When a Lawmaker's Quote Scores 100% AI, What Does That Actually Prove?
By Dusan Boljevic · AI/ML Engineer at TheChecker.AI
Quick answer
On November 2, 2026, Forbes ran 21,848 congressional press releases, more than 12 million words, through the AI detector Pangram. One release from freshman Rep. Everton Blair Jr. of Georgia came back 100% AI-generated, including the quote Blair's office attributed to him directly: "He fought for our country, so now it is time for us to fight for him and bring him home." His chief of staff denied the release was "100% AI-generated" while defending AI as a drafting tool his office uses and edits. That exchange is the real story here, not the percentage. A detector score on an attributed quote doesn't tell you whether the named person said those exact words. It tells you the text matches a statistical pattern common in machine output, which is a narrower claim that still deserves a real answer, not a denial or a shrug.
What Forbes actually measured
Forbes scraped up to 50 recent press releases from every sitting member of Congress, excluded anything from before ChatGPT's November 2022 launch and anything under 200 words, then ran the remaining 21,848 releases through Pangram 4.1. The headline numbers split cleanly by chamber and party: 26% of releases came back AI-written (at least 80% AI text) or Mixed (over 10% AI text) overall. House releases hit that mark at 29%, Senate releases at 12%. Republican offices landed at 29% Mixed-or-AI-written against 22% for Democratic offices, with 12% of Republican press-release text flagged against 8% of Democratic text. Releases from before ChatGPT existed all scored as human-written, the control Forbes used to argue Pangram isn't just flagging normal political-speak at random.
Three days before the congressional story broke, Forbes had already run the same check on the White House: 8,702 releases, three administrations, all the way back to 2020. About one in ten of the second Trump administration's releases landed as AI-written or Mixed overall, though that average hides a steep climb. Isolate just the 2026 releases and the share nearly doubles to 18%. Only ten releases in that whole stretch crossed the stricter, full "AI-written" threshold rather than the looser Mixed category, and every single one came from this year. First was a presidential-health memo in July 2025. A Ukraine peace-talks readout hit the same mark that November. Oddest of all, so did a press release in May 2026 celebrating a repaired fountain near the Capitol. The Biden years and the first Trump term barely register on that scale, unsurprising since most of both predate ChatGPT's spread into everyday government work.
When people get asked about it, the answers split four ways
Two weeks before the Forbes pieces ran, City & State New York hit the same wall from a different angle. Reporters there emailed 29 New York City Council members for a housing story, got 18 responses, and ran every one through Pangram before quoting it. All 18 showed at least partial AI involvement. Seven came back flagged as fully or partly AI-generated; three of those, from Council Members Frank Morano, Mercedes Narcisse, and Shahana Hanif, scored 100%.
What happened next is the part worth studying. Morano admitted it outright: "I sometimes use AI to help me draft or clean up responses, just like I use staff input, notes, and other tools... The positions and opinions I sent you are mine, and I approved the language before sending it." Hanif's spokesperson gave a similar answer, saying the office drafted a statement first and used AI afterward to condense it, with the final text still reflecting the office's actual position. Narcisse gave a third kind of answer: the flagged response came from a departed fellowship staffer, and she said she only uses AI for research into unfamiliar subjects herself. Chris Banks and Farah Louis denied using AI at all, with Louis's team calling the finding a false positive outright.
Four named officials, four different explanations, and every one of them is a plausible account of how a 100% score can show up on a real statement. That spread matters more than any single number in the dataset, because it's the same range of outcomes TheChecker.AI has documented in false-positive cases outside politics: some flags catch exactly what they claim to catch, some catch heavy editing rather than full authorship, and a smaller number are the detector getting it wrong on a short, formulaic piece of text. A press release is short, formulaic writing almost by design, which is exactly the length and genre where a detector's confidence and its actual accuracy can drift apart the most.
The part that makes an attributed quote different from an essay
Student essays and manuscripts raise one question: did a human write this. A press release with a direct quote attached to a named person's name raises a narrower, sharper one: did this person actually say these specific words, or did a model generate a quote and someone decided it was close enough to attribute to them anyway. That's not a hypothetical distinction in the Blair case. The flagged 100% score covered the release's full text, "right down to the quote attributed to him," as Forbes put it. His chief of staff's response didn't deny AI involvement in the drafting process. It denied that the office's output, as a whole, qualifies as "100% AI-generated," while confirming AI serves as "a starting point" that staff then "review and edit." That's a meaningful distinction a detector can't make on its own. A detector can tell you a passage carries a strong AI-text fingerprint. It can't tell you whether a staffer used AI to draft a quote and then checked it with the member by phone, used AI to draft a quote and never checked it with anyone, or pasted a model's output directly into "according to Rep. X" with zero review. All three produce similar detector scores. Only one of them is the kind of ghostwriting-without-oversight that would actually trouble most readers.
A flagged quote-bearing document raises one real question for anyone in communications, journalism, or research: did the named person actually review and approve the specific language before it went out under their name? Accepting or rejecting the raw score skips past that question entirely. Rep. Blair's office gave a direct answer that held up fine against scrutiny, while Council Member Louis's team went the other way and flatly denied AI touched the statement at all, a claim a newsroom could, in principle, keep checking on its own. Both count as legitimate responses. Neither gets proven or disproven by Pangram's percentage alone. That pattern is worth taking seriously rather than brushing off as a quirk of press-release writing, since it lines up with independent research: TheChecker.AI's own work on where AI writing actually concentrates put government and corporate communications near the top of AI involvement across ten document categories, well above what survey respondents themselves guessed, tracking closely with what Forbes found separately in Congress and at the White House. None of this is really new. What changed is that newsrooms now check the raw text of official communications before they quote it, the same scrutiny they started applying to freelance opinion pieces a while back, and a short, formulaic genre like a press release is a harder case for any detector besides, since there isn't much text to establish a reliable pattern. That's exactly why Forbes excluded anything under 200 words and leaned on pre-ChatGPT releases as a human-written baseline. Running a statement through a detector before it ships costs nothing and buys the same early warning Forbes and City & State got after the fact. TheChecker.AI's free demo gives any newsroom, comms shop, or research team that same first pass. It can't tell you whether a named official personally signed off on every word. It can tell you whether the text carries an AI fingerprint strong enough that the follow-up question is worth asking before the quote runs, not after.
Dusan Boljevic
AI/ML Engineer at TheChecker.AI
Dusan Boljevic writes at TheChecker.AI, covering how AI-text detection works and how students, writers and teams can use it responsibly.
Related posts
Why AI Detector Accuracy Has an Expiration Date
A 99% accuracy claim has a shelf life: new models, paraphrasing, and time itself quietly erode it. Here's why — and how to read a score.
Read more
What Turnitin's Own Margin of Error Means for the Palo Alto AI-Cheating Lawsuit
Turnitin's own documentation shows a 76% AI-flag score carries a 15-point margin of error — what that means for the Palo Alto lawsuit.
Read more
Fabricated Citations, Not a Detector Score, Got This AI-Written Paper Retracted
A detector score didn't trigger this retraction. Fabricated citations did. Here's why that matters.
Read more