How Many Expert Pitches Are Written by AI?

What one platform counted

Expert-source platforms are the plumbing of a large share of modern coverage. A journalist posts a request — a dermatologist on a new skincare trend, a CFO on interest rates — and sources reply with a pitch: a short answer they hope will be quoted. HARO, the oldest of these, was closed by Cision and then bought and relaunched by Featured, which is the subject of what happened to HARO. The question this guide asks is how much of what comes back through that plumbing is now written by a machine, and what happens when some of it reaches print.

The prompt was a piece distributed by Stacker on 29 September, “How journalists feel about citing AI-generated PR, according to data”, carried by several US local news sites. Its figures come from Featured, which runs three platforms — Featured, HARO and Connectively — and used the AI detector Pangram to score incoming pitches.

Between June 2025 and July 2026, those three platforms received 1,394,740 pitches in response to 42,392 journalist requests. Pangram flagged 292,895 of them as fully AI-generated: 21.0 per cent. Pitches scored as mixed or AI-assisted were excluded from the comparisons that follow, and the piece does not say how many there were.

The comparisons are stark. Journalists selected roughly one in ten AI-flagged pitches against roughly one in three human-written ones. Out of a pool of more than 2,500 journalists on HARO, 35 per cent chose to screen out pitches flagged as fully AI-generated, an option that is off by default. And accounts that were later banned were far more likely to have submitted a fully AI-generated pitch: 24.7 per cent of banned accounts had at least one, against 6.8 per cent of active accounts, a ratio of 3.6.

Who counted, and why it matters

The byline on this data matters as much as the data. The Stacker piece states that “Featured supplied aggregate data” and that “this story was produced by Pangram Labs and reviewed and distributed by Stacker”. Pangram sells AI detection. Featured sells access to journalists, and its pitch since relaunching HARO has been that its filters keep low-quality, machine-written responses out. Both parties benefit from a finding that AI pitches are common, detectable and unsuccessful.

That does not make the figures wrong. It does put them in the same category as most numbers in this field, which, as how many pitches a journalist gets sets out, are published by companies selling to the people they measure. It also means some things a neutral analyst would report are missing: the number of pitches in the excluded mixed and AI-assisted categories, the definition of “selected”, how selection is recorded, and the detector's error rate on this particular kind of text.

One further detail in the piece is worth noting for accuracy. It says HARO “was officially discontinued in 2024” and revived by Featured in 2025, which is consistent with the account in our HARO guide of Cision closing the Connectively platform and Featured announcing its acquisition on 16 April 2025.

What the arithmetic allows

Because the piece gives totals as well as ratios, a little arithmetic is possible, and it is worth doing before anyone repeats the one-in-three figure.

The 1,394,740 pitches answered 42,392 requests, which is about 33 pitches per request. Take the most generous reading of the human pool: every pitch that was not flagged as fully AI-generated was human, ignoring the excluded mixed and AI-assisted category entirely. That gives at most 1,101,845 human pitches. One in three of those is about 367,000 selections. One in ten of the 292,895 AI-flagged pitches is about 29,000 more. Together that is up to roughly 396,000 selections across 42,392 requests — up to about nine selected sources per request.

The true figure is lower, because the excluded category is not empty, but by how much the piece does not say. Nine sources per request would be a lot for a news story and unremarkable for a roundup article that collects dozens of short expert answers. Without knowing what “selected” means on the platform — shortlisted, contacted, or published — the ratios describe relative performance, which is the useful part, and say little about the absolute chance that any pitch reaches print.

The relative finding survives every reading: on Featured's figures, a pitch the detector calls fully AI-generated is about a third as likely to be picked as one it calls human.

What the detector can and cannot see

Every number above depends on Pangram deciding what is “fully AI-generated”. Pangram's own account of its accuracy is published on its blog. It says it has “an industry-leading 1 in 10,000 false positive rate” overall, and gives rates by domain measured internally — 0.001 per cent on news articles, 0.004 per cent on academic essays, 0.01 per cent on social media question-and-answer text, and higher on short or formulaic text such as recipes, at 0.23 per cent.

The same post gives the caveat that matters here. Pangram says it performs best when “the text is long, in complete sentences, and requires more original input from the writer”, and recommends against screening “very short (e.g. single sentences) responses” and “template based writing”. An expert pitch is often a few sentences long and follows a template the platform supplies. We do not know the length distribution of Featured's pitches and so cannot say how much this matters, but it is the condition under which the vendor itself advises caution.

The error in the other direction is not measured at all in the release. A detector tuned to avoid accusing humans will let some machine writing through, and Pangram's blog says it treats false positives as the higher priority. The 21 per cent should therefore be read as the share the detector was confident about, not as the share written by AI.

One AI expert, retested

The Stacker piece cites one case by name: “Dr. Eleni Nicolaou, a supposed art therapist whose fully AI-generated quotes appeared more than 30 times in publications including Vice, Forbes, and Glamour”. The case was reported by Press Gazette in September. Its investigation found that her profile photograph and much of her output appeared to be AI-generated, that she had no record on the Art Therapy Credentials Board, that her online presence linked back to a paint-by-numbers retailer, and that she could not be verified to exist. The source platform Qwoted removed her profile and a second persona's after being alerted.

Exposure is one thing; what happens to the published quotes afterwards is another, and it is the part that matters to anyone relying on coverage. So on 1 October 2026, about two weeks after Press Gazette's report, we fetched every article URL Press Gazette linked as carrying her quotes — ten in all — and checked each for her name and for any correction or editor's note.

Only one outlet did what a correction policy describes. Tom's Guide's article now opens with “Update 9/18/26: This article has been amended from its original form following our discovery that some of the information it contained was inauthentic. We apologize for the error and have removed the offending information to ensure utmost accuracy.” Her name no longer appears on the page.

Vice's article no longer mentions her either, but the change is silent. The URL still describes painting your walls a colour to calm an anxious dog; the headline now reads “If Your Dog Is Always Anxious, Toys In This Color Can Help”; and we found no editor's note, update line or correction on the page. We cannot see the earlier version, so we cannot say exactly what changed, only that a reader arriving now is not told anything did.

Four pages still quote her with no note of any kind. Glamour's piece on “crying therapy” still describes her as “an art therapist and creative wellness expert” who sweeps clients' cheeks with a paintbrush dipped in cool water. Psychology.org still credits her with a Ph.D. in clinical psychology. AOL still lists her among its expert sources.

The fourth is the most instructive. Yahoo's copy of the Tom's Guide article — same headline, with Tom's Guide named as the provider in its markup — still quotes “Dr. Eleni Nicolaou a clinical psychologist and creative wellness expert”. Her name appears eight times in the page's HTML. Tom's Guide corrected its original on 18 September. The correction did not travel to the syndicated copy. That is the mirror image of what we found for press releases in where a wire release actually ends up: syndicated copies are a separate object with their own life, and changes to the original do not reliably reach them. How outlets handle corrections in general is in how publications handle corrections.

Press Gazette also named a second persona from the same retailer, quoted as a business expert on the website PR On the Go. That page still quoted him when we checked.

What this means for real experts

For a genuine expert pitching through these platforms, the data is good news in a narrow way and bad news in a broad one.

The narrow good news is that, on the only large dataset published, human-written pitches are picked about three times as often as machine-written ones, and a third of journalists on one platform filter machine-written pitches out altogether. Writing your own answers, specifically, from real experience, is not just more honest; on these numbers it works better.

The broad bad news is that the channel is degrading. When a fifth of the replies to a request are machine-generated and some of them are invented people, journalists respond the way Press Gazette reported Pressflow did — restricting access to paid, media-trained experts — or by leaving the open platforms for verification-gated ones. Every such move raises the cost of being quoted for a real expert without a profile that can be verified. Qwoted told Press Gazette it is rolling out video verification for that reason.

That makes verifiability an asset in its own right. A source whose credentials can be checked in a minute — a named employer page, a licensing record, published work, prior coverage — is the kind journalists can still use quickly. How editors weigh sources is covered in how editors decide what to cover, and the risk of AI material in coverage from the search side in does Google penalise AI-generated coverage.

Methodology and limitations

The platform figures are Featured's, as published in the Stacker-distributed piece of 29 September 2026, which we read on 1 October on ABC17 News's site; we had no access to the underlying data. The arithmetic in the third section is ours and uses only the totals and ratios stated there. Pangram's accuracy claims are quoted from its own blog post on false positives, read 1 October 2026; we did not test the detector. The Nicolaou case is taken from Press Gazette's report, read in full through a text-extraction proxy because Press Gazette returned an empty page to our client, and from its newsletter of 21 September. For the retest we extracted every article link from that report, excluded links to social profiles, tools and Press Gazette's own pages, and fetched the remaining ten article URLs, plus the one page quoting the second persona, on 1 October 2026 with curl and a standard desktop browser user-agent string. For each page we recorded the HTTP status, counted occurrences of the name, and searched the text for update lines, editor's notes and corrections. Forbes and Parade returned HTTP 403 to direct requests and no article text through the proxy, and are reported as unread. Five limitations apply. The selection ratios exclude an unstated number of mixed and AI-assisted pitches. “Selected” is undefined. The detector's error rate on short pitches is unknown. Our retest covers ten of the more than 30 quotations Press Gazette counted. And we saw each page only as it is now, so we can say whether a correction is visible, not what was changed.

Conclusion

On the one large dataset published, about a fifth of the pitches sent to journalists through HARO and its sister platforms are confidently machine-written, and they are picked about a third as often as human ones. Those figures come from a platform and a detection vendor that both benefit from them, they exclude an unstated middle category, and they depend on a detector its own maker advises against using on very short or templated text, which is what many pitches are. The relative finding is still the safest one: machine-written answers do worse. The retest adds the part the data leaves out. Two weeks after an AI-generated expert was exposed, four of the ten articles linked as quoting her still did so without a word of correction, one had removed her silently, and the one outlet that corrected its article properly had the correction stop at its own page while Yahoo's syndicated copy carried on quoting her. Exposure does not clean the record. For a real expert the lesson is to write your own answers and make your credentials checkable in a minute; for anyone relying on coverage, it is to remember that a quote in print is not the same as a source that exists.

Frequently Asked Questions

Sources

1. vendor research Pangram Labs for Featured, distributed by Stacker, How journalists feel about citing AI-generated PR, according to data, 29 September 2026, as carried by ABC17 News. 1,394,740 pitches, 42,392 requests, 292,895 flagged; selection and ban ratios. Read 1 October 2026.
2. vendor Pangram, All About False Positives in AI Detectors. Claimed 1-in-10,000 false-positive rate; per-domain rates; advice against screening very short and template-based text. Read 1 October 2026.
3. trade press Press Gazette, Prominent art therapist quoted by Vice, Forbes and others is AI-generated, September 2026. The investigation, the article links we retested, Qwoted's and Pressflow's responses. Read 1 October 2026.
4. trade press Press Gazette newsletter, UK mag giant returns to growth | Widely quoted art therapist is AI creation, 21 September 2026. Read 1 October 2026.
5. live read Tom's Guide, The 15-minute sunlight rule ended my morning energy slump. “Update 9/18/26” correction; name absent. Read 1 October 2026.
6. live read Yahoo, syndicated copy of the same Tom's Guide article. Provider marked as Tom's Guide; name present eight times; no correction. Read 1 October 2026.
7. live read Glamour, crying therapy; Psychology.org, How psychologists change lives; AOL, 5 Low-Pressure Ways to Reconnect With Friends. Each still quoting her, no correction. Read 1 October 2026.
8. live read Vice, If Your Dog Is Always Anxious, Toys In This Color Can Help. Name absent; no correction visible; URL slug refers to painting walls. Read 1 October 2026.
9. blocked source Forbes, two articles by Bryan Robinson, 10 May 2026 and 15 July 2026, and Parade, Psychologists explain the right way to retire. All returned HTTP 403 on 1 October 2026; not read. Netmums' nursery colour article returned 404.