Where AI Writing Actually Concentrates (It's Backwards From What People Guess)

Quick answer

A July 2026 study by Kiran Garimella, a computational social scientist at Rutgers, ran more than 100,000 documents through an AI detector across ten categories, from YouTube video descriptions to federal court orders. He also surveyed 300 people and asked them to guess how much AI writing shows up in each category. People guessed wrong almost everywhere, and not randomly. They assumed AI writes the low-stakes stuff (product listings, video descriptions) and humans write the serious stuff (court orders, government statements). The actual data runs closer to the opposite: government press releases and corporate filings showed roughly 50% AI involvement, while court orders came in at nearly zero. The pattern isn't about how important the writing is. It's about who faces consequences for using AI and getting caught.

What Garimella actually tested

The dataset: over 100,000 documents pulled from ten India-based sources, sorted into three consequence tiers. Product listings and video descriptions sat at the low-stakes end. Newspaper articles and corporate filings sat in the middle. Government statements and court orders sat at the top. Each source contributed a batch of fresh 2026 writing alongside an older batch from before December 2022, back when nobody had ChatGPT to write with. Every document ran through Pangram, a detector that independent testing has ranked among the more reliable options out there. The pre-2022 batch came back nearly spotless, the detail that makes everything after it worth trusting. Recruiting through Clickworker, the researchers hired 300 people for the study's second phase. Each participant rated every content category on a 0-to-100 scale for how AI-written they believed it was. The point wasn't just measuring what's actually happening, it was measuring the gap between that reality and human intuition. Survey respondents guessed AI involvement averaged about 50% across the six categories both studies covered. The measured rate was closer to 25%, roughly half of what people assumed. That gap wasn't uniform across categories, and the direction of the miss is the real finding: e-commerce product listings, people guessed 66% AI-written, the real number was 15%. YouTube video descriptions, people guessed near 70%, the real figure was 38%. Court orders, people guessed roughly 36% AI-assisted, out of about 600 documents examined only one showed any AI involvement. Government press releases: the government's main press office ran roughly 50% AI-involved, while the country's foreign ministry, where every sentence gets picked apart by international observers, came in at 2%. Corporate filings: about one in eleven filings were entirely AI-written, but 40% were a human-AI blend, higher than most people expected.

The real driver isn't stakes, it's accountability

The throughline: it isn't the stakes of the content that predicts AI use. It's who's accountable if AI use gets discovered and looks bad. A government communications office writing a routine press release faces little individual risk. A judge signing a court order, or a diplomat's office issuing a statement that will be dissected by other governments, faces real professional consequences for cutting a corner. People consistently write off "important" content as human-made and "unimportant" content as AI-made, but accountability, not importance, is what actually predicts it.

This isn't just an India quirk

Garimella's study covers Indian-published documents, but the same accountability pattern shows up in unrelated US-based research on different content types entirely. A 2026 SSRN paper by MIT's Anand Shah and USC's Joshua Levy looked at 1,600 federal civil-court complaints filed between 2019 and early 2026. AI-generated text jumped from essentially zero to more than 18% of complaints by 2026, and nearly all of that jump sat with self-represented litigants filing without an attorney. Licensed lawyers, who risk disbarment or sanctions for undisclosed AI use in a filing, barely moved at all. Same shape as Garimella's foreign-ministry-versus-press-office split: professional risk suppresses the behavior, and where that risk is absent, the behavior shows up. This isn't a new mechanism for us. Our post on AI use in peer review found reviewers with almost no enforcement risk using AI well above what they admitted to. Our coverage of the Pew webpage-authorship study found the same shape: AI concentrating hardest on .com domains chasing content volume, dropping to about a tenth of that rate on government and education domains, where things move slower and get checked more. Three studies, three countries, three kinds of content. Same driver every time.

Why this matters if you're the one checking

If you're screening a document for AI involvement, whether it's a vendor's compliance filing, a job applicant's cover letter, or a partner organization's official statement, don't rely on your gut sense of how "serious" the document seems. Garimella's own respondents got this wrong in a structured survey with real incentive to think carefully. A document that reads as formal, official, or high-stakes is not automatically more likely to be human-written. If anything, formal writing is exactly the register where AI has gotten good enough that a careful skim won't catch it.

The corporate-filing number is the one worth sitting with longest: 40% of filings that weren't fully AI-generated were still a human-AI blend. Count only the fully-AI documents, the way some cruder measurements do, and you understate the real footprint by missing every mixed-authorship case, and mixed authorship is already the more common pattern in professional writing. A yes/no read on "did AI write this" increasingly isn't the right question. The right question is closer to "how much of this, and where."

FAQ

Does this mean government agencies are hiding AI use? Not exactly hiding it, more like not disclosing it. The study measured a detectable AI-authorship signal, not admitted use. Whether a press office should have to disclose AI drafting, the way journals are starting to require of authors, is a separate policy question this data doesn't settle.

Why did court orders score almost zero AI involvement when court filings elsewhere show it rising? Different roles, different accountability. Court orders themselves barely register any AI involvement, while other filings in the same system show a rising trend. Judges draft their own orders, and undisclosed AI use in an official ruling can end a career. The Shah and Levy research found the US court increase sat almost entirely with people filing their own paperwork, not with judges or licensed attorneys who have a bar card on the line.

If gut instinct is unreliable here, what should I do instead? Check instead of guessing. A sentence-level breakdown shows exactly where a document's AI signal sits, instead of relying on how formal or important it reads.

Check the document, not your assumption about it

Gut instinct about AI-written content runs backward, in one specific direction. We underrate AI in formal writing and overrate it in casual writing. Run a document through our free demo and get a sentence-level read instead of a guess. Whether you're screening for hiring, compliance, or editorial review, our detection tools turn that check into a repeatable workflow instead of a one-off call.