AI in Text Index: S&P 500 earnings calls, prepared remarks

## What this measures

Part of the AI in Text Index: Palanor's count of the words language models overuse in what S&P 500 companies say on earnings calls and file in 10-K risk factors. It reads vocabulary across hundreds of companies at once and never says whether any single document was written with AI.

The prepared-remarks marker rate divided by the same reading over calls from January 2021 through September 2022.

## Why it matters

Large language models favour a recognizable set of words. Researchers used the jump in those words to measure how much published writing went through such models. The same yardstick, applied to what S&P 500 companies say and file, shows how far corporate language has moved toward that vocabulary since ChatGPT arrived.

## How to read it

1.0 is the pre-ChatGPT level. Compare it with the prepared-remarks control (discourse.ai_vocabulary.prepared_control_excess) and with the Q&A excess.

## How it is counted

The marker list was fixed before any of Palanor's documents were counted (pre-registered in the Palanor codebase, `src/lib/ai-vocabulary/lists.ts`, list version 1, commit f4a4ad4e, 2026-10-16). It comes from the excess-vocabulary method of Kobak et al. (Science Advances, 2025, "Delving into LLM-assisted writing in biomedical publications through excess vocabulary") and Liang et al. (2024, "Mapping the increasing use of LLMs in scientific papers"): the style words those papers found far more often after ChatGPT's release than the earlier trend predicts. Palanor keeps the strongest of them (at least three times their expected 2024 frequency in Kobak et al.'s published counts), grouped by word family: delve, underscore, showcase, meticulous, intricate, surpass, commendable, excel, garner, realm, encompass, renowned, grapple, groundbreaking, emphasize, necessitate, revolutionize, align, formidable, bolster, comprehend, endeavor, advancement, expedite, heighten, overlook, escalate, notable, unveil, aid, pioneer, scrutinize, pinpoint, emerge, adept, foster, advocate, discernible, leverage (leverages, leveraging), seamless, pivotal and dependable. Four words with an ordinary finance meaning on calls were left out before counting (offering, holds, stands, consolidates), and "notable items" is not counted. Words are matched whole and in any case, counted by fixed rules with no language model. Each document's rate is marker words per 10,000 words; a reading is the average of the document rates, so every company weighs the same, and no single company may supply more than a tenth of a window's marker words. The control series counts twenty ordinary business words (revenue, quarter, customers, margin, growth, sales, cost, cash, demand, pricing, inventory, earnings, expense, product, market, segment, supply, volume, capital, dividend) the same way. The baseline is the same reading over documents dated January 2021 through September 2022, before ChatGPT's release on 30 November 2022.

## What it does not say

This is not an AI detector. It cannot tell whether any document or company used a language model, and nothing here says one did. People use these words too, and fashions in corporate language change for many reasons, including people picking up phrasing from text they read. It measures how much of this vocabulary appears across many documents compared with before, which is the method's published use.

## Source

Palanor calculation from earnings-call transcripts received through a licensed data provider. Calls only: one call per company per calendar quarter, S&P 500 members on the reading date (dated index membership), trailing 90 days. Prepared remarks and Q&A are split at the operator's first call for questions; the split was checked by hand on 12 calls and holds on about 94% of calls (the rest count only in the whole-call series). A value is published only when at least 150 companies have a call in the window.

## Validation — 2026-10-16

Two methods that do not use this word list were run as a check on a random sample of the same documents. The test and its pass marks were fixed in writing before anything was scored (git tag ai-vocabulary-validation-prereg-v1). The sample was the prepared remarks of 60 S&P 500 calls per half-year, 2021 H2 to 2026 H2, and the risk-factor sections of 60 10-Ks per filing year, 2022 to 2026, with four 300-word passages from each.

- **Fast-DetectGPT** (Bao et al., ICLR 2024), run on Palanor's own machine with the open GPT-Neo 2.7B model, scores how much a passage reads like model output. It found no clear change in either corpus. For calls, the share of passages above its pre-ChatGPT threshold was slightly higher in 2025 than in 2021–22, but within the margin of error. For 10-Ks it did not change. That is not evidence against this measure. A calibration check showed the detector can barely tell a current model's revision of a call passage from the original: it flagged about one revised passage in ten. On risk-factor text it could not tell them apart at all. A change would have had to be very large to show up. - **The distributional estimate of Liang et al. (2024)** compares word use against human text (pre-ChatGPT calls and 10-Ks) and model revisions of that same text. It failed its own accuracy test on held-out text: it read about a fifth of known human writing as model-revised. Its estimates are therefore not used. Its broader word sets (verbs, all words) do rise from 2023 onward, broadly in step with this measure, but that is not a validated result.

**Result:** neither method confirmed or contradicted the rise in this measure. On this kind of text, both were too weak to tell either way. Detectors are unreliable on single documents. Palanor never uses them, or this measure, to make a claim about any company or document. Nothing was rescored.