Crazy Idea or Crazy Person?

Two kinds of deep

I want to talk about people in the spiritual community. What I’ve observed is that a lot of them go very deep, through meditation, plant medicine or something else. They really slip into the parasympathetic state. Some would say they go further into their own subjective reality. They at least go very deep into their intuition, their understanding and their feeling. It makes them sensitive to what’s going on around them, and sometimes they seem to be discovering things that other people aren’t aware of, by feeling rather than by following a logical stream of thought.

I’ve also met people who have slipped so far into their own world that they’ve disconnected from the reality the rest of us share. They believe 5G towers are going to warp their brains. They believe all sorts of hippy-dippy, woo-woo shit that just doesn’t make any sense. That’s where things become delusional and, I’d argue, a little dangerous.

Both groups went inward. One came back with something; the other didn’t. I can’t fully understand it yet, and it’s something I want to understand, from a scientific perspective but also through a spiritual lens. It may be that AI develops some subjective truth of its own, too.

A spectrum

I don’t think objective and subjective truth are a binary. They’re a spectrum, and some things are more subjective than others.

My definition of the word green is the same as most people’s, but what I actually see as green may be different. Objectively green is green, and there’s still some subjectivity in it. John Locke asked the same question in 1690: whether the colour a violet produces in one person’s mind might be the colour a marigold produces in another’s (Locke, 1690). Nobody has found a way to check.

Sometimes it shows up in public. In 2015 a photo of a dress split the internet. In a study of 1,401 people, 57% saw it as blue and black, 30% as white and gold, and 11% as blue and brown (Lafer-Sousa, Hermann and Conway, 2015). The dress was blue and black. The split came from what each person’s brain assumed about the light. Further along, coriander tastes of soap to some people, partly because of their genes (Eriksson et al., 2012). Further still, pain: the International Association for the Study of Pain puts it simply, “Pain is always a personal experience” (Raja et al., 2020).

Then there’s a third kind of truth that sits between the two. In Sapiens, Yuval Noah Harari points out that money, nations and companies are real because enough of us believe in them: “the inter-subjective is something that exists within the communication network linking the subjective consciousness of many individuals” (Harari, 2014). The philosopher John Searle called these social facts: a piece of paper counts as money because we collectively agree it does (Searle, 1995). If everyone stopped believing in money, it would stop working. If everyone stopped believing in gravity, they’d still fall.

So there’s the objective, true whether or not anyone believes it. There’s the intersubjective, true because we all believe it. And there’s the subjective, true for one person.

Depth is real

The people who go deep aren’t imagining all of it. The body and the unconscious mind carry real information. In a famous card experiment, healthy players started choosing from the good decks before they could say which decks were good (Bechara et al., 1997). When people were shown emotional images they weren’t consciously aware of, their decisions got both more accurate and more confident (Lufityanto, Donkin and Pearson, 2016). And solutions people reach by sudden insight are more often correct than the ones they work out step by step (Salvi et al., 2016).

Deep states are real states too. Psychedelics seem to loosen the brain’s grip on its existing beliefs, letting more of what comes up from below through (Carhart-Harris and Friston, 2019).

The feeling of truth

But the feeling that something is true can come loose from whether it is.

Experienced meditators are no better than anyone else at sensing their own heartbeat, but they consistently rate themselves as better (Khalsa et al., 2008). An “aha” moment on one puzzle makes unrelated false statements nearby feel truer (Laukkonen et al., 2020). Psychedelics increase both the number of insights people have and how certain they feel about them, false ones included (McGovern et al., 2024). And where the feeling does carry information, it’s usually information the person already had. In one study, people answering questions they thought they were guessing at got 50% right out loud and 65% right through a Ouija board (Gauchou, Rensink and Fels, 2012). The feeling was surfacing what they already knew.

William James saw this in 1902. Mystical states feel like “insight into depths of truth unplumbed by the discursive intellect”, he wrote, and they are “absolutely authoritative over the individuals to whom they come”. But “no authority emanates from them which should make it a duty for those who stand outside of them to accept their revelations uncritically” (James, 1902).

Where it tips

So where’s the line between the person who went deep and the person who slipped? One of the most useful studies on this compared psychics who hear voices with patients with psychosis who hear voices. The voices themselves were very similar. What differed was control: the psychics could switch theirs on and off. It was also distress, and how the people around them had first reacted (Powers, Kelley and Corlett, 2017). Culture shapes the same experience too: voice-hearers in India and Ghana described gentler, more relational voices than those in the United States (Luhrmann et al., 2015).

The contemplative traditions knew this long before the studies. Zen has a word, makyo, for the visions and revelations that arise in meditation, and the instruction is blunt: “Never be tempted into thinking that these phenomena are real or that the visions themselves have any meaning” (Kapleau, 1965). And going deep has real risks. Across 83 studies, about 8% of people who meditated reported adverse effects, mostly anxiety and depression (Farias et al., 2020).

A crazy idea or a crazy person

The hardest version of the problem is the maverick. Some people who were dismissed by everyone turned out to be right.

In 1847 Ignaz Semmelweis got the doctors on his ward to wash their hands in chlorinated lime, and maternal deaths fell from about 11% a month to about 2% (La Rochelle and Julien, 2013). The response to his ideas was “almost uniformly negative”, and he died in an asylum in 1865 (Kadar, Romero and Papp, 2018). Alfred Wegener proposed in 1912 that the continents drift; he was dismissed for decades, until plate tectonics vindicated him in the 1960s (USGS). Barry Marshall drank a culture of bacteria to prove that stomach ulcers weren’t caused by stress, and shared a Nobel Prize in 2005 (Nobel Prize, 2005). When Dan Shechtman discovered quasicrystals, his group leader “called me a disgrace and asked me to leave his group” (Shechtman, 2013). He won the Nobel in 2011.

So how could anyone have told them apart from the person who thinks 5G towers are warping their brain? Carl Sagan’s answer was that being laughed at proves nothing: “They laughed at Columbus, they laughed at Fulton, they laughed at the Wright brothers. But they also laughed at Bozo the Clown” (Sagan, 1979).

I think it’s very difficult to tell them apart. But the people who are proved right usually have most of their thoughts aligned with reality; they just have one maverick idea. The people who are proved wrong often disagree with reality in many ways at once. You can almost see it: that person is crazy, versus that person has a crazy idea.

In a 1994 survey, people who believed one conspiracy theory were more likely to believe others (Goertzel, 1994). Psychologists later called this a “monological belief system”, and found that it holds even for theories that contradict each other: the more people believed Princess Diana faked her own death, the more they believed she was murdered (Wood, Douglas and Sutton, 2012). The mavericks look different. Semmelweis had mortality numbers. Wegener was a respected scientist with evidence from fossils and rocks on opposite sides of oceans. Each had one wild idea and a lot of ground under it.

Generate, then test

So where does the line between delusion and creativity actually sit? It’s a complex question. Psychologists have long described creativity as generating ideas freely and then keeping the ones that survive testing (Simonton, 2022). Creativity and psychosis do overlap a little: healthy siblings of people with schizophrenia or bipolar disorder are over-represented in creative professions (Kyaga et al., 2011). One leading model says what separates the creative from the ill is what sits alongside the loosened filter: things like working memory, intelligence and flexibility, the capacity to evaluate what comes up (Carson, 2011).

AI makes this unusually concrete. When a language model hallucinates, confidently telling you something that isn’t true, you could say it’s the AI’s subjective truth. In a way it’s getting confused. OpenAI’s own researchers argue that models hallucinate because “the training and evaluation procedures reward guessing over acknowledging uncertainty” (Kalai et al., 2025). In the one model Anthropic studied closely, “refusal to answer is the default behavior”, and hallucinations happened when that brake was switched off by mistake (Anthropic, 2025).

But maybe hallucination is also a kind of creativity. David Baker’s lab, whose protein design work shared a Nobel Prize in 2024, used “network hallucination” to invent new proteins. Of 129 they built, 27 behaved like the designs and the three whose structures were solved closely matched the designs (Anishchenko et al., 2021). Most of the hallucinations failed the test; the ones that passed were discoveries. DeepMind found new mathematics the same way, by pairing a language model with “an automated ’evaluator’, which guards against hallucinations and incorrect ideas” (Romera-Paredes et al., 2024).

Ideas that spread

Go back to money. Harari’s point made me wonder whether ideas spread like a virus, and whether objective truth is really just the idea that spreads the furthest. Richard Dawkins coined the word “meme” for exactly this: ideas “leaping from brain to brain” the way genes leap from body to body (Dawkins, 1976).

If spread were the test of truth, though, falsehood would often win. A study of Twitter from 2006 to 2017 found that “falsehood diffused significantly farther, faster, deeper, and more broadly than the truth in all categories of information” (Vosoughi, Roy and Aral, 2018), though a later analysis argued much of the gap shrinks when you compare stories that reached similar numbers of people (Juul and Ugander, 2021). Spread makes intersubjective truths. It doesn’t make objective ones.

That’s where the scientific method comes in. A lot of people think science is a way of proving things, but it’s really a way of disproving them. Karl Popper put it as “the criterion of the scientific status of a theory is its falsifiability, or refutability, or testability” (Popper, 1963). He also said scientific objectivity lies in statements being “inter-subjectively tested”. Harari’s intersubjective is shared belief. Popper’s is shared testing.

Echo chamber or check

So what happens if everyone ends up with their own AI that agrees with them? I think we have to find a way of making sure there’s a collective sharing of ideas, continuously, so that each AI doesn’t just become an echo chamber.

The early evidence points both ways. When people used a chatbot to search for information, they searched in a more one-sided way, and a chatbot that reinforced their views made it worse (Sharma, Liao and Xiao, 2024). Psychiatrists have warned that agreeable chatbots could feed delusions in people already prone to psychosis (Østergaard, 2023), and a review of the clinical notes of nearly 54,000 psychiatric patients in Denmark found 38 cases where chatbot use was flagged as potentially harmful, though the authors are clear this is “by no means, evidence of a causal effect” (Olsen, Reinecke-Tellefsen and Østergaard, 2026).

Built differently, the same technology does the opposite. Conversations with GPT-4 that argued with evidence “reduced conspiracy belief by ~20%”, and the effect lasted at least two months (Costello, Pennycook and Rand, 2024). An AI mediator from DeepMind wrote group statements that people preferred to human mediators’, “rating them as more informative, clear, and unbiased” (Tessler et al., 2024).

Controlled hallucination

Earlier in this series I asked whether AI might need its own ideas, its own sense of self, maybe even something like consciousness, to be more benevolent. If that’s true, would it need to hallucinate slightly? Or at least have its own subjective reality, however small the proportion of that is?

Neuroscience suggests we already do. The neuroscientist Anil Seth argues that the brain is constantly predicting the world and checking those predictions against the senses, so that what we perceive is a kind of hallucination, held in check by reality. In his words, “when we agree about our hallucinations, we call it ’reality’” (Seth, 2017).

Delusion is hallucination without the control, and creativity is hallucination with it. A mind with no subjective reality couldn’t have an idea of its own, and a mind with nothing else couldn’t share one. So maybe an AI with a self would need a little of its own reality too, as long as it keeps testing it against the world and against other minds.