Passages Flagged as Likely Heavily AI-Written: S&P 500 Earnings-Call Prepared Remarks

## What this measures

The percentage of passages from the prepared remarks of S&P 500 earnings calls that an open AI-text detector scores above a fixed cutoff, over calls in the trailing 90 days. Prepared remarks are the scripted part of a call, before analysts ask questions; the unscripted Q&A is left out.

## Why it matters

Language models can now write a whole passage of a script or a filing. If companies started handing that writing over wholesale, the share of flagged passages would climb well above the level human writing alone produces. This series is a tripwire for that shift. A separate Palanor measure, the words language models overuse (discourse.ai_vocabulary.* and filings.ai_vocabulary.*), tracks a lighter change: vocabulary picked up from edited or assisted text.

## How to read it

The cutoff was set so that 5% of a sample of passages from before ChatGPT (released 30 November 2022) cross it. Human writing alone therefore reads near 5% (across every call in the baseline period the reading was about 4%). Read the series against that level, not against zero: a reading near it means no detectable change. A sustained move well above it means more passages read like model output; on its own it does not show how they were written (see Findings). Each value carries a 95% interval in its metadata.

## How it is made

Calls: S&P 500 members on the reading date (dated index membership), one call per company per calendar quarter, trailing 90 days. The prepared remarks run from the start of the call to the operator's first call for questions; operator turns and speaker names are removed, and calls where that split is unclear or the remarks are under 500 words are left out. Each document is cut into 300-word passages and up to four are picked by a fixed rule (a seeded shuffle keyed on the document), so the same document always gives the same passages. Each passage is scored by an open AI-text detector, TMR (Oxidane/tmr-ai-text-detector on Hugging Face, a RoBERTa classifier, MIT licence), at a pinned model revision (0ceddea903015ef99cbaa040a4d8a216aed9c683). The score is the model's probability that the passage is AI-written, reading the first 512 tokens. A passage is flagged when its score is above the cutoff: 0.9829, the 95th percentile of scores on prepared remarks of S&P 500 calls from July 2021 to June 2022. Every passage counts once.

The detector was chosen in a test on 600 passages from these same calls and 10-Ks, each paired with a rewrite of the whole passage by a current Claude model. TMR told originals from rewrites with an AUROC of 0.91 on calls and 0.89 on 10-Ks. At a cutoff that flags 5% of the human passages, it caught 71% of the call rewrites and 66% of the 10-K rewrites. Another detector tested (an ensemble) failed on 10-K text and is not used. Changing the model revision requires re-running that test and the baseline; the cutoffs only hold for the revision they were fitted on.

## What it does not say

It counts passages flagged as likely heavily AI-written. It does not see light AI editing: a person's text that a model polished, or a model draft a person reworked, mostly scores as human. A flag on a single passage is a probability, not proof, and about one human passage in twenty is flagged anyway. That is why Palanor publishes only the share across hundreds of companies. It never shows a score for any company or document, and nothing here says that any company used AI.

## Findings — 2026-10-17

The first full scoring (2026-10-17) covered 9,237 S&P 500 calls from July 2021 to October 2026, about 900 per half-year. The share of flagged passages was about 4% in the second half of 2021 and the first half of 2022, rose slowly through 2023 and 2024, and reached about 8% in the second half of 2025 and the first half of 2026, before easing in the second half of 2026. The 95% intervals of the early and late readings do not overlap. Three checks: all transcripts come from one provider throughout, so a change of supplier does not explain it; the rise is spread across many companies (the number of companies with a flagged passage rose by about three-quarters), not driven by a few; and the AI in Text Index (discourse.ai_vocabulary.prepared_excess), which counts the words language models overuse, rose and eased over the same period. One caution limits what this says. The detector scores most human prepared remarks close to its upper end (the median passage from 2021 already scores about 0.90), and the whole range of scores drifted upward gradually from 2022, not in a jump. Because the cutoff sits in that crowded upper end, a gradual shift in how remarks are written, by people, by models or both, moves the share a lot. The series shows that more prepared remarks read like model output than before ChatGPT. It cannot say how much text was written by a model. An earlier check on a sample of 60 calls per half-year found no clear change; that sample was too small to see a rise of this size.

## Source

Palanor calculation from earnings-call transcripts received through a licensed data provider. A value is published only when at least 150 companies have a scored call in the window.