Science Popularization Writing in the Age of Generative AI: Efficiency,Authorship,Trust,and Fidelity

Introduction

Generative artificial intelligence (AI) has moved quickly from a technical novelty to a practical writing tool. Scientists, educators, journalists, public information officers, and science communicators increasingly use large language models to summarize research, generate outlines, revise drafts, translate technical concepts, and produce audience-specific versions of the same message. In science popularization writing, these functions appear especially attractive. Many public-facing science texts require speed, clarity, analogy, narrative framing, and linguistic accessibility. Generative AI seems able to provide all of these at low cost and with little delay.

Yet the usefulness of generative AI in science popularization should not be confused with communicative authority. Popular science writing is not only a matter of making expert knowledge easier to read. It also involves selecting what counts as relevant, deciding how much uncertainty to reveal, preserving the difference between evidence and interpretation, and maintaining a responsible relationship between scientific institutions and public audiences. These tasks cannot be reduced to fluent textual production. They require accountable human judgment.

Recent work in science communication has begun to examine generative AI as a tool for science communication practice. Hendriks et al. (2025) describe generative AI as a potential writing assistant, idea generator, and collaborative support tool, while also emphasizing risks such as misinformation, prompt dependency, and the reproduction of biases. Barel-Ben David et al. (2026) further examine whether generative AI can support higher-level science communication strategies in writing, showing that its role needs to be assessed in relation to established science communication goals rather than mere textual fluency. These studies suggest a central question for science popularization: not whether generative AI can write, but under what boundaries its writing support can remain scientifically accurate, ethically responsible, and publicly trustworthy.

This commentary argues that generative AI introduces four major boundary issues for science popularization writing: efficiency, authorship, trust, and fidelity. It then proposes a human-led, AI-assisted workflow that can help science communicators use generative AI without transferring responsibility to the tool itself.

From Writing Tool to Communication Infrastructure

Science communication scholarship has long moved beyond the simple deficit model, in which public communication is treated as the one-way transmission of expert knowledge to uninformed audiences. Public understanding, engagement, trust, values, and social context are now widely recognized as part of the communication process (Brossard & Lewenstein, 2010; Bucchi & Trench, 2021; National Academies of Sciences, Engineering, and Medicine, 2017). This shift matters for generative AI because AI-assisted writing may easily return science communication to a narrow transmission logic. If a system is asked to “explain this paper to the public,” it may produce a fluent summary, but fluency alone does not ensure relevance, responsibility, or trust.

Generative AI is not simply a neutral instrument added to the end of the writing process. It can influence topic selection, framing, tone, metaphor, structure, headline style, and the perceived certainty of scientific claims. In this sense, it becomes part of the communication infrastructure. It shapes not only how scientific information is expressed but also how science is interpreted by non-specialist audiences. For this reason, its use in science popularization requires explicit norms.

A central problem is that generative AI systems are optimized to produce plausible language, not to guarantee scientific truth. Bender et al. (2021) warned that large language models can produce convincing text without grounded understanding. In scientific writing contexts, fabricated or inaccurate references have been repeatedly identified as a serious risk (Alkaissi & McFarlane, 2023; Walters & Wilder, 2023). For science popularization, the problem may be even more subtle. Errors may not appear as obviously fake references. They may appear as overconfident simplifications, missing caveats, unsupported causal claims, or analogies that distort the underlying mechanism.

Therefore, generative AI should be treated as a conditional writing assistant. It can accelerate textual labor, but it cannot assume responsibility for scientific interpretation.

Boundary 1: Efficiency Without Substitution

The strongest appeal of generative AI lies in efficiency. It can help science communicators generate alternative titles, simplify terminology, produce summaries at different reading levels, create question-and-answer structures, and compare several narrative frames. For researchers with limited time and communication training, this support may lower the threshold for public engagement (Hendriks et al., 2025). For non-native English writers, generative AI may also reduce linguistic barriers and allow scientific ideas to be expressed more smoothly across languages.

However, efficiency becomes problematic when it substitutes for communicative judgment. Science popularization is not merely the compression of expert knowledge. It requires choices about audience needs, public concerns, social context, and the limits of available evidence. An efficient text may still be a poor science communication product if it hides controversy, exaggerates novelty, or presents early-stage research as settled knowledge.

The responsible use of generative AI should therefore distinguish between low-risk linguistic assistance and high-risk interpretive work. Low-risk uses include grammar polishing, structural suggestions, translation drafts, and alternative wording. Higher-risk uses include claim generation, literature synthesis, policy implication writing, health or environmental risk explanation, and public recommendation. In high-risk contexts, human review should not be optional. It should be built into the workflow.

The practical question is not whether AI can make science popularization faster. It clearly can. The more important question is how to prevent speed from becoming the primary value of public science writing. In science communication, slowness sometimes has epistemic value. Checking sources, preserving uncertainty, consulting experts, and revising misleading analogies take time because they protect the integrity of the message.

Boundary 2: Authorship and Accountability

Generative AI also complicates authorship. Popular science writing often depends on the perceived credibility of the author. A scientist, journalist, educator, or institutional communicator is not only a language producer but also a source of accountability. Readers may trust a text because they assume that a responsible human author has selected evidence, checked claims, and accepted responsibility for errors.

AI systems cannot meet this standard. Publication ethics organizations have generally emphasized that AI tools should not be listed as authors because authorship implies responsibility, consent, accountability, and the ability to respond to criticism (Committee on Publication Ethics, 2023; International Committee of Medical Journal Editors, 2026; Zielinski et al., 2023). Although these guidelines are mainly written for scholarly publishing, the principle is also relevant to science popularization. Public-facing science writing affects how audiences understand health, climate, technology, energy, risk, and social decision-making. Responsibility for such texts must remain human.

The authorship issue is not limited to whether AI is named as an author. More often, the issue is invisible delegation. A communicator may allow AI to generate the central explanation, select examples, write the headline, or frame the public implication, while presenting the final text as entirely human-authored. This practice can blur accountability even if the human later edits the draft.

A more responsible approach is role clarification. Generative AI may be acknowledged as a tool used for brainstorming, language revision, audience adaptation, or translation assistance. The human author should remain responsible for the argument, evidence, interpretation, and final wording. In institutional science communication, this responsibility should be assigned clearly among the researcher, communication officer, editor, and approving organization.

Boundary 3: Trust and Transparency

Trust is a central concern in science communication. Public trust is shaped not only by factual accuracy but also by perceived honesty, competence, openness, and respect for audience concerns. Generative AI may support trust when it helps produce clearer and more accessible explanations. It may also undermine trust when audiences suspect undisclosed automation, discover factual errors, or encounter texts that sound polished but lack specificity.

Transparency is therefore necessary, but it must be meaningful. A vague statement such as “AI was used in preparing this text” may not be enough. Readers, editors, and institutions need to know how the tool was used. Did it polish language? Did it generate the first draft? Did it summarize a scientific paper? Did it create analogies? Did it translate the text? Did it suggest claims that were later verified by experts? These distinctions matter because different uses carry different risks.

At the same time, transparency should not become a substitute for quality control. Disclosing AI use does not make an inaccurate explanation acceptable. Nor should AI disclosure be used to transfer responsibility away from human authors. The purpose of transparency is to support accountability, not to excuse weak verification.

Science communicators should also be careful about over-reliance on AI detection tools. Detection systems can produce uncertain results and may create fairness problems, especially for non-native writers whose English style may be wrongly classified as AI-generated (Liang et al., 2023). A trust-centered approach should focus less on detecting whether a text “sounds like AI” and more on whether the claims are accurate, sources are traceable, uncertainty is preserved, and responsibility is clear.

Boundary 4: Fidelity to Scientific Detail

The most important boundary in AI-assisted science popularization is fidelity. Fidelity does not mean that every technical detail must be retained. Popular science writing necessarily involves selection, simplification, and translation. Rather, fidelity means that simplification should not change the scientific meaning of the claim.

Generative AI can threaten fidelity in several ways. First, it may overgeneralize from limited evidence. A study involving a narrow population may be presented as if it applies universally. Second, it may convert correlation into causation. Third, it may remove uncertainty because confident prose sounds more readable. Fourth, it may produce attractive analogies that misrepresent the mechanism. Fifth, it may make a technology appear more mature than it actually is.

These problems are especially serious in emerging scientific and technological fields, such as AI, biotechnology, energy storage, climate engineering, quantum technology, and medical innovation. In these areas, public audiences need not only enthusiasm but also boundary conditions. They need to know what is known, what remains uncertain, what is technically possible, what is economically or socially constrained, and what claims are speculative.

A fidelity-centered approach should therefore preserve at least five distinctions in science popularization writing: evidence versus interpretation, current capability versus future possibility, laboratory result versus real-world application, expert consensus versus active debate, and scientific uncertainty versus ignorance. Generative AI may assist in drafting these distinctions, but it cannot decide them independently.

A Human-Led, AI-Assisted Workflow

To use generative AI responsibly in science popularization, science communicators need a workflow that places human judgment before, during, and after AI assistance. The following model is proposed as a practical framework.

First, define the communication purpose before using AI. The author should identify the audience, public question, key scientific claim, evidence base, and intended level of certainty. Without this step, AI may optimize for general readability rather than communicative responsibility.

Second, ground the draft in verified sources. AI should not be asked to invent references or summarize unspecified evidence. The author should provide validated source material, such as peer-reviewed articles, official reports, expert notes, or institutional documents. When possible, AI should be used to transform provided material rather than retrieve facts independently.

Third, use AI for multiple controlled drafts rather than one final text. A useful practice is to ask for several versions with different communication goals: a plain-language version, a question-led version, a metaphor-light version, a version preserving uncertainty, and a version for readers with no background knowledge. This allows the human author to compare rhetorical choices instead of passively accepting one output.

Fourth, conduct expert verification. A subject expert should review the scientific claims, technical mechanisms, numerical values, analogies, and uncertainty statements. This step is essential when the topic involves health, environment, safety, public policy, or emerging technology.

Fifth, perform communication review. A science communication editor should assess whether the text is understandable, respectful, non-exaggerated, and suitable for its audience. This review should include checking whether simplification has distorted meaning.

Sixth, disclose AI use in a precise way. The disclosure should state what tool was used, for what purpose, and what human review was conducted. For example, a disclosure may say that generative AI was used to produce alternative plain-language drafts and improve sentence clarity, while all scientific claims, sources, and final wording were checked and approved by the human author.

Seventh, maintain a correction pathway. Public science texts should allow post-publication correction when errors are found. AI-assisted writing may increase the speed of production, so correction systems must also become more responsive.

This workflow does not reject generative AI. It treats AI as useful but bounded. Its core principle is simple: AI may assist the writing process, but humans must retain responsibility for purpose, evidence, interpretation, and accountability.

Implications for Science Communication Practice

The rise of generative AI requires science communication training to change. Traditional training often emphasizes clarity, narrative, jargon reduction, audience awareness, and media skills (Baram-Tsabari & Lewenstein, 2017). These remain important, but they are no longer sufficient. Future science communicators also need AI literacy. They must understand what generative AI can do, where it fails, how hallucinations occur, how bias may be reproduced, and how to verify AI-generated text.

Institutions also need editorial policies for AI-assisted science popularization. Many universities, journals, museums, science centers, media organizations, and public agencies already produce large volumes of public-facing science content. Without shared standards, AI use may become inconsistent and invisible. Some communicators may use AI only for language polishing, while others may use it to draft entire explanations. This inconsistency can create uneven quality and unclear accountability.

A useful institutional policy should not simply ban or permit generative AI. It should classify uses by risk. Low-risk uses may be allowed with minimal disclosure. Medium-risk uses may require internal review. High-risk uses should require expert verification, source documentation, and explicit disclosure. This graded approach is more realistic than a universal rule.

For science popularization writers, the main task is to protect the public value of science communication. Generative AI can help make science more accessible, but accessibility without fidelity is not communication; it is simplification without responsibility. The goal should not be to produce more content, but to produce more trustworthy, useful, and context-aware communication.

Conclusion

Generative AI will not disappear from science popularization writing. Its advantages in speed, language adaptation, brainstorming, and translation are too significant to ignore. However, science popularization cannot be reduced to efficient text generation. It is a public act of interpretation, selection, explanation, and responsibility.

This commentary has argued that the boundaries of AI-assisted science popularization can be understood through four concepts: efficiency, authorship, trust, and fidelity. Efficiency should support but not replace judgment. Authorship should remain human and accountable. Trust should depend on transparency and verification rather than polished language. Fidelity should preserve the scientific meaning of claims, including uncertainty and limits.

A human-led, AI-assisted workflow offers a practical way forward. It allows science communicators to benefit from generative AI while maintaining the standards that make science communication socially valuable. In the age of generative AI, the central question is not whether machines can write popular science. The central question is whether humans can use them without weakening the responsibility that public science writing requires.