The scientific attempt to measure the world's soul

In 1998 a network of devices scattered around the world began to be assembled for an unusual task: to listen for whether the whole planet felt something at once. It was called the Global Consciousness Project (GCP).¹ It is, in a way, a distant relative of what this Observatory does, which is why it is worth telling carefully what it did, what it said it found and why its critics do not accept it.

The idea

It was coordinated by Roger Nelson, who had been Coordinator of Research at the PEAR laboratory at Princeton University from 1980 to 2002; the GCP grew out of that laboratory's experiments. It was funded by private donations through the Institute of Noetic Sciences.¹

The hypothesis was this: when a great many people share an intense emotion at the same moment, the output of a random number generator drifts slightly away from chance.

What exactly it measured

The network was made of physical random number generators —not programs, but devices that draw their bits from a natural phenomenon— connected over the internet to a server. Once per second each device delivers a "trial": the sum of 200 bits, whose expected value is 100.¹² In 2001 there were about 38 devices;³ by 2004 there were between 65 and 70 around the world.²

The aim was not for a single device to drift on its own, but for the devices to move together more than chance allows. The main statistic, the "network variance", measures that excess of correlation between them.²

Each "event" that is analysed —an attack, an earthquake, a ceremony, a football final— is chosen because it produces "widely shared, deep engagement" among many people. Its duration and the calculation to be applied are entered in a formal registry before anyone looks at the data.¹²

What its authors claim

After some 17 years and nearly 500 events, its authors say the cumulative result departs from chance by about seven standard deviations,⁴ a probability of chance of the order of 10⁻¹³ by their own calculations.² Nelson himself describes it as a "small but persistent" effect.⁵

September 11: where the problem showed

The most famous case is the attacks of 11 September 2001. Nelson published that the network had drifted from chance for about four hours, starting shortly before the first impact.⁶²

That same year, Edwin May and James Spottiswoode, two independent researchers, repeated the analysis with the data the project publishes openly.³ They found that the formal result was, in their word, "fortuitous": with the four-hour window fixed by the protocol the test reached p = 0.035, just under the usual 0.05 threshold; had the analysis window been a few minutes shorter or 30 minutes longer, the formal test would not have achieved significance.³

They also pointed out that the protocol was unclear at the time: there were two contradictory descriptions, one in the published account and one on the project's website (an inadvertent ambiguity that was later corrected), and that Nelson's preliminary look at the data could have introduced an inadvertent bias in his choice of parameters. The analysis described on the website "showed chance deviations throughout". And the 6,000-to-1 odds computed by another researcher, Dean Radin, were accounted for by a local deviation about three hours before the attacks. They concluded that the data were "consistent with mean chance expectation".³

Nelson replied that testing other windows was "unacceptable post hoc data selection".² In the same issue of the journal, Jeffrey Scargle published a critical review of the analyses by Nelson and by Radin.⁷

The deeper critique: choosing the window afterwards

With random data, any long series contains stretches that look odd: that is what "random" means. If the stretch is chosen after seeing the data, one "significant" is almost always found. That is why the project registers its hypotheses beforehand, and why the discussion about 11 September turned on whether that registry was clear enough. May and Spottiswoode calculated by simulation that a cumulative curve like the project's has roughly a 50% chance of crossing the p = 0.05 envelope somewhere even if there is only chance.³

Try it yourself. This block generates a whole day of pure noise in your browser —there is no hidden effect— and lets you choose the window:

It is a simplified model, not the project's exact statistic. It serves one idea only: the significance of a result depends on when it was decided what to look at.

What Bancel himself concluded

Peter Bancel had analysed the data alongside Nelson for years.⁵ In 2017, after 17 years, he published his assessment in the journal Explore: the cumulative result over nearly 500 events rejects the null hypothesis by seven standard deviations, but the experiment has not found evidence for global consciousness per se; it has provided clear evidence for a goal-oriented effect, of the kind seen in other experiments with random number generators. An alternate interpretation, he writes, is that the result is due to an anomalous effect associated with persons directly engaged with the experiment.⁴

Nelson replied that this conclusion is "premature at best", and that there is structure in the data that Bancel's model cannot explain.⁵ In other words: even inside the project there is no agreement about what the result would mean, if it were real.

Outside the project it is classified as parapsychology. Its critics point out that spikes and fluctuations are to be expected in any random data, and that there is no set time frame for how close a spike has to be to an event.¹ They also acknowledged one virtue: May and Spottiswoode noted that operating in such an open way is "a testimony to the integrity and curiosity of those involved".³

How a claim like this is checked

The lesson is less about consciousness than about method:

• Fix it before looking. The hypothesis, the event and the window are written down before the data is opened.
• Have someone else repeat it with the same data. That is how the 11 September problem was found, and it was possible because the project published its data.
• The result should survive changing the details. If moving the window by half an hour destroys it, it is fragile.
• There should be a theory that says what should happen and what should not. Without one, any result can be read as confirmation.

What we measure, and what we do not

Anima Mundis measures something much more modest and much easier to check: how many people pressed "transmit", and at what UTC hour. There is no hidden signal to interpret: every dot in the Observatory is a transmission someone chose to make. We do not claim that people are connected by a field; we count how many times someone said how they were.

And the world hour, at 00:00 UTC, is not an experiment: it is an appointment. All we are after is for people to coincide.