Rewritten Charity Mission Statements That Read as AI-Written (Quarterly)

## What this measures

The mission and program text of a fixed 1-in-20 panel of filers (by EIN) on Form 990 and 990-EZ, up to 1,500 returns a quarter, is scored by an open AI-text detector; the series is the share scoring above a fixed cutoff, by the quarter the return was filed. Only text that changed since the organization's previous scored return counts, which is where newly written text shows. Detector: Oxidane/tmr-ai-text-detector (MIT), the revision pinned for Palanor's AI-written text share. The cutoff is the score that 5% of returns filed from January 2020 to September 2022 exceed, so those quarters read about 5% by construction.

## Why it matters

It shows how fast text that reads like language-model output spreads in a corpus written by people who care about wording: researchers describing their work, charities describing their mission.

## How to read it

The series starts in 2021 Q2 at 3.2%. The latest reading, for 2026 Q2, is 8.7%. The high is 8.7% (2026 Q2) and the low 2.5% (2021 Q4).

## What it does not say

A high score means text reads like model output, not that a model wrote it: edited, formulaic or translated text can score high too. The detector learned from the output of earlier models, so text from newer models, or model drafts edited by hand, can read as human to it: a falling share can mean the detector recognises less, not that less text is written with AI. Palanor never publishes a score for one paper or one organization.

## Source

Palanor scoring of IRS Form 990 and 990-EZ e-file returns (public domain). Source: Internal Revenue Service.