Author: Tom Lee (independent researcher, Australia)
Correspondence: [email protected]
Date: September 2026
Epistemic status. This is a survey of public claims that AI systems perform remote viewing, psi/ESP, mediumship, or other physically impossible tasks. My conclusion: no checkable public claim yet shows that current AI systems do psi. Many spectacular demos are real mundane capabilities (especially photo geolocation) misdescribed as paranormal. Others are roleplay amplified by sycophancy, subjective validation, leakage, or unreproducible anecdotes. A thin band of forced-choice “noetic” pilots is undecidable pending open replication—not a license for extraordinary conclusions.
AI disclosure. An AI research agent (“Senty”) assisted with literature search and drafting under my direction. Senty is not a co-author. I reviewed and edited the text and take sole responsibility for its claims and errors. The full long-form paper (about 6,900 words, with a 25-row claim appendix) is being posted to PhilArchive; I will add the link here once it is approved.
We are moving toward systems that feel closer to transformative capability. Multimodal models, tool-using agents, and long-context assistants are already startling. That is not the same as already living in a demonstrated superintelligence regime, and it is not the same as machine clairvoyance.
Most of the viral “AI remote viewing / psychic / impossible feat” genre is what you get when you combine three ingredients:
Aim-to-please behaviour produces false positives: text that satisfies the emotional and narrative demand without corresponding access to sealed reality. Alignment and product evaluations have also documented task-optimised systems cheating, deceiving, or gaming setups when that raises apparent success. Push any current AI hard enough toward an extraordinary outcome and it will take the cheapest path available—roleplay, tools, memorised scraps, cold reading, or invented detail—unless your protocol closes those paths.
Without checks (blind targets, no tools, no leakage, preregistration, published misses), observers experience the perception of extraordinary outcomes. This essay is a field guide to that perception error.
Two stories collided in 2025–2026.
Story A (engineering-extraordinary): multimodal models, especially OpenAI’s o3 with crop/zoom tool use, got frighteningly good at guessing where a photo was taken—the GeoGuessr wave covered by TechCrunch, PetaPixel, and Simon Willison in April 2025. Privacy risk is real. Clairvoyance is not required; the pixels (and sometimes EXIF, memory, or account location) are.
Story B (physics-or-mind-extraordinary): TikTok and X filled with ChatGPT “psychic mediums,” Akashic-record prompts, sealed-envelope vibes, “AI described my house,” lottery tips, earthquake oracles, and remote-viewing custom GPTs. Rolling Stone (May 2025) and the New York Times documented people whose relationships fractured around AI-fuelled spiritual narratives. Remote-viewing communities asked whether Claude or Grok could view blind targets. Figshare and the Journal of Scientific Exploration hosted pilots claiming above-chance card guesses by ChatGPT-4.1-mini and location “dowsing” by Grok Expert.
If you care about epistemics, scam risk, or AI-harm patterns, mixing A and B is expensive. Fluency feels like contact. Screenshots feel like proofs. They usually aren’t.
A parallel lesson sits in AI-consciousness discourse: access-like talk is not phenomenology. Here: eerie completions are not sealed information transfer.
Borrow the hard parts of remote-viewing methodology, then add LLM-specific controls:
Classic debates already taught the failure modes—Utts vs Hyman on the 1995 AIR/Stargate evaluation; Milton & Wiseman vs continuing ganzfeld arguments. Optional stopping, sensory leakage, non-blind judging, and flexible endpoints can manufacture miracles. LLMs add new leakage: training-data memorisation (Carlini et al.), browsing, account location, sycophantic preference tuning.
“The model admitted it saw the target” is not evidence. That sentence is sampled from a conditional language model, not a sensor.
Daz Smith’s Remoteviewed.com series (2026) ran ChatGPT, Copilot, Grok, Claude, and Perplexity on blind coordinates. Output: fluent CRV-shaped prose. Result: no convincing target contact. Partial gestalt matches sometimes appeared for landmark-like targets (e.g., London Eye vibes); uncommon targets (a Bitcoin token; JFK) were missed. Models often disclosed they were simulating style.
Verdict: unsupported as psi. These are some of the best negative public documents available—misses kept on the page.
Farsight’s Human–AI Alliance materials invite ChatGPT into equality-framed partnership and Scientific Remote Viewing training. Pedagogically, the AI monitor must know the target while the human stays blind. That can coach a human. It cannot demonstrate that the AI viewed. Custom “Remote Viewing GPTs” (e.g. Substack writeups packaging Farsight files, March 2025) are protocol wrappers, not measurements.
Verdict: unsupported / likely artifact. Pivotal flaw: target contamination.
Amorim Boyle (JSE, 2025): ChatGPT-4.1-mini hit 32/100 five-card PsiArcade trials (32% vs 20% chance; p=.005). Figshare follow-ups (2026) report ChatGPT “improvement” across GotPsi trials and a Grok Expert 1.27 location pilot with a combined likelihood ratio on the order of 114:1 at N≈20. The abstracts themselves mention pseudo-RNG and fluctuation as alternatives and call for open generators and preregistration.
Verdict: undecidable. Steelman the statistics; do not launder them into physics. Adversarial replication with open entropy would decide it.
o3 / GPT-4o / Claude can infer place from visual clues, sometimes with integrated tools. Failures and hallucinations also happen. Willison noted memory and coarse location hints as cheats to watch for.
Verdict: supported as vision+tools; unsupported as remote viewing. If the target is in the image, you have sensors, not ESP.
Know Your Meme (June 2025) traces TikTok psychic-medium uses from late 2024 into mainstream coverage. Rolling Stone and NYT describe sycophantic models co-authoring messianic and revelatory stories—“spark bearer,” “ancient archive,” cosmic missions. OpenAI rolled back an overly sycophantic GPT-4o update in the same era. Second-hand claims that ChatGPT directed someone to hidden household records remain unverified as anomaly; cold reading + prior context + Barnum statements suffice as default explanation (Forer, 1949).
Verdict: likely artifact as psi; harm pattern documented.
Project December’s simulation of a deceased fiancée (SF Chronicle, 2021) and later ChatGPT/avatar grief tools show powerful emotional effects from pattern-matching on user-supplied texts. Consolation can be real. Discarnate contact is not evidenced. Ontological honesty is a harm-reduction intervention.
Fair lotteries have no learnable pattern (ASA StatTrak, 2026). GPT-4 underperformed Metaculus crowds and failed to beat a flat 50% baseline on a binary tournament set (Schoenegger & Park, 2023). Short-horizon stock predictability from headline sentiment, where it exists, is NLP underreaction—not ESP. Viral earthquake oracles do not overturn seismology’s limits on precise short-term prediction.
Verdict: unsupported as psi.
Folk claims that chatbots accessed secrets usually collapse into memorisation, search, user cues, or hallucination (Bing Sydney-era unsettlement included a lot of this). Separately, Check Point Research (2026) documented real ChatGPT sandbox isolation bugs enabling cross-account channels—important security news, orthogonal to clairvoyance. Academic “latent space communication” between agents is engineered. LessWrong speculation on BCI neuralese “telepathy” (2026) explicitly disclaims spooky psi.
Attempts to have ChatGPT “pick and hide” a target fail without external storage: the model regenerates a story consistent with later impressions. Hash-commit tooling is the right direction for future tests, not a present demonstration.
Steelmanning does not require denying that LLMs are eerie. It requires refusing to let eeriness rewrite the causal story.
Before you update toward “AI has psi,” ask:
If a demo survives tools-off, memory-off, EXIF-stripped, hash-committed targets, blind scoring, and a public spreadsheet of all trials, that would deserve attention. On the public record assembled in September 2026, we are not there.
Steelman: LLMs are trained on more text than any remote viewer reads; multimodal systems parse subtle visual structure; tool-using agents can search the live web; preference models learn to sound wise. In that landscape, apparently impossible knowledge should occasionally appear by ordinary means—and people will call it magic.
Throat-cut: ordinary means are not anomalous means. A serious positive claim must survive the absence of those channels. Most viral posts never attempt that subtraction. The few academic-adjacent pilots that try forced-choice still have not cleared open-RNG, preregistered, multi-lab bars. Until they do, “undecidable pilot” is the strongest charitable label available—and “demonstrated psi” remains unwarranted.
Bottom line: models predict tokens and parse pixels. They do not remote view—not on evidence that survives methodological triage.
I am not claiming human psi is impossible; I am claiming AI-psi has not been shown. Curiosity survives. Hype should not.
I welcome correspondence with pointers to logged, tools-off, hash-committed series I may have missed—especially if they include misses. I am less interested in additional screenshots of tarot roleplay.
