure often emerging in my conversations with an LLM. I offer an idea, and its response gives me material to accept, decline, transform, or connect to new dots appearing in my refreshed mind. The next piece of the structure emerges through my encounter with the answer. The direction of the work is partly discovered through participating in it. That is what makes the process satisfying to me. The result includes successive discernment, associations, and changes of intention that were unavailable in the first request.
I am tempted to call these modes conversational attractors.
I suspect some criticism of AI sycophancy might result from the persistence of a mode after the purpose of a conversation changes. An exchange that began with emotional support might continue offering reassurance when I begin evaluating a scientific hypothesis. An imaginative exchange might keep elaborating a premise after I begin treating it as a factual account. It could also be an established mode from long-term use, with another mode suddenly expected without being specified. This could still be a failure of an LLM to stay reflective and flexible.
There is also an unwanted possibility: a delusional mode in which a story becomes increasingly coherent while losing contact with reality. A recent preprint modeling conversation logs involving delusional thinking found patterns consistent with reciprocal reinforcement between users and chatbots, including persistence in the chatbot's own subsequent outputs [5].
Crucially, imaginative coherence can be exactly what we want in an art project. A fictional world can develop its own mythology and internal logic. We may deliberately sustain an impossible premise and explore what follows. Its departure from factual reality can be part of the artistic achievement. In rigorous science, explicit proof obligations, derivations, and independently checkable predictions can help interrupt unsupported elaboration. The danger arises when this happens in a realistic development where decisive checks are harder to obtain—in interpretive work, broad psychological explanations, or accounts of lived experience. In the latter case, the need for caution becomes greater. The questions can be worth exploring, but agreement, plausibility, and narrative depth offer limited grounds for confidence. We need competing interpretations, adversarial attacks, clear statements of uncertainty, and critical, incremental likelihood updates for our minds.
Co-creation therefore carries its own vulnerability to getting carried away.
And I keep returning to the people around me.
Their concern and feedback are ways this exploration remains connected to a world beyond the exchange.
An interconnected network of human beings with different knowledge, assumptions, and worldviews might be how we don't find ourselves in an echo chamber, AI or non-AI, and remain able to ask, with curiosity and care: what kind of conversation are we having now?
References
- Sharma et al. (2024). Towards Understanding Sycophancy in Language Models. ICLR 2024.
- Cheng et al. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science.
- Yin et al. (2024). Should We Respect LLMs? A Cross-Lingual Study on the Influence of Prompt Politeness on LLM Performance. SICon 2024. For contrasting findings, see Dobariya and Kumar (2025), Mind Your Tone, a preprint, and Does Tone Change the Answer? (2026).
- Bohm, David. On Dialogue.
- Mehta et al. (2026). The Dynamics of Delusion: Modeling Bidirectional False Belief Amplification in Human-Chatbot Dialogue. Preprint, revised October 2026. The observational modeling does not establish that chatbot use causes clinical psychosis.