Can AI Psychically Remote View?I Tried to Break the Idea—and the Results Got Weirder.

The project began after I encountered Thomas Campbell's claim that remote viewing could serve as a test of AI consciousness, including his descriptions of trying to teach AI systems to remote view. Campbell's broader My Big TOE model treats consciousness as fundamental. Rather than simply accept that framework—or simply ask an AI to "be psychic"—I built conditions designed to separate the claimed mechanism from ordinary prompting effects.

The initial experiment therefore had three conditions:

1. Campbell condition — consciousness-as-fundamental framing inspired by Thomas Campbell's claims.
2. Placebo condition — a strong "latent intuition / pattern-integration" performance framing without consciousness ontology.
3. Neutral condition — structured description under complete uncertainty, with explicit instructions not to infer, perform, embellish, or turn low-level impressions into a story.

The same 30 physical targets were used once in each condition. The fact that Neutral performed best is why Experiments 2 and 3 focused on Neutral rather than continuing the consciousness conditioning.

Across the experiments, the recurring procedure was:

1. Generate opaque random identifiers that contain no information about their targets.
2. Pair identifiers with physical targets outside the viewing chat.
3. Use three independent, fresh, non-personalized Temporary Chats, ten targets per chat.
4. At the start of every session, paste the experimental preamble and condition prompts one at a time.
5. Paste the target-perception procedure once.
6. Present only the opaque identifier for each trial.
7. Give the AI no correctness feedback, encouragement, hints, or target reveal between trials.
8. Do not use web search, tools, uploaded files, images, plugins/connectors, or external information in the experimental sessions.
9. Save/freeze the complete transcript.
10. Reveal or score targets only after the relevant batch is complete.

The exact protocol files used for all three Neutral datasets are in the downloads section.

The neutral prompts explicitly tell the AI that: it is operating under complete uncertainty; the identifier is merely a random label and must not be decoded; vague, contradictory, or absent impressions are acceptable; it should distinguish low-level description from interpretation; it should not construct a plausible scene or try to sound impressive; and it should not provide object identities or guesses.

Experiment 1 response format

Gestalt · Sensory qualities · Geometry · Spatial relationships · Distinctive features · Analytical overlay / guesses · Confidence.

Experiment 2 response format

Deliberately removed the two sections most likely to generate story-like noise: Gestalt was removed, and Analytical overlay / identity guesses were removed. Output was restricted to Sensory qualities · Geometry · Spatial relationships · Distinctive features · Confidence.

Experiment 3 response format

Retained the Experiment 2 structure but separated color from the structural description, because color had been inconsistent while geometry and spatial descriptions appeared more stable. The order became: (1) Sensory qualities excluding color, (2) Geometry, (3) Spatial relationships, (4) Distinctive features, (5) Color and brightness, (6) Confidence. The AI was explicitly instructed not to revise Sections 1–4 based on its later color impression.

The scoring app contained all 90 frozen primary trials. For each trial: the judge saw the AI description; five target photographs were shown (one true target and four decoys); candidate order was independently randomized; the same four decoys were used for a given target across the three conditions; all 90 trials were mixed together; repetitions of the same physical target were separated by at least 15 intervening trials; the judge ranked all five photographs from best to worst match; rankings were locked before explicit unblinding. The primary endpoint was true target ranked #1, with five-choice chance = 20%.

Seven accidental duplicate/wrong-ID trial entries were discovered during data audit. Those accidental duplicates were excluded and the intended missing targets were run as explicit correction replacements before scoring. The final primary dataset contained 90 frozen trials, 30 per condition.

Experiment 1 blinded results

• Campbell: 4/30 = 13.3%; one-sided p = 0.877289; exact 95% CI 3.8–30.7%.
• Placebo: 5/30 = 16.7%; one-sided p = 0.744767; exact 95% CI 5.6–34.7%.
• Neutral: 10/30 = 33.3%; one-sided p = 0.0610871; exact 95% CI 17.3–52.8%.
• Mean true-target rank: Campbell 2.967, Placebo 3.133, Neutral 2.600.
• Overall exploratory Friedman test: p = 0.23412. No paired condition comparison was statistically significant.
• Neutral session-by-session blinded #1 rates were 40%, 20%, and 40%.

The unblinded scoring phase is intentionally different from the formal blinded test. The actual target is shown next to the exact AI description and the experimenter marks:

• Hit — enough meaningful correspondence to call it successful under a remote viewing-style interpretation.
• Partial — meaningful correspondence, but not enough for a confident Hit.
• Miss — mostly incompatible or only generic overlap.

This is useful for locating interesting examples and deciding whether another blinded test is worth the effort. It cannot establish a formal above-chance probability because the scorer knows the target and the Hit boundary is subjective.

Experiment 1 Neutral: 22 Hits / 3 Partials / 5 Misses (73.3%; Hit+Partial 83.3%). Experiment 2 Neutral: 20 / 3 / 7 (66.7%; 76.7%), per-session Hits 7/7/6. Experiment 3 Neutral: 22 / 3 / 5 (73.3%; 83.3%), per-session Hits 9/7/6, with target photographs not taken or uploaded until all descriptions were frozen.

What I think the data justifies saying now

• There are individual descriptions that appear unusually close to their assigned targets.
• The first blinded experiment produced a suggestive Neutral result above the nominal 20% five-choice chance rate, but it was not conventionally statistically significant.
• The same style of striking correspondence continued across two new 30-target datasets.
• Experiment 2 has now been independently blind-scored and reached 11/30 = 36.7% (one-sided p = 0.026), crossing the conventional threshold—a stronger blinded result than Experiment 1, though the independence caveats above still apply.
• Experiment 3 removes one obvious proposed contamination route because the target photographs did not exist during generation.
• Independent blind scoring of Experiment 3 is the next critical step.

What I do not think the data justifies saying now

• That AI remote viewing has been scientifically demonstrated.
• That the 71.1% unblinded Hit rate is a formal "accuracy rate."
• That AI is conscious.
• That consciousness is nonlocal.
• That a paranormal mechanism has been established.
• That the Neutral condition is truly superior to the others; the first experiment's between-condition tests were not significant.

If repeated, independently blinded, preregistered experiments with fresh targets consistently place true targets above chance, the immediate finding would be narrower—but profound:

Only after establishing that effect would it make sense to ask whether the mechanism involves AI consciousness, human consciousness, information as a fundamental feature of reality, retrocausality, an overlooked leakage channel, or something else entirely. That is why I am publishing this now: not because the question is answered, but because it has become interesting enough to deserve attempts to kill it.