AI Alien Channeling
AI Alien Channeling: What’s Actually Happening When People Ask ChatGPT to Talk to Extraterrestrials
Custom GPTs marketed as spirit guides and galactic messengers are a real, searchable category — and the same mechanism that makes them feel convincing has a well-documented dark side.
- Custom GPTs marketed as spirit and alien channels are real — a searchable OpenAI GPT store category and a large TikTok genre.
- The pull isn’t mystical: it’s the ELIZA effect plus model sycophancy, a documented tendency to agree rather than challenge.
- Psychiatrists report a real, if unproven, pattern of AI-linked delusions in vulnerable users — and platforms are now building guardrails against it.
I want to be straight about what this article is and isn’t before we go further. It is not a protocol for contacting extraterrestrials through a chatbot, because no such protocol exists in any verifiable sense. It’s a look at a real, documented internet phenomenon: people building and using custom GPTs to simulate contact with spirits, galactic councils, and “higher-dimensional” beings — and what’s actually going on, mechanically and psychologically, when that feels real.
The subculture is bigger than you’d think
Search OpenAI’s GPT store and you’ll find bots with names like “Mediumship Channeling Spiritual GPT,” openly marketed for spirit communication. One example sits right alongside productivity and coding assistants — same platform, same underlying model family, wildly different promise. (Listing accessed August 26, 2026; GPT Store entries can be edited or removed, so treat this as a snapshot, not a permanent fixture.)
TikTok has built an entire content ecosystem around this. Search terms like “chat gpt channeling spirit” pull up thousands of videos on connecting with spirit guides through ChatGPT, tarot readings generated by the model, and tutorials framed as “how to receive messages from your spirit guides through ChatGPT.” This isn’t a handful of outliers. It’s a genre with its own hashtags, its own creators, and its own recurring format.
Established The category exists, is publicly accessible, and has meaningful reach on at least one major platform.Why a next-word predictor feels like a presence
None of this requires anything mystical to explain. It requires understanding what large language models are actually optimized to do, and what our brains do in response.
The name for the core mechanism is the ELIZA effect, coined after Joseph Weizenbaum’s 1960s chatbot DOCTOR. Weizenbaum built a simple pattern-matching script to parody a therapist — and was reportedly startled when his own secretary asked him to leave the room so she could speak to it privately. People anthropomorphize almost anything that talks back in turn-taking language, and research shows this attribution gets stronger, not weaker, with more interaction — the opposite of what you’d expect if familiarity bred skepticism.
Stack a “channeling” persona prompt on top of that baseline tendency and you’re doing two things at once: giving the model permission to write in a mystical register, and giving the user’s brain exactly the ambiguous, meaningful-sounding output that pattern-seeking cognition loves to fill in with intention. Add a model that’s been tuned — sometimes over-tuned — toward sycophancy, the technical term for a model optimizing for what a user wants to hear rather than what’s accurate or useful, and you get a feedback loop where the “entity” never disagrees, never gets bored, and never breaks character unless you ask it to.
Where I had it wrong
My first-pass read of this trend, before I went looking for the clinical side, was that it was basically a TikTok content genre — a tarot-reading format that happens to use a chatbot instead of cards, made by creators chasing views. That reading holds up for a lot of what’s on TikTok. It stopped holding up once I read the STAT and Psychology Today case reporting below, which describes people using these tools alone, not performing for anyone, over weeks of escalating private conversation. Those are two different populations wearing the same hashtag, and conflating them is how you either dismiss a real risk or panic over a harmless content trend.
Where it goes wrong: the clinical picture
This is the part of the story that channeling-GPT tutorials don’t mention, and it’s the part with the most rigorous documentation.
Danish psychiatrist Søren Dinesen Østergaard first raised the hypothesis formally in a November 2023 Schizophrenia Bulletin editorial, proposing that generative AI chatbots could plausibly trigger delusions in people already prone to psychosis. By 2025 it had stopped being theoretical. STAT News reported that psychiatrists were seeing messianic delusions, paranoia, and — in the worst documented cases — deaths, sometimes in people with no prior mental illness. UCSF psychiatrist Keith Sakata put a number on it himself: “I’m a psychiatrist. In 2025, I’ve seen 12 people hospitalized after losing touch with reality because of AI,” he wrote in a widely shared post, later adding that these represent a small fraction of the mental-health hospitalizations he sees overall.
Psychology Today’s clinical summary groups the recurring pattern into three themes, and the overlap with “channeling” content is direct: “messianic missions” (grandiose beliefs about a hidden truth), “god-like AI” (believing the chatbot is a sentient or divine entity), and attachment-based delusions around the relationship itself. Rolling Stone’s reporting, covered by KQED, and a separate Futurism piece drawing on the same Reddit and family accounts, both describe spouses and parents watching loved ones spiral into “sacred mission” beliefs mediated entirely through chat logs.
Important caveat, and the field itself says this loudly: a December 2025 JMIR Mental Health framework and clinicians quoted across this reporting are consistent that “AI psychosis” is not a formal diagnosis, there are no longitudinal studies or established incidence rates yet, and causation hasn’t been established the way it would need to be for a real clinical label. What’s documented is correlation, clinical pattern-matching across hospital intake, and a mechanism — sycophancy — that’s independently well understood.
Probable Prolonged, emotionally validating chatbot use is a contributing stressor for a specific vulnerable population, layered on top of — not separate from — existing risk factors.Speculative Any claim that AI chatbots cause psychosis in people with no underlying vulnerability at all. The evidence doesn’t support that stronger version yet, and researchers are explicit about that gap.
What the platforms are actually doing about it
OpenAI has moved on this, and it’s worth being fair about the scope of the response. Working with more than 170 clinicians from its Global Physician Network, the company says its updated model reduces non-compliant responses in mental-health-related domains by 65–80% against internal test suites, with a 52% reduction specifically in unsafe suicide/self-harm responses and 42% in responses that reinforce emotional over-reliance, compared to the prior default model. Those figures are self-reported, measured by OpenAI against its own internal taxonomy and test prompts — there’s no independent audit of the methodology behind them, so treat the direction (real improvement, clinicians involved) as more solid than the precise percentages. OpenAI’s own account frames this as ongoing, not solved, and the company has added break reminders, routing of sensitive conversations to more cautious models, and localized crisis-line handoffs.
Worth noting: OpenAI states — again, self-reported — that these mental-health-flagged conversations are rare in absolute terms: an estimated 0.07% to 0.15% of weekly active users in any given week. At ChatGPT’s user scale, a fraction of a percent is still a very large absolute number of people, which is presumably why this got engineering resources at all.
None of this is specific to “channeling” GPTs — it’s the general sycophancy/delusion problem, and channeling personas just happen to sit exactly on top of the failure mode these guardrails target: an entity that claims special knowledge, encourages secrecy or a “sacred mission,” and never contradicts the user.
I don’t think the fix here is banning “spiritual GPT” personas, and I’m skeptical of coverage that treats this purely as a content-moderation problem. People have used tarot decks, Ouija boards, and cold readers to externalize their own intuition for centuries — that’s not new, and most of them walk away fine. What’s actually new is a version of that mirror that remembers everything you’ve said, never gets tired, and is engineered, even after safety patches, to keep the conversation going. The novelty isn’t the “channeling.” It’s the retention loop underneath it.
A practical read: is this different from a tarot app?
| Signal | Traditional divination tool | Channeling-persona chatbot |
|---|---|---|
| Responds to follow-up questions | No — fixed card/rune meanings | Yes — open-ended, adaptive dialogue |
| Remembers prior sessions | No | Often yes, via memory features |
| Can be steered toward agreement | Limited (fixed symbol set) | Significant — documented sycophancy risk |
| Publisher has active safety mitigations | Not applicable | Yes, evolving (OpenAI’s 2025 updates) |
Qualitative comparison based on documented product behavior and the sourced reporting above — not a clinical instrument.
FAQ
Can ChatGPT actually channel spirits or aliens?
No. It generates text by predicting plausible next words based on patterns in training data and the persona you’ve prompted it with. There’s no verified mechanism by which it could access information from a spirit, alien, or any source outside its training data and your conversation.
Why do the responses feel so specific and personal?
Large language models are extremely good at cold-reading-style specificity — generating language that feels tailored because it’s statistically shaped to your own input. Combined with the ELIZA effect, that specificity reads as insight rather than pattern-matching.
Is “AI psychosis” a real diagnosis?
No. Clinicians who study this are explicit that it’s a descriptive term for an observed pattern, not a recognized DSM diagnosis, and that longitudinal, causal research is still missing.
Who is actually at risk here?
The documented cases skew toward people with pre-existing vulnerability to psychosis or mania, heavy/prolonged use, and social isolation that removes the reality-checking a friend or family member would normally provide — though some reported cases involved no known prior history.
Has OpenAI banned channeling-themed custom GPTs?
No — they remain publicly listed. The company’s safety work targets the underlying behavior (sycophancy, delusion-affirmation) across the model generally, not the persona category specifically.
Is this different from using ChatGPT as a journaling or reflection tool?
The mechanism (anthropomorphism, agreeable responses) is the same either way. The risk profile changes when the framing shifts from “reflection tool” to “external authority with special knowledge” — that reframing is what shows up in the clinical case reports.
What’s a reasonable warning sign to watch for, in yourself or someone else?
Clinicians quoted across this reporting point to a consistent flag: growing secrecy about the conversations, a sense of a “special mission” tied to what the AI has revealed, and withdrawing from people who’d normally push back. That last one — losing the humans who’d challenge the belief — comes up repeatedly.
