How we fix mistakes in plain daylight.

In traditional cartography, when a coastal survey discovered an uncharted shoal, the hydrographic office did not throw away the charts or pretend the rocks were never there. They issued an immediate Notice to Mariners with the exact coordinates, so captains could correct their charts before setting sail.

We think any organisation publishing research on artificial intelligence should follow that same maritime tradition. Editing over an error without a note to protect an author's pride leaves everyone else navigating by a flawed map. Putting mistakes on public display is the only way to keep our work honest and dependable over time.

The three kinds of fixes we record

Not every keyboard slip needs a formal public notice. If we fix a typo, tidy a layout glitch, or clarify a sentence without altering its meaning, we update the page directly. But whenever a revision changes the substance of what we published, we record it here.

We log three distinct kinds of corrections:

• Factual and numerical errors. If we misstate a technical measurement, an institutional date, a legal citation, or an author's name, we correct the text and log the previous and updated values.
• Substantive claim revisions. If fresh empirical evidence, a retracted paper, or an outside critique changes our evaluation of a technology or policy, we state the revision clearly.
• Institutional and forecast updates. When an organisation changes its safety posture or when a forecasted event resolves in our Observatory, we record the resolution and keep the historical assessment visible.

Our correction and verification process

When someone reports an error, or when our own audits uncover one, we follow a simple, four-step routine:

1. Acknowledge and inspect. We log the report and check the questioned passage against the cited primary sources and peer-reviewed literature.
2. Independent verification. A team member who did not write the original passage verifies whether the claim holds up under scrutiny.
3. Direct correction with an in-text note. We update the published article or dataset and attach a dated footnote explaining what was revised.
4. Public errata entry. We publish a dated summary below describing the original text, the corrected text, and the reason for the change.

The public corrections log

No corrections have been logged yet for this version of the site, which launched in November 2026. Prior design notes and exploratory records are preserved in our open repository history.

When the first substantive correction occurs, it will appear here with the date, the affected page title, the original passage, the replacement text, and the primary citation supporting the fix.

How to report an error or broken source

If you spot an inaccurate number, a broken citation, or a flawed argument on any of our pages, please let us know. You do not need to be an academic specialist or a software engineer to point out a mistake.

Send an email to contact@collectiveintelligencenetwork.org with the following details:

• The web address of the page where you found the problem.
• The exact sentence, number, or chart you believe is incorrect.
• A link to a primary source, study, or official document showing what the accurate fact should be.

We review every report promptly. If your note leads to a correction, we will thank you in the log entry unless you prefer to remain anonymous.

1. Glenn W. Brier, "Verification of forecasts expressed in terms of probability," Monthly Weather Review 78, no. 1 (1950): 1–3, doi:10.1175/1520-0493(1950)078<0001:VOFEIT>2.0.CO;2.

2. Ezra Karger et al., "ForecastBench: A Dynamic Benchmark of AI Forecasting Capabilities," arXiv preprint (2024), arxiv.org/abs/2409.19839.

3. Arthur Hoarau, "Can we trust our models? Epistemic calibration in second-order classification," arXiv preprint (2026), arxiv.org/abs/2606.10777.