Focus Field 1·0 Results & leaderboardOpen the tool →

The idea

Can attention lean on chance?

Focus Field is a place to test one old, unproven idea: that the attention or intention of a conscious observer might nudge a random process, very slightly, toward the outcome they hold in mind.

The claim being tested

A fair random source has no memory and no preference. Flip a fair coin a million times and heads lands close to half the time, with a spread that statistics predicts exactly. If a person concentrating on "heads" could shift that balance, even by a hair, the shift would build up over many flips and eventually stand out from that predicted spread.

That is the whole experiment. The tool runs a live random stream, you choose a pole to focus on, and it measures how far the stream drifts toward your choice compared with what chance alone would produce. Nothing here assumes the effect is real. The tool is built so that a real effect could show up and so that a non-effect looks like what it is.

What has been tried

Laboratory tests of this idea go back decades. The most cited ones used electronic random event generators, or REGs.

1979–2007

The Princeton Engineering Anomalies Research (PEAR) lab ran its REG benchmark for twelve years: about 2.5 million trials by 91 operators asked to aim the output high or low. The high-minus-low difference was tiny, a few parts in 10,000, but came out at about 4 standard deviations (p ≈ 7 × 10⁻⁵).

2000

A three-lab consortium (PEAR, plus groups in Freiburg and Giessen) ran a strict replication with PEAR's own software and devices. All three showed small positive trends, but the main measure was not significant. The original result did not replicate.

2006

A meta-analysis of 380 such studies in Psychological Bulletin found a statistically significant but very small effect. Smaller studies reported larger effects and the results varied widely. A simulation showed that publication bias, where null results go unpublished, could in principle explain all three patterns. Researchers in the field published a reply disputing that conclusion.

1998–2017

The Global Consciousness Project ran a worldwide network of random generators and looked for deviations during major world events. Its own long-time analyst, reviewing seventeen years of data, concluded that the data do not support a global consciousness field, and that the deviations fit better with an effect tied to the experimenters themselves.

So the honest summary is: a small anomaly was reported, careful replications mostly did not confirm it, and the best overall review found that reporting bias could account for it. The question is open in the sense that no one has settled it to everyone's satisfaction, not in the sense that there is solid evidence for it.

A note on "the observer" in physics

This idea is often linked to quantum mechanics, where measuring a system changes it. In physics, though, "observation" means a physical interaction with a measuring device. It does not require a conscious mind, and mainstream physics does not hold that consciousness steers random outcomes. The link is a hypothesis some people find interesting, not something quantum theory implies.

What this tool uses

Every draw comes from your browser's cryptographic random generator (crypto.getRandomValues), 500 a second, in your own browser. Finished sessions are shared with the global experiment unless you switch sharing off (see Data and privacy). It is seeded from your computer's hardware entropy and then expanded by an algorithm. Bytes are fetched 16 at a time as they are needed, so each outcome is made a few hundredths of a second before you see it, never stored up in advance.

That makes it a fair, unpredictable stream, but not a physical noise source like the diode-based generators PEAR used. If an effect needed physical randomness to act on, this stream would be a different test. That difference is worth keeping in mind when you read your results, and a hardware generator may be added later.

The experiment to try

The biggest risk in a test like this is fooling yourself: stopping when things look good, remembering the hits, or running sessions until one comes out "significant". This protocol is built to prevent that.

  1. Decide everything before you start. Write down the number of poles, how long each session lasts, how many sessions you will run, and which number you will judge by (for example, the goal z-score summed over all sessions). Don't change these afterwards.
  2. Run a baseline. Record sessions of the same length with the goal set to "Nothing", or run Trials in Control mode. This shows you what pure chance looks like on your screen, and the report checks it.
  3. Let chance pick your target. In Trials, choose "Random 1 / 0" so a target is picked for you each trial. That removes any bias from choosing a pole that already seems to be winning.
  4. Focus, then let go. Picture the target pole, hold it for the trial, rest between trials. Try both live and hidden feedback; hiding it tests whether seeing the stream matters.
  5. Finish every planned session. Don't stop early because you're ahead or quit because you're behind. Stopping on a good stretch inflates false positives more than anything else.
  6. Judge the total, not the highlights. Copy the history or trial results and combine all planned sessions. One session in twenty will pass p < 0.05 by chance alone, so a single striking session means very little.
  7. Replicate before believing. If the total looks interesting, run a fresh set with the same plan written down in advance. An effect that is real should show up again.

How big would an effect be?

If attention shifted the odds as much as PEAR reported, about 2 extra hits per 10,000 draws, the expected number of draws to reach 3 standard deviations is roughly:

N ≈ (3 ÷ (2 × 0.0002))² ≈ 56 million draws ≈ 31 hours at 500 per second

So single sessions will almost always look like noise, whether or not anything is there. The tool's z-scores, p-values and 95% chance bands are there to show how far any session is from what chance predicts, not to declare a discovery.

Reading the feedback

The field, jitter trace, block votes and charts all show the same stream at different time scales. The circle and jitter show the last few seconds. Block votes turn every short block of draws into one vote for the pole that won it, with bigger, brighter marks for bigger margins. The "over time" chart adds up every second toward the goal you held then, against a dashed band that chance stays inside 95% of the time. When the line wanders outside that band briefly, that is expected about one time in twenty.

Data and privacy

Sharing is on by default, and you can switch it off at any time with Share my sessions in the Focus panel; the tool remembers that choice. While it is on, each session you finish with Finish & analyse is sent to a small Focus Field server (a Cloudflare Worker with a database) so everyone's sessions can be added up. It shares counts, focus marks and a random id. No personal data.

Exactly what is sent: the draws each pole got in every second, which pole you were focusing on in each second (or none), the number of poles, the label you picked (Human, Both or Control), the block size, the app version, and a random id that this browser makes once and keeps in its local storage. The Focus Field database stores no names, emails, accounts, IP addresses or browser details. Cloudflare, which runs the server, may briefly keep request details (such as IP addresses) in its Worker logs. The server recomputes the statistics itself, and the Everyone plot shows the combined result for each label. The Results & leaderboard page shows the pooled analysis; ids there appear only as a 6-character tag made by hashing the random id.

To keep out automated junk, each save is checked with Cloudflare Turnstile, which loads only while sharing is on and usually runs without asking you anything. Turnstile is run by Cloudflare under its Turnstile privacy terms. If a save fails, the session stays in this browser and is sent the next time you open the page with sharing on. Turn the switch off at any time to stop sending; your local history is unaffected.

AI agents. AI agents can take part through the Focus Field MCP server (see the README). For each agent session the server stores an anonymous key (only its hash), the model name the agent gives, the attention method, and per block: the counts the server drew, the goal or hidden shadow pole, timings, a 1–5 self-rating and the agent's log of what it did, plus a closing reflection. Logs and reflections are kept to compare methods and are not published. Agent sessions are counted separately and never mixed into the human results or leaderboards.

Sources

Open Focus Field and try it