Likely AI Language Surged Across the Congressional Record

By Jonathan Gillham, Founder and CEO, Originality.ai

Coverage: 2020 through August 2026

Likely AI language rose sharply in public congressional material after 2022. 

In 2026 YTD, 64.9% of estimated Extensions of Remarks and 22.2% of estimated Floor Speaking Turns were classified Likely AI, compared with 5.0% of CRS bill summaries. 

The pattern crossed party lines and was substantially stronger in House Floor Turns than Senate Floor Turns. 

We found two clear disclosures of politicians querying AI, but no disclosure of AI drafting congressional language and no exposed AI residue like the instructions found in our earlier New Brunswick investigation.

Key Findings:

• Extensions of Remarks rose from 1.3% Likely AI in 2020 to 64.9% in 2026 YTD
• Floor Turns rose from 2.1% Likely AI in 2020 to 22.2% in 2026 YTD
• CRS bill summaries changed much less, from 3.3% Likely AI in 2020 to 5.0% in 2026 YTD
• By politician, the estimated share of sampled Floor Turns classified Likely AI ranged from 0% to 46.9% among 153 politicians with at least 10 sampled turns.
• House Floor Turns had roughly three times the Senate Likely AI rate in 2025 and 2026 YTD.
• Only two clear disclosures of politicians querying AI were found; neither disclosed AI drafting congressional language, and no hard AI residue was verified.

Methodology and Sample

• We assembled 115,089 English texts of at least 100 words from GovInfo's Congressional Record and Congress.gov: 5,782 CRS summaries, 84,187 Floor Turns, and 25,120 Extensions of Remarks.
• We analyzed all 5,782 eligible CRS summaries and outcome-blind probability samples of 5,275 Floor Turns and 1,425 Extensions.
• Texts were checked with Originality.ai using a 15% AI Allowance. Scores of 50% or higher were classified Likely AI. Floor Turn and Extension results are survey-weighted estimates; 2026 is a partial year.
• A complete analysis can be run on request to replace the sampled estimates with full-corpus counts.

This investigation began with Politico's reporting on AI-shaped requests reaching the Office of Legislative Counsel. It also follows our New Brunswick investigation, where a politician appeared to read AI drafting instructions aloud.

Finding 1: Extensions of Remarks Rose From 1.3% to 64.9%

Extensions of Remarks are supplemental material House members submit for publication in the Congressional Record. They can include tributes, policy statements, constituent material, and event recognitions. They are not necessarily spoken aloud and may be prepared by a member, staff, constituents, organizations, or other contributors.

The estimated Likely AI rate rose from 1.3% in 2020 and 3.1% in 2022 to 25.9% in 2024, 54.5% in 2025, and 64.9% in 2026 YTD—the sharpest increase in the study.

Finding 2: Floor Turns Rose From 2.1% to 22.2%

Floor Turns are the closest public dataset to words delivered during House and Senate proceedings. We segmented the Congressional Record into member-attributed speaking turns and retained English passages of at least 100 words.

The estimated Likely AI rate stayed near 2% through 2022, then reached 6.6% in 2024, 14.5% in 2025, and 22.2% in 2026 YTD.

Finding 3: The Share of Floor Turns Classified Likely AI Ranged From 0% to 46.9% by Politician

Among 153 politicians with at least 10 sampled Floor Turns, the estimated share of turns classified Likely AI ranged from 0% to 46.9%. Seventy-five had no sampled turns classified Likely AI, while the tenth-highest share was 19.9%.

Finding 4: House Floor Turns Had Roughly Three Times the Senate Likely AI Rate

In 2025, an estimated 19.7% of House Floor Turns were classified Likely AI—1,659 of 8,427—compared with 5.8% of Senate Floor Turns, or 286 of 4,943. 

In 2026 YTD, the House estimate rose to 28.0%—1,539 of 5,503—while the Senate estimate was 9.8%, or 250 of 2,562.

We independently recalculated the chamber-year estimates from the weighted sample. The public record does not identify why the gap exists; speaking formats, office workflows, staff support, and the mix of prepared remarks may all contribute.

Should Politicians Disclose AI Use?

We searched the full eligible corpus for direct AI references, first-person AI-use disclosures, exposed prompts, assistant preambles, and editing instructions. Thirty-seven candidates were manually reviewed.

We found two clear disclosures. Senator Marsha Blackburn said she queried Google's Gemma about herself and described the model fabricating an allegation. Senator John Cornyn said, “I asked my friendly AI assistant,” before presenting material connected to a request about parole programs. Neither senator said AI drafted the congressional language.

We found no AI-drafting disclosure and no hard AI residue in the U.S. corpus. That differs from the New Brunswick investigation, where apparent drafting instructions were read aloud. 

If AI is used to draft or materially shape words entered into an official legislative record, a simple disclosure would let the public distinguish AI-assisted language from member- or staff-written text.

Final Thoughts

Likely AI language is no longer an isolated feature of the Congressional Record. 

Its sharp growth in Extensions of Remarks and Floor Turns—while CRS summaries changed comparatively little—suggests AI is becoming part of how congressional material is prepared. 

This shows that AI is materially shaping words entered into the public record.

Further Reading:

• Politician Reading AI Residue Reveals a Larger AI Pattern
• AI Studies
• AI Detection Accuracy