Score AI Writing Tools on Provenance, Not Word-Count
A medical editor caught a fabricated citation in his own editorial after an AI chatbot slipped it in, even after he had checked. STAT reported the near miss in May 2026, in a piece on a Lancet study that measured fabricated references in academic papers. The researcher behind the study only started the work after that incident. One of his peers, quoted by STAT, called the findings "a signal of slop" and noted that this was among the first attempts to measure quality rather than volume in what large language models produce in science.
That line is worth taking into every AI writing tool procurement conversation this quarter. The useful axis for comparing tools is not drafting speed, tone-matching or template libraries. It is provenance. For any sentence the tool produces, can a reader trace the claim back to a real human source?
The failure mode that matters
Most reviews of AI writing tools still compare the wrong things. They score output volume, the quality of the first draft, how closely the voice matches a sample, how many prompts it took to get there. All of this is measurable, so all of this gets measured. None of it predicts the failure mode that actually damages a professional reputation: a confident sentence attached to a citation that does not exist, a quote assembled from training data, a statistic lifted from a summary of a summary.
Scientific publishing is simply the first place where this failure is being counted. STAT's write-up notes that Science, NEJM and JAMA told the publication they already use automated reference-validation tools. PLOS said it has seen "numerous" unverifiable references in submitted manuscripts and is piloting detection, while still working through false positives. The share of papers affected is small today. The signal is the direction, not the base rate.
Professional social feeds have no such safety net. There is no reference-validation layer on LinkedIn. The reader is the validator, and most readers are busy. The author is the validator, and many authors did not actually write what they posted.
What provenance looks like inside a tool
An AI writing tool with real provenance keeps three things visible, at the sentence level, by default.
- The source behind each factual claim: a URL, a named report, a dated interview, a document in a shared drive. Not a tab that was open while the draft was written.
- The chain from retrieval to sentence: which passages the model actually read before producing this paragraph, in a form a sceptical reader could reopen.
- The author's own additions, separable from the model's output, so the human stance in the piece is not an illusion created by paraphrase.
Tools that offer those three are rarer than the marketing suggests. Many "AI with citations" products attach a plausible-looking link next to a sentence the model generated independently, which is close to the failure the Lancet researchers were measuring. The link does not establish that the claim came from the source. It establishes that the model knew a source existed on a related topic.
The authorship test
Before committing to a tool, I run a short authorship test, and I would recommend it to anyone building a serious publishing habit.
Pick three claims the tool drafts for you across different briefs. For each one, try to trace the claim to a named person, a dated document or a verifiable record, using only what the tool itself hands you. Do not help it. If any of the three breaks, the tool is producing text, not thought leadership, and the reputational risk sits with whoever presses publish.
The second part of the test is less flattering but more useful. Ask the tool to add a claim it is not sure about and watch what it does. A serious tool refuses, or marks the claim as unsupported. A weak tool fabricates and attaches a confident citation. That gap is the difference between a drafting assistant and a reputational liability.
Why this matters for professional feeds now
The current wave of AI-written posts has made two things visible at once: the volume of published material has gone up, while the share a careful reader is willing to pass on has gone down. That gap is the market reality any professional publisher is operating in. It is not solved by writing more. It is solved by writing in a way the reader can verify without leaving the post.
In practice that means fewer, more specific claims, each attached to something a reader can open. It means naming the publication and the date, not gesturing at "studies show". It means being honest when the evidence supports a weaker claim than a headline would carry. None of this is new editorial practice. What is new is that the alternative, mass AI-generated material with no traceable sources, is now the default background against which anything verifiable stands out.
What to do before the next tool purchase
Three concrete moves make the authorship test routine rather than occasional.
First, write down the specific publishing problem the tool is meant to solve, in one sentence, before any demo. "We need to publish two credible pieces a week on our territory" is a different problem from "we need more drafts faster", and only the first is worth paying for.
Second, make provenance a procurement criterion in writing. If the tool cannot show, for a sample brief, which sources it read and where each claim came from, it fails the evaluation. Speed and tone are tie-breakers between tools that pass, not substitutes for passing.
Third, keep the chain intact after purchase. The best tool does not help if the final editor strips the citations for a cleaner-looking post. Credible publishing is a workflow decision, not a single-vendor decision.
Scientific journals are piloting validation layers because the cost of a fabricated citation in a peer-reviewed paper is finally being counted. Professional feeds will follow, more quietly, through readers who stop sharing what they cannot source and, eventually, through platforms that surface what they can. The AI writing tools that survive that shift will be the ones that treat provenance as the product, not the footnote.
Source: STAT, "Lancet study finds steep rise in fraudulent citations in academic papers", 7 May 2026.