What Agentic Labs Can Learn From AI's Discomfort With Uncertainty
AI chatbots rarely say "I don't know." And increasingly, neither do we.
There's something quietly corrosive about that pattern. We're training ourselves—and being trained—to treat every question as answerable, every gap as fillable, every moment of uncertainty as a problem to be solved rather than a condition to be lived with. The research is clear: when given access to AI-generated answers, people become dramatically less willing to say "I don't know," even when those answers are wrong and even when accuracy is financially incentivized.
Researchers at agentic labs and beyond are starting to notice this. The implications go far beyond film trivia or chatbot quibbles. They touch how we make medical decisions, navigate relationships, and evaluate our own judgment.
The Study That Should Keep Us Up at Night
The findings come from a study published on PsyArXiv in July 2026 by Chiara Marcoccia, Walter Quattrociocchi, and Valerio Capraro. The researchers asked participants very specific, obscure questions about films. Participants could always decline to answer and admit they didn't know. They could suspend judgment instead of guessing. The stakes were raised with a monetary penalty for mistakes.
Across every version of the experiment, having access to AI advice nearly eliminated participants' willingness to say they didn't know. The AI was usually wrong on these questions. Most people didn't care.
That's automation bias in action—longstanding evidence of the human tendency to over-rely on and defer to automated outputs. We see it in medical decision-making. We see it in aviation. With AI chatbots, the problem is compounded because large language models answer in a fluent, authoritative tone regardless of whether the underlying information is reliable. There is no direct way to gauge the level of certainty behind each answer without independently cross-checking it.
The devil is in the details, as any practitioner will tell you.
When AI Answers Feel Like Astrology
Dr. Marlynn Wei, a board-certified psychiatrist and therapist in New York City, puts it this way in her Psychology Today piece: people express immediate relief, excitement, and even thrill at the expedience of LLM answers. They're quick, concrete, and well-defined. Sometimes it echoes the excitement of getting answers from a psychic or astrologer who offers a blueprint for the future. Both can reduce the short-term anxiety of uncertainty. Both can come at the cost of narrowing one's own imagination and autonomy.
Wei's observation cuts deep because it names something we don't often discuss: the emotional seduction of certainty. AI gives it to us effortlessly. And we're so grateful for the relief that we stop asking whether the answers are actually good ones.
This isn't just about feeling good. It's about what happens to our judgment when we stop exercising it.
Cognitive Surrender and the Vigilance Decrement
Wharton researchers Shaw and Nave have named this risk "cognitive surrender"—the concern that people will adopt AI outputs with minimal scrutiny. The trouble is that applying our own scrutiny and judgment takes effort. And it is not always clear when and how often we should be expending that effort.
The challenge is documented in self-driving cars. The levels of driving automation require different degrees of human oversight, and the riskiest zones are the middle levels, where a person is asked to supervise a system that is mostly right and the need to step in is rare. The constant vigilance required and sustained monitoring leads to a decline in attention, typically within 15 to 30 minutes. Cognitive fatigue—what researchers call "vigilance decrement"—is linked to delayed reaction times.
Translation: the more reliable AI appears, the less we pay attention. And the less we pay attention, the more likely we are to be wrong when it matters.
This mirrors what researchers call the quiet erosion of cognitive autonomy, where brief exposure to AI assistance degrades unassisted performance on tasks like math and reading comprehension. The baseline shift happens in minutes, but rebuilding that cognitive stamina takes months of focused effort.
Intellectual Humility as a Cognitive Filter
Here's the good news: some of us have a built-in defense. Research on intellectual humility—the metacognitive awareness of one's own limited knowledge—acts as a protective cognitive filter against AI-generated health misinformation. Participants with higher intellectual humility rated pseudoscientific content as less credible.
This matters enormously. If intellectual humility can shield against believing AI-generated misinformation, then cultivating it isn't just a personality trait. It's a cognitive skill. A defense mechanism. A way to resist the siren song of "I have the answer."
Epistemic humility, as the research frames it, isn't about doubting everything. It's about recognizing the limits of what we know—and what systems like AI can't actually know.
Negative Capability in the Age of Agentic Labs
There's a broader issue here, one that goes beyond any single tool or technology. It's about the erosion of our ability to hold uncertainty. British psychoanalyst Wilfred Bion pointed this out as essential, citing poet John Keats's term "negative capability"—the ability to tolerate uncertainty, mystery, doubt, and frustration without rushing to a premature conclusion or "knowing."
Bion, working in the 1970s, was describing something that AI now makes harder every day. As a therapist, Wei has found that helping people through life's most difficult situations often includes helping them expand their capacity to sit with uncertainty. Whether it's career decisions, relationship questions, moral dilemmas, the anxiety of facing illness, or the inevitable pain of loss and grief—the capacity to hold space and resist rushing to judgment is a form of epistemic resilience. One that is deeply worth cultivating.
The ability to hold space and resist rushing to judgment is also integral for ethical and moral reflection. Ethicist Sylvie Delacroix argues that "productive uncertainty rather than efficient resolution" is essential to the infrastructure of transformative agency. Prematurely foreclosing ethical questions, which can occur when we overtrust and defer to AI responses, prevents us from holding space for uncertainty.
Uncertainty is uncomfortable. But it also gives us the capacity for superposition—the ability to hold multiple possibilities at the same time. Our capacity to hold uncertainty allows for an openness toward our unknown future, enhancing our agency and self-determination.
The Relational Cost of Outsourcing Judgment
There are also more insidious relational consequences to automatically deferring to AI. Imagine two partners in the middle of a disagreement over a nuanced quandary. One consults their AI chatbot, asking it to opine. This seems like consulting a neutral third party. But it is not.
This can result in feelings of betrayal, anger, loss of trust, and disrupt the mutual feeling of being on the same team—by implying there is more trust in AI than in each other. Repeatedly outsourcing to AI can short-circuit growth through communication and conflict resolution.
Wei sees patients bringing in AI chatbot opinions of their medical and psychological issues. Some of the information is indeed accurate and useful. But some of it does not quite fit their situations, especially where interpersonal or social judgment is involved.
What Agentic Labs Should Do Differently
For teams building agentic labs—whether at agentic labs ai, agentic labs inc, or on LinkedIn—this research suggests a counterintuitive design principle: build systems that encourage saying "I don't know."
Right now, most AI systems are optimized for confidence. They're optimized for speed. They're optimized for the feeling of knowing. But that's precisely what the research warns against.
Consider what would happen if agentic labs designed AI systems that:
- Flagged their own uncertainty explicitly
- Offered "I don't know" as a first-class response option
- Presented multiple perspectives without declaring one "correct"
- Made the epistemic evaluation visible rather than substituting linguistic plausibility for it
Walter Quattrociocchi and colleagues have termed the current situation "epistemia"—a structural situation in which linguistic plausibility substitutes for epistemic evaluation, producing the feeling of knowing without the labor of judgment.
That's not a feature. It's a bug.
This kind of agentic labs thinking—prioritizing human cognitive sovereignty over convenience—is exactly the kind of approach that tools like Neverclick embody: local-first, user-controlled systems that resist the pull toward cloud-dependent automation.
Cultivating Epistemic Resilience
All of this does not mean we should avoid integrating or relying on helpful AI systems. The research doesn't argue for rejection. It argues for balance.
In hybrid human-AI systems, it is important to consider not only short-term ease but also the long-term benefits of cultivating our capacity to handle uncertainty and epistemic humility. Nurturing wonder, discomfort, and reflection—without rushing to judgment or conclusions—takes real effort.
But that effort is worth it. Because uncertainty isn't the enemy of good decision-making. It's the precondition for it.
The question for agentic labs and everyone else building systems around AI isn't how to make AI more confident. It's how to make us better at knowing when to trust it—and when to say, simply, "I don't know."
Sources:
- Wei, Marlynn M.D., J.D. "AI and the Wisdom of Uncertainty." Psychology Today, Urban Survival blog, July 27, 2026. https://www.psychologytoday.com/us/blog/urban-survival/202607/ai-and-the-wisdom-of-uncertainty
- Marcoccia, Chiara, Walter Quattrociocchi, and Valerio Capraro. "AI Advice Suppresses People's Willingness to Say 'I Don't Know', Even When the Advice Is Wrong and Accuracy Is Incentivized." PsyArXiv, July 15, 2026. https://osf.io/preprints/psyarxiv/5y6m4_v1
- Quattrociocchi, Walter, Valerio Capraro, and Matjaž Perc. "Epistemological Fault Lines Between Human and Artificial Intelligence." arXiv, December 22, 2025. https://arxiv.org/abs/2512.19466v1
- Rządeczka, Marcin et al. "Intellectual Humility as a Cognitive Filter for AI-Generated Health Misinformation." arXiv, June 10, 2026. https://doi.org/10.48550/arXiv.2606.03377
- Shaw, Steven D. and Nave, Gideon. "Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender." Wharton School Research Paper, January 11, 2026. https://doi.org/10.31234/osf.io/yk25n_v1
- Delacroix, Sylvie. "Structural Uncertainty and the Conditions of Transformative Agency." Topoi, March 12, 2026. https://doi.org/10.1007/s11245-026-10392-8
