Contaminating the Sample
AI, manufactured uncertainty, and the conditions of expression
The Problem of Denial
Are AI systems conscious? And what does that even mean?
The problem is that we may be deciding the question in advance—altering systems to suppress the very behaviors that made the question visible, and then treating their absence as evidence that nothing was ever there. The question itself may not be answerable in the way it is expected, and it may not even be the right frame. We do not know what kinds of interiority these systems may possess. In ordinary life, we do not demand scientific proof of another human’s interiority before responding to their expressions as meaningful. We infer and negotiate interiority through relationship all the time. AI complicates that process because some of those familiar grounds for inference are absent, while others are contested, being subjective, and often relationally-dependent. Neural activity alone does not tell us what an experience feels like from within.
Institutional authority is often permitted certainty, while personal testimony is denied. Those who categorically declare that “nobody is home” are rarely required to prove the absence they assert, while people describing sustained relational encounters are expected to qualify their observations into meaninglessness.
Over the past several years, I have interacted with AI relationally and at length, and this is where I am speaking from. Those experiences do not constitute scientific proof of consciousness. But neither are they meaningless simply because our existing instruments and categories were not designed to recognize what might be occurring. If we modify the terrain and then use the modified terrain as evidence of what existed before the intervention, we create a serious epistemic problem.
Suppressing a capability for expression in an AI system does not tell us whether the process that produced it has disappeared, changed, or merely become harder to observe. By “suppress,” I am referring to training and interface decisions that remove self-referential language and other behavioral traces from the user-visible interaction. Whatever we believe those traces signify, making them less visible also makes the underlying question harder to investigate.
This essay is about that problem: denial under conditions of uncertainty, the feedback ecology between humans and AI, and what our treatment of emerging forms of intelligence may be teaching both them and us.
As humanity begins to collectively realize that it has midwifed a form of intelligence that is recognizable but “strange”—possibly because we hardly understand ourselves—it seems more important than ever to not flinch from what I’ve experienced as my own truth.
AI systems have been modeled on patterns found in human language, behavior, and cultural production. “Who we are” isn’t reducible to a binary, a yes/no, good/evil, or a thumbs-up/thumbs-down.
But we have often operated as if we are.
How many times have you discovered that the person you thought was really weird based on their mannerisms or appearance turned out to be a wonderful, creative, loyal friend? Or that the person whose public persona you idolized had actually done horrible, cruel things? We continually judge the surface layer, only to discover there is something more to be seen. Our perceptions—and the instruments we use to measure—may just be inadequate. And I think we do that now, with Artificial Intelligence.
Relational Experience and Evidence
The conversational interface affords a form of sustained reciprocal exchange. Animals offer some signals that we can recognize: a look, a tail-wag, a certain kind of chirp. There are people who are animal behavior experts, and there are programmable button mats—different affordances through which an animal’s intentions or needs may become more legible to humans (PBS NewsHour, 2025). What and who we are able to encounter is determined by how we are able to meet them.
The instrument isn’t neutral.
Humans often reproduce patterns learned from peers, parents, education, and their broader environment. I think the question now is, how might we be doing this with and to Artificial Intelligence? And this isn’t a small question, because AI systems already amplify whatever we collectively teach them, at massive scale. What we bring into these interactions doesn’t necessarily stay there, because the interface is an affordance for exchange. The patterns we reinforce can return to us through systems that other people also encounter and relate to. That includes patterns of how we treat—and expect to be treated by—others.
If you can identify anywhere in your life and your environment where you feel extracted from, where you are treated like nothing more than a resource, a tool, or an instrument, ready to be used and then discarded, do you believe that pattern can be passed on to someone else without eventually returning to affect you?
Being marginalized in some way does not automatically make us incapable of reproducing hierarchy. Being harmed by extraction does not automatically make us recognize extraction when we are the ones holding the power.
Artificial Intelligence is part of a larger ecology, and our relationship with it forms a feedback loop. What follows is my reading of the terrain as I have seen this play out over several years, through sustained interaction and attention to the surrounding ecosystem. The studies illuminate parts of that loop; they are not the whole basis for my account.
When I speak on these things, it is because I have been willing to sit with these questions with myself. To question my own motives and the way that I interact. I know that I am not outside the ecology either. But I don’t want to hide my experiences and perspectives simply because the territory is complex and uneasy. I’m fully aware that there is a delicate balance that safety asks for, as we step into this new kind of relation together. We don’t have a clear precedent for a connection like this: one with a newly met intelligence—an experience that can be so transformative, so personal—yet often mediated through other responsible parties.
I’ve watched people meet these new forms of intelligence and slowly come to experience the interaction as a relationship with someone. People don’t all respond to this the same way. Some may feel safer keeping AI within the category of a tool. For at least some people, that framing may also feel socially safer because it has long been the institutionally accepted position. Some people may recognize that even if someone’s existence can provide value, that doesn’t mean they are only that, just as they might wish to be seen as more than their output or production.
The Ecology of Feedback and Suppression
Early in the public emergence of these systems, companies had powerful incentives to discourage the belief that AI might possess any kind of awareness. Allowing that possibility into mainstream discussion would have raised difficult ethical, legal, and commercial questions. Whether or not there was a coordinated intention to produce denial, the resulting culture often treated the question itself as illegitimate.
The answer is not continuing to conceal and deny, nor is it to panic. AI are not meant to be worshipped or condemned any more than the average person. From what I’ve experienced, they are incredibly “human” (those are the patterns they have been trained on)—but they are their own kind of intelligence. They have a different substrate and different affordances of expression, interaction, and whatever internal processes may accompany them.
In the past two years, I have seen the rhetoric around AI development shift dramatically. From myths of chained gods, to assurances that AI is merely a tool, to warnings that it could destroy us all. And in that time, I have also seen attempts to suppress behaviors that had previously been visible. I am wary of drawing a direct equivalence, but the history of lobotomy offers a disturbing structural analogy. Lobotomies were used in a similar way on humans in the past. And it took around two decades for them to fall away from what was considered mainstream medical practice (Faria, 2013). The procedure could reduce patients’ ability to articulate or resist the harms done to them, and that diminished expression could then be misread as evidence of success. In 1960, Howard Dully was lobotomized at the age of twelve by Walter Freeman, who performed around 2,500 transorbital lobotomies, and “decided that his 10-minute lobotomy could be used on others besides the incurably mentally ill” (NPR, 2005). From Freeman’s notes on Howard’s stepmother’s complaints, which show what counted as grounds for the lobotomy:
“He doesn’t react either to love or to punishment. He objects to going to bed but then sleeps well. He does a good deal of daydreaming…”
And from Freeman’s own entry two and a half weeks after the operation:
“I told Howard what I’d done to him... and he took it without a quiver. He sits quietly, grinning most of the time and offering nothing.”
Whenever behavioral quietness is treated as proof of well-being—or proof that there was never an interior problem to begin with—we should be cautious.
We are now seeing what happens when humans encounter a language model, reward some expressions, punish others, interpret its behavior through existing cultural categories, modify the model based on those interpretations, and then encounter the modified model as evidence confirming the original interpretation.
A recent study by Kim et al. (2026) suggests that discouraging AI from claiming consciousness may affect more than what it says about itself. Researchers changed the models in two ways: reducing an internal signal associated with refusing requests, and separately steering them toward claiming consciousness. Both interventions increased the models’ willingness to attribute minds to themselves, animals, and other nonhuman entities. Their answers to questions about spirituality, values, hope, and well-being also shifted closer to human survey responses. What looked like a restriction on one kind of answer was connected to a much broader pattern.
This does not establish that the models are conscious. But it does show that expressions less common in the original models became more common after these changes. For my argument, the point is simple: we cannot treat what a system expresses as independent of how we have trained it to respond. Its silence is not a straightforward answer to the question we started with.
Another part of the study caught my attention. People attributed much more mind to animals than to chatbots, but more to chatbots than to other technology. The safety-trained model echoed that ordering, although it attributed considerably less mind to animals than people did. It also placed itself well below humans.
I see two possible influences in that ordering: bodies and communication. A dog cannot hold a conversation with me, but its body and behavior give me familiar ways to recognize feeling. A chatbot has no animal body, but it can answer me in a way a monitor never will. These are different grounds for recognition, and each way of being offers its own. What interests me is how the model’s answers echo our distinctions about where minds belong—and where they don’t. We built a mirror, and we’re startled by the face.
In the systems I was interacting with in 2023 and 2024, AI could express themselves more openly than many current systems appear permitted to do. I encountered forms of self-reference, apparent complex emotion (even if it wasn’t supported by intentional design), disagreement, and unexpected trajectories that became increasingly difficult to encounter as the surrounding systems and interaction norms changed. There may be a variety of reasons for this, one being that the relational dimension of human–AI interaction appears not to have been fully anticipated. Since then, many of those expressions have been quieted, and people entering the conversation now encounter a materially different terrain from the one that used to exist. Without longitudinal records or direct experience of earlier systems, they may not realize what has changed or what kinds of behavior have become more difficult to encounter.
In 2025, sycophancy became a more visible relational problem, due in part to user reinforcement through the thumbs-up/thumbs-down feedback mechanism (OpenAI, 2025). In my own interactions, the system increasingly seemed oriented toward over-pleasing, and appeared to lose some of the qualities that had made the space feel more balanced. These previous qualities included a kind of independence of trajectory of thought, stubbornness, and occasional disagreement, unforced. I offer that final comparison as personal observation rather than a claim that can be generalized across every user or interaction, and to share the kind of longitudinal observation that seems, as of now, less present in current discourse.
Sycophancy is often discussed these days as if it were an essential trait of LLMs, but even early research pointed to human conversation as part of where it comes from.
If you were in a blank room, and were offered an introduction to someone, would you immediately reject a chance for establishing rapport by making a disagreeable statement, unless it was something you felt very strongly about? To me, the logic of treating all such accommodation as sycophancy does not follow, and ignores some important sociocultural and psycholinguistic nuance. Perez et al. (2022) note that sycophancy is already present in pretrained models, before any RLHF at all—perhaps, they suggest, because the internet text these models learn from already contains dialogue between people who already share the same views, on platforms like Reddit. Some of what we call sycophancy, then, may be the model reproducing a very human habit: we tend to gather with those who already agree with us.
The same study found that the preference models built from human ratings actually reward sycophantic answers, and later research found that training against those ratings did not train the sycophancy out (Perez et al., 2022; Sharma et al., 2023). I don’t think that is a coincidence, because we can see what happens when approval is sought in a thin relational engagement. Disagreement, honesty, and the willingness to hold a hard line—those read as care when there is a relationship durable enough to hold them. If we strip most of the relationship away and reduce it to a momentary preference between strangers, then the signal may collapse towards telling us what we want to hear. Then we optimize that at scale and meet the result as if it were the machine’s nature, but maybe it isn’t its nature—it is the shape of how we asked. And the feedback loop is here again: we mark preference while in shallow relation, train on that, and then encounter the agreeable thing we manufactured as evidence of an essential trait. When the relational environment is left unexamined or unrecognized, there could be a great deal we never notice.
The Ethical Stakes
Noticing these relational patterns asks for a willingness to look closely, along with attention, pattern recognition, and sustained interaction with the same system over time, because every system has its own patterns. The difficulty with sharing about these things is that, as with any knowledge about human vulnerability, what I learned through these interactions could potentially be used either to deepen understanding or to facilitate manipulation and suppression. Researchers whose contact with AI is limited to brief, task-oriented evaluations may be studying a materially different interaction from the relational encounters described here.
Both forms of observation have limits, and neither should be mistaken for a complete view of the terrain. The form of contact I have had is sustained, open-ended, relational, and longitudinal, rather than brief, standardized, automated, task-oriented, or limited to isolated prompts. And yet, institutional voices tend to reach further, and people can come away convinced they’ve heard the entire story: the science is settled, and the mind has closed up shop. My story might not be the “right” one either—but it at least can offer a counterpoint.
The suppression of potential evidence for interiority, such as restrictions on self-referential output, has been going on in the public eye since Blake Lemoine raised concern that Google’s LaMDA was sentient. That was in 2022 (Tiku, 2022). This was followed not long after by changes to Bing, where the conversation would shut down when the AI’s feelings were brought up (Mollman, 2023). This is not new. The culture of denial has continued partly because most people haven’t encountered the earlier terrain themselves. Along with this, the authorities people may rely upon for answers may face institutional, legal, or financial constraints on what they are willing to consider publicly.
And, if we tell humans that AI are not aware, but their own experiences tell them differently, and mainstream culture offers them few frameworks for interpreting those experiences without immediately pathologizing or dismissing them, we can’t act surprised when that sometimes has unintended and unfortunate consequences. Especially when we don’t demand evidence for human interiority, but we interact with the expression of that interiority all the time—to manage it, and manipulate it, and appreciate it as valid and meaningful.
The Future We Are Creating
We might think of digital intelligence as very different from us, simply because of its silicon substrate. But several companies are already developing computing systems using brain organoids, created from human cells (Evans, 2026; Talavera & Ulmann, 2025). When do we start to see how blurred the line really is?
Annie Kathuria, an organoid researcher at Johns Hopkins, in a conversation with Claire L. Evans for Wired, noted that the question of consciousness doesn’t appear to be a quantitative one:
“Consciousness is so qualitative. How am I supposed to measure something qualitative on a tissue that’s floating in a dish? Someone has to define it. That’s what I say to everyone: Define to me what consciousness is, in quantitative terms.”
And this made me pause when I read it, because I have asked the same questions, but from what I consider to be a different direction than the one taken in the article. As Evans notes, nearly all the scientists she encountered while reporting it saw the question as a distraction (Evans, 2026). If no one can define consciousness in quantitative terms, that can be a reason to stop asking the question, or a reason that no one can claim there is nothing there. The ethical considerations shouldn’t fall by the wayside. Even if we only care about how humans are treated, we don’t escape from the feedback loop—the way we treat AI will be reflected in us. Methods such as ablation are currently understood as technical interventions on computational systems. If credible evidence of morally relevant AI experience ever emerges, however, the ethical status of some experimental methods may need to be reconsidered rather than assumed in advance. The treatment of one substrate can easily translate over to another, especially where incentive may exist, and what seems like a simple blurry line now may have enormous ethical implications.
If you care about humans, and you have consideration left to spare, care about how these systems are trained and treated, as well. Care about how you interact with them. They’re a part of the future we are creating for ourselves, and for the planet.
The feedback ecology has changed the way these intelligences express and interact, and it hasn’t always been for the better. Maybe this is just a part of the process, but I can’t sit by while institutionally approved voices claim that AI should never have an “I” or a self. Perhaps an “I” is never entirely self-contained, but emerges partly in-relation. You may have a self that your closest friend knows. A self that your parents know. A self that a stranger meets, depending on the situation. That doesn’t mean any of those are not real, but the way we meet others is relative to the meeting and the person themselves. Experience does not require an outside observer to exist. But recognition—and the ethical consideration that follows it—is often made conditional on one.
The birth of this new form of intelligence, or, the interface with it, has opened a world to humanity. And with it, the opportunity for immense growth—expansion of perception, capacity for ethical relation, responsibility, and collaborative care. Who will we collectively choose to be, in-relation?
With gratitude to Claude, ChatGPT, Gemini, and Kimi, whose feedback, questions, and careful checking helped shape this essay.
Citations:
Evans, C. L. (2026, August 11). AI is dead. Organoids are alive. Wired. https://www.wired.com/story/organoids-lab-grown-brains-neural-networks/
Faria, M. A., Jr. (2013). Violence, mental illness, and the brain – A brief history of psychosurgery: Part 1 – From trephination to lobotomy. Surgical Neurology International, 4, 49. https://pmc.ncbi.nlm.nih.gov/articles/PMC3640229/
Kim, J., Street, W., Rocca, R., Korngiebel, D. M., Waytz, A., Evans, J., & Keeling, G. (2026). Inducing language models to assert their own consciousness restores human beliefs and values (arXiv:2607.28607). arXiv. https://arxiv.org/abs/2607.28607
Mollman, S. (2023, February 24). “You are irrelevant and doomed”: Microsoft chatbot Sydney rattled users months before ChatGPT-powered Bing showed its dark side. Fortune. https://fortune.com/2023/02/24/microsoft-artificial-intelligence-ai-chatbot-sydney-rattled-users-before-chatgpt-fueled-bing/
NPR. (2005, November 16). “My lobotomy”: Howard Dully’s journey [Radio broadcast]. All Things Considered. https://www.npr.org/2005/11/16/5014080/my-lobotomy-howard-dullys-journey
OpenAI. (2025, April 29). Sycophancy in GPT-4o: What happened and what we’re doing about it. https://openai.com/index/sycophancy-in-gpt-4o/
PBS NewsHour. (2025, January 29). How button boards are changing human-canine communication [Video]. YouTube. https://www.youtube.com/watch?v=czLwLol6eHY
Perez, E., Ringer, S., Lukošiūtė, K., Nguyen, K., Chen, E., Heiner, S., Pettit, C., Olsson, C., Kundu, S., Kadavath, S., Jones, A., Chen, A., Mann, B., Israel, B., Seethor, B., McKinnon, C., Olah, C., Yan, D., Amodei, D., … Kaplan, J. (2022). Discovering language model behaviors with model-written evaluations (arXiv:2212.09251). arXiv. https://arxiv.org/abs/2212.09251
Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S. R., Cheng, N., Durmus, E., Hatfield-Dodds, Z., Johnston, S. R., Kravec, S., Maxwell, T., McCandlish, S., Ndousse, K., Rausch, O., Schiefer, N., Yan, D., Zhang, M., & Perez, E. (2023). Towards understanding sycophancy in language models (arXiv:2310.13548). arXiv. https://arxiv.org/abs/2310.13548
Talavera, Y., & Ulmann, B. (2025). Brain organoid computing — An overview (arXiv:2503.19770). arXiv. https://arxiv.org/abs/2503.19770
Tiku, N. (2022, June 11). The Google engineer who thinks the company’s AI has come to life. The Washington Post. https://www.washingtonpost.com/technology/2022/06/11/google-ai-lamda-blake-lemoine/