Here is an uncomfortable pair of facts. More than half of researchers now use AI tools somewhere in their writing. And when audiences find out, they trust the result less. Not because the text is worse, but simply because a machine touched it. Researchers who are honest about their tools pay a price that the quiet users avoid. Call it the honesty penalty.

This is the situation science communication has to work in. Pretending AI is not in the room is no longer an option, and neither is using it carelessly. The interesting question is what a credible middle path looks like.

The quiet shift has already happened

In a study published in Science Advances, Dmitry Kobak and colleagues analysed the vocabulary of more than 15 million biomedical abstracts indexed in PubMed between 2010 and 2024. Certain words, "delves" and "pivotal" among them, spiked so abruptly after ChatGPT arrived that they work as a fingerprint. By this measure, at least 13.5 percent of 2024 abstracts had been processed with a language model, and in some fields and countries the share reached about 40 percent.

The LLM fingerprint in biomedical abstracts. Source: Kobak et al., Science Advances, 2025.

Surveys tell the same story from the inside. In a Springer Nature survey of over 2,000 researchers, 52 percent said they had used AI to help write papers or grant applications. A Nature survey of some 5,000 researchers found the community split on where the line sits, but with one point of near consensus: over 90 percent were comfortable using AI to edit or translate scientific text.

So the tools are in use, widely and mostly invisibly. The problem starts when the use becomes visible.

The transparency dilemma

You would expect honesty about AI use to build trust. The evidence points the other way. In a series of experiments published in Organizational Behavior and Human Decision Processes, Oliver Schilke and Martin Reimann found that people who disclosed using AI were trusted less than those who said nothing, across professions and tasks, and regardless of whether the disclosure was voluntary or required. The researchers named it the transparency dilemma.

Audiences bring the same instinct to what they read. In the Reuters Institute's 2025 report, only 12 percent of respondents said they were comfortable with news produced entirely by AI. People do not just evaluate the text in front of them. They evaluate how it was made, and who stood behind it.

The trust gap: how many researchers use AI to write, versus how many readers are comfortable with fully AI-made content. Sources: Springer Nature survey, 2025; Reuters Institute Digital News Report, 2025.

For anyone communicating science, this is the trap. Hide the tool and you risk being caught by readers who, by now, can smell machine text. Disclose it clumsily and you hand back the trust you were trying to earn.

What AI is actually good for

It would be easy to conclude that AI has no place in science communication. That would mean ignoring the strongest result in its favour.

David Markowitz, in a study in PNAS Nexus, took abstracts of published papers and had GPT-4 rewrite them as plain-language summaries. The AI versions were markedly simpler than the lay summaries the scientists had written themselves. Then came the twist. Readers who got the simpler AI summaries understood the science better, and rated the scientists behind the work as more credible and more trustworthy than readers who got the original human-written versions.

Read those two findings together and the picture sharpens. Audiences do not distrust clarity, whoever drafted it. They distrust outsourced judgment, the sense that no person stands behind the words. AI used as an editor tends to produce the first. AI used as an author produces the second.

Where to draw the line

The Nature survey suggests researchers are already converging on a workable rule. Editing, translating, tightening: broadly accepted, and most saw no need to announce it, any more than one credits a spell checker. Generating content, arguments, or text from scratch: that is where respondents said disclosure becomes necessary, and where discomfort grows.

That maps onto a distinction worth keeping. The thinking is the work. Deciding what matters in a result, what an audience needs, what an honest caveat looks like, which claim the evidence actually supports. None of that can be delegated, because it is exactly what readers are trusting you to have done. The wording is the vehicle, and help with the vehicle has never been scandalous. Researchers have always had their text improved by colleagues and editors.

A practical line: what AI can carry, and what has to stay human.

In practice, a few habits keep the line visible. Never let a model supply a fact, a number, or a reference; it will happily invent all three. Keep your own first draft, or at least your own outline, so the argument is yours. Read every AI-touched sentence as a skeptic, because your name absorbs the blame, not the tool's. And when AI has done more than polish, say so plainly and specifically. The transparency research carries a hopeful footnote here: the trust penalty shrinks when people explain what the tool actually did and what remained in human hands. Vague labels frighten readers. Specific ones reassure them.

Trust is the product

At Formidla, this is not an abstract debate. We use AI tools openly, for the vehicle, never for the judgment. A researcher checks every fact against its source, and nothing goes out that we cannot stand behind line by line.

The honesty penalty is real, but it is not an argument for silence. It is an argument for doing the human part so visibly well that the tools become a footnote. Audiences have never really been asking whether a machine helped with the words. They are asking whether someone knowledgeable is standing behind them. That question has only one good answer, and no model can give it for you.

Sources

- Kobak, D. et al. (2025). Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Science Advances. https://www.science.org/doi/10.1126/sciadv.adt3813

- Nature (2025). Is it OK for AI to write science papers? Nature survey shows researchers are split. https://www.nature.com/articles/d41586-025-01463-8

- Springer Nature (2025). Perspectives on AI in scholarly communications. https://stories.springernature.com/AI-perspectives/

- Schilke, O. & Reimann, M. (2025). The transparency dilemma: How AI disclosure erodes trust. Organizational Behavior and Human Decision Processes. https://www.sciencedirect.com/science/article/pii/S0749597825000172

- Reuters Institute (2025). Digital News Report 2025.

- Markowitz, D. M. (2024). From complexity to clarity: How AI enhances perceptions of scientists and the public's understanding of science. PNAS Nexus. https://academic.oup.com/pnasnexus/article/3/9/pgae387/7750129