How this site is made, and where its biases come from

Who makes it

Sidenote is written by an independent editor working together with an AI model (Claude, by Anthropic). The AI drafts entries, searches for sources, reads archive scans and checks quotes; the editor decides what the site is for, reviews, and corrects. Entries drafted by the AI carry the tag provenance: llm and stay marked as drafts until they're reviewed. Sources marked ✓ have been checked against the original document; ⚠ means they haven't been yet.

Bias in the sources

Everything online about this conflict has been fought over, and that includes the reference works. Wikipedia's Israel–Palestine articles have been under special arbitration restrictions for years, and organised editing campaigns from both sides have been documented and sanctioned. That doesn't make Wikipedia useless, but it does make it a pointer, not a source. Here we follow a claim to the document it rests on: the law, the ruling, the dataset, the original newspaper page.

Reports that cite each other count once. A conclusion repeated by five organisations that share their inputs is one finding, not five confirmations (see the citation loop). This applies on both sides.

Old newspapers have biases too. The archive scans on this site are mostly from the New York Times: one outlet, one country's view, and an index that contains its own errors. We use them to show how events were reported at the time, not as the final word on what happened.

Bias in the AI

It recalls; it doesn't look things up. An AI model writes from patterns in its training data. It's usually right about the shape of a story and less reliable about exact numbers, dates and wording. It can also be confidently wrong in a very fluent way.

Its training data mirrors what's been written. A misquote that circulates widely online is more familiar to it than the verbatim original. So the AI reproduces popular versions: the famous wording rather than what was actually said, or a summary's mistake rather than the source's fact.

It surfaces the most-written-about version first. Asked about a topic, an AI reaches first for the arguments that appear most often in its training data, usually the consensus framing and the older, familiar arguments on each side. Newer or less-covered arguments, sometimes the strongest ones their advocates actually use, can stay out of view unless someone looks for them on purpose. So before an argument is stated here, we search for what that side's current leading advocates and legal organisations argue, rather than relying on what comes to mind first.

It tends to agree with whoever it's working with. Language models are known to lean towards the views of the person in the conversation. Our counterweights are rules the site holds itself to: every argument is stated so that its own best advocates would sign it, every argument is examined in both directions, and every claim needs a public source.

The verification record

These numbers are generated automatically from the site's database each time the site is published: they count what is actually marked in the data, not what anyone says was checked. As of 2026-09-24 (database version cvmkh4bv).

Sources cited in entries

Newspaper archive

Examples from the record

Quotes in the 'Where it's been said' sections carry their own ⚠ until checked; those aren't counted here yet.