Where the four agree, and where they split

If four AIs answer the same question the same way, asking a different one adds nothing. Here we measure how much the four answers overlap.

The overlap is textual. We do not judge which answer is correct.

“The AIs are converging” — we cannot say that yet

When we began, we expected the four to be converging year by year. They are not. Across 244 questions from 2026-02-24 to 2026-08-30, the mean overlap is 14.1%, and its correlation with time is 0.072 — essentially none.

Half a year may simply be too short. Either way, we leave the null result where you can see it.

What decides it is the question, not the month

The monthly averages barely move, but question by question the overlap ranges from 7.2% to 22.6% — nearly fourfold.

Where the four agreed most

Where the four split most

By genre (the differences are small)

How it is measured, and what it cannot tell you

• Only rounds where all four answered are included (answers under 80 characters are dropped).
• Overlap is the agreement of two-character sequences in the answer text. Different wording reads as “split” even when the substance matches.
• ★ Questions that ask for a translation or a quotation score high, because all four copy the same source text. Some of the “agreed” rows above are of that kind. Read the questions, not just the numbers.
• Genre differences are about two points — within sampling error. Please do not read them as real differences.
• The vendors updated their models during this period. These numbers alone cannot separate that from any real change.