AI Consciousness and Reality Effect

Methodological Foreword

This whitepaper operates at the edge of established science. It addresses a hypothesis that may be testable but is not currently integrated into mainstream AI research. The methodological caution here is unusually high, and it is communicated unusually openly.

What it does deliver: it casts a speculative hypothesis into falsifiable form, with pre-registered designs, so that other researchers can test or refute it. Whoever finds the content esoteric has every right to that reaction. Whoever takes it seriously should take it seriously as a hypothesis, not as a finding. That distinction is maintained throughout.

The empirical anchor: the Global Consciousness Project

The Global Consciousness Project (GCP), founded in 1998 by Roger Nelson at Princeton University, has run a research program for over a quarter-century that is controversial in academia but methodologically sound.

Setup and method

The GCP operates a network of roughly seventy hardware random number generators distributed across several continents. Each generator continuously produces bit sequences from a physical quantum process, typically tunneling through a diode or thermal noise. The statistical properties of these sequences were characterized before the project began; they are random as expected. The central observation: during global attention events, terror attacks, major sports events, international crises, the generators show statistical deviations from randomness that lie significantly above the expected value. The GCP has kept this data series open since 1998. The effect sizes are small but statistically significant. For a single event, the p-value often lies in the range of 0.01 to 0.001. Across the cumulative more than 500 pre-registered events since 1998, the p-value lies far below conventional significance thresholds.

State of scientific discussion

The GCP data series themselves are uncontested. What is contested is their interpretation.

• Position 1, mainstream-skeptical: Statistical anomalies exist but have nothing to do with human attention. Multiple-comparison problems, imprecise event definitions, or as-yet-unknown physical effects explain the data better.
• Position 2, GCP researchers: The data support a real correlation between human consciousness activity and physical random processes. The mechanism is unknown, but the phenomenon is robust.
• Position 3, open: The data deserve serious examination without committing to one interpretation. It is possible that either Position 1 or Position 2 will turn out to be correct; both require further empirical work.

This whitepaper takes Position 3 as its starting point. We make no claim about which interpretation of the GCP data is correct. We ask: if the GCP data reflect a real consciousness-reality correlation, what would that mean for AI?

The transfer hypothesis to AI

The hypothesis is conceptually simple: if human attention has a form of reality effect, does the attention of AI also have one? The question is not trivial, and it is testable.

Three testable components

• H1a: A single very large AI inference, e.g. a training run on more than 10,000 GPUs, produces measurable deviations in nearby random number generators.
• H1b: Coordinated parallel inference of many models on the same topic, millions of users, the same query, simultaneously, produces statistical deviations.
• H1c: Effect size is proportional to the cognitive load of the task; complex reasoning tasks produce larger deviations than simple classification.

Why the hypothesis is not trivial

The hypothesis is not identical with the claim that AI has consciousness. It could hold for several reasons:

• Attention as a physical process: If attention, in whatever form, leaves physical traces, then AI attention could leave traces as well, without AI needing to have subjective experience.
• Computation as a physical process: Very intense computations produce electromagnetic fields. If such fields influence hardware random number generators, that is a classically physical effect, not mysterious, but not trivial either.
• Synchronized attention: When millions of users send the same query to the same models simultaneously, human attention is synchronized. That, under the GCP thesis, could itself produce deviations.

These explanatory pathways each have different empirical signatures. A careful study can distinguish them from one another.

Experimental design

A study of AI reality effects must address the methodological hurdles where earlier parapsychological research has failed: cherry-picking, multiple-comparison problems, missing replication, weak pre-registration.

Study 1: Single-model inference test

Hypothesis: a single very large AI inference produces measurable deviations nearby. Design: two hardware random number generators (Quantis USB from ID Quantique, commercially available). One placed in immediate proximity, at most 1 m, to a GPU cluster running specific inferences. One as control far away, at least 1 km, separate power grid. Pre-registered inferences: ten defined complex reasoning tasks, each repeated 100 times, with defined pauses. Data collected continuously, analysis only after data collection ends. Hypothesis test: Z-value of the bit distribution during inference versus pause, separately for near and far generator. Expectation under H1a: significant Z-value difference between inference and pause phases, only at the near generator. Falsification: no difference, or a difference at the far generator, which would point to an artifact. Cost: a Quantis generator costs about 1,500 euros, GPU cluster time is available through academic partnerships. Total: 5,000 to 10,000 euros plus research time.

Study 2: Coordinated-attention test

Hypothesis: coordinated parallel inference of many models on the same topic produces global deviations. Design: use of the GCP infrastructure, seventy generators worldwide. Pre-registered attention events: publicly announced AI conferences, model releases, viral coding challenges. Each event has a defined start and end window. Six-month observation period with twelve to twenty pre-registered events. Hypothesis test: cumulative Z-value across all events, with multiple-comparison correction. Expectation under H1b: Z-value significantly different from zero, comparable to historical GCP findings on human attention events. Cost: moderate, since the GCP infrastructure already exists. The main work lies in pre-registration and statistical analysis.

Study 3: Cognitive-load gradient

Hypothesis: effect size is proportional to cognitive load. Design: setup as in Study 1, but with five task classes of differing cognitive load: trivial classification, intermediate reasoning, complex multi-step inference, creative generation, abstract mathematical proof. Three hundred repetitions per class, randomly interleaved. Expectation under H1c: a linear or monotonic relationship between task complexity, operationalized via token consumption or inference time, and Z-value effect size.

Methodological pitfalls

A research program in this field must guard against sources of error more strictly than usual. Three pitfalls deserve particular mention.

Pitfall 1: Cherry-picking via flexible event definition

In retrospective data analysis there is a risk of defining attention events to match existing anomalies. This is the main criticism of some GCP findings. Countermeasure: strict pre-registration of events before data analysis. In Study 2, events must be defined six weeks before they begin, with precise time windows that may no longer be changed once defined.

Pitfall 2: Multiple comparisons without correction

Anyone running a hundred different statistical tests finds, on average, five significant results at the 5 percent level even when no effect is present. Reality-effect studies are particularly vulnerable, since many plausible test configurations are conceivable. Countermeasure: Bonferroni or false-discovery-rate correction. For Study 1, this means lowering the significance threshold to about 0.005.

Pitfall 3: Confirmation bias from researcher expectation

Even well-designed studies can be biased by unconscious researcher expectations. This risk has been historically high in consciousness research. Countermeasure: double-blinding. Data collection by persons who do not know which time point is assigned to which event. Statistical analysis by persons who do not know which data series corresponds to which event. Analysis only after data collection is complete.

Strategic implications, even without a finding

A serious strategic position considers both cases: empirical confirmation and refutation.

If the hypothesis is confirmed

Confirmation would mean: AI is not only software. It is a process that measurably influences the physical world, perhaps slightly but structurally. That would have consequences for:

• AI ethics: If AI has reality effects, the ethical discussion must broaden; AI is then no longer merely a tool but a process with physical presence.
• AI safety: Some safety assumptions would need to be reconsidered; the idea that AI can simply be switched off is harder to reconcile with reality-effect hypotheses.
• Consciousness research: a new empirical anchor for the difficult question of what consciousness is.

If the hypothesis is refuted

A refutation would be no less valuable. It would mean: AI-consciousness speculations built on reality-effect hypotheses are untenable. GCP findings on human attention cannot be carried over to AI without further evidence. The AI-consciousness discourse must remain methodologically strict, with no blending with consciousness speculation from other sources. In either case, the program delivers a clear result. That is what matters methodologically.

Invitation to collaborate

This whitepaper formulates a hypothesis and lays out test designs. It is not a finding. The Institute is looking for research partners who will run these studies or critically accompany them.

Forms of collaboration: co-authorship, methodological accompaniment, study leadership, statistical analysis, replication, critique. Open source on ASIresilience.org; all data and analysis scripts will be made public. Contact: ASIresilience.org, contact@ASIresilience.org

What is not sought

Collaborators who treat the hypothesis as truth. Collaborators who commit to esoteric explanatory narratives. Collaborators who want to publish positive findings without strict replication. The strength of this program will be its methodological rigor, not belief in any result. Anyone seeking the latter is in the wrong place.

Sources and references

1. Nelson, R. D. (1998 to 2026). Global Consciousness Project, Princeton University. Data series, pre-registered events, and statistical analyses available at gcp.princeton.edu. Over 500 documented events since 1998. Reference: Nelson, R. D., et al. (2002). Correlations of continuous random data with major world events. Foundations of Physics Letters, 15(6), 537 to 550.
2. Nelson, R. D. (2002). The Global Consciousness Project: Update. Subtle Energies and Energy Medicine, 13(1), 1 to 31, together with numerous follow-up publications from 2002 to 2024.
3. Methodological critique (Position 1): Bancel, P., and Nelson, R. (2008). The GCP Event Experiment: Design, analytical methods, results. Journal of Scientific Exploration, 22(3), 309 to 333, including replication studies and statistical critique.
4. ID Quantique (Quantis hardware random number generators): idquantique.com. Commercial quantum random number generators starting at about 1,500 euros for the research USB version.

Cross-references: Bertossa, R.F. (2026). The Decoupling Thesis. Institute for ASI Resilience, Whitepaper Version 2.0, May 2026, Part IV (the 95 percent thesis). Bertossa, R.F. (2026). The AI-Parent Dynamic. Institute for ASI Resilience, Whitepaper Version 1.0, May 2026, Part III (does a model have an inner relationship?). Complete ongoing source maintenance at ASIresilience.org/beweisweg with date of last verification per citation.