
Andrii Yurychko, journalist, media trainer, Candidate of Philological Sciences and lecturer at the Educational and Scientific Institute of Journalism, Taras Shevchenko National University of Kyiv
Imagine a note beneath a piece of journalism: ‘Text typed in a licensed copy of Office 365. Video edited in the free version of DaVinci Resolve. Illustration processed in Adobe Photoshop.’ And if Photoshop turns out to be pirated – heaven forbid – we will have to disclose that too.
We could add the browser, the voice recorder model and the spelling service. Eventually, the production notes would be longer than the story. Readers would know almost everything about the software and still have no idea where the author’s main conclusion came from.
I am not persuaded by demands to list every programme a newsroom uses. A Word licence does not certify the truth of what is written, and a free video editor can produce a report based on verified facts. Explaining what work a programme actually did, however, helps us assess the author’s contribution and understand their responsibility.
I do not credit the word processor in which I type and format a page with my conclusions. Generative AI is more complicated: a system can propose an explanation of an event, select arguments and formulate a position. A person remains responsible for publication, but readers should know about that contribution. The label ‘AI was used’ does not explain it: it could mean corrected commas or a borrowed central conclusion.
On 8 November 2026, Career Communications Group publisher Tyrone D. Taborn published ‘AI Is Part of My Editorial Team. It Is Not the Author’. He compares AI assistance to the work of editors and researchers: an author may use it while developing their own idea, deciding the direction and making the final decision. If a system shapes the position and writes the finished article without meaningful human involvement, Taborn believes readers should know how it was produced.
I find that argument convincing. Several people often work on an author’s text, and the proportion of words changed tells us little about the origin of an idea. Extensive editing may preserve it; one added sentence may alter the entire conclusion.
The editorial-team analogy does not, however, resolve the question of accuracy. Editors can make mistakes too. A source suggested by a language model must be opened and checked to see whether it supports the text.
In October 2025, the European Broadcasting Union and BBC published research in which journalists from 22 public service media organisations in 18 countries assessed responses from four AI assistants in 14 languages. At least one significant issue appeared in 45% of assessed responses; 31% had serious sourcing issues. These results concern a particular test of news-related answers. Applying them to every task, system or later version would be inappropriate.
A mistake can arise in ordinary editing too. Remove ‘possibly’ or ‘according to a preliminary estimate’, and a tentative assessment becomes a categorical statement. Put a paraphrase in quotation marks, and it looks verbatim, although the interviewee may never have used those words. A stylistic edit has changed what readers understand as fact.
‘Help me express my thought more precisely’ and ‘Suggest an opinion I will agree with’ leave different work for the person. In the second case, you might recognise yourself in the finished text and stop there. Or you might challenge its explanation, find other evidence and construct your own. An AI suggestion can certainly begin an independent examination; approving a draft does not show whether that examination took place.
For me, trust in a news organisation depends partly on understanding the work it has done. ‘The author checked it’ does not explain who constructed the argument. Every date and figure can be verified in a text whose logic was supplied entirely by software.
Imagine a report on the closure of a district hospital: accurate budget figures, a correctly quoted official and an explanation of rationalisation. It may lack an answer to one question: how will someone without a car reach the new treatment location? The article can be factually accurate while leaving readers without an explanation of a major consequence.
Journalism knew this problem before AI. With a ready-made machine draft, it is easy to limit yourself to checking the facts already included. To notice omissions, journalists need subject knowledge, fresh questions and conversations with people who may be absent from the draft.
The Paris Charter on AI and Journalism sets requirements for newsroom use of AI. Reporters Without Borders (RSF) and 16 partner organisations published it on 10 November 2023. RSF describes it as the first global ethical benchmark in this field. Individual newsroom rules existed earlier: Associated Press announced its own in August 2023.
The fourth of the Charter’s ten principles establishes media responsibility for publication. The fifth requires disclosure when AI significantly affects the production or distribution of journalism. It also requires a public record of AI systems used, describing their purposes and conditions of use. Other principles cover human oversight, independent evaluation, content provenance, source protection, distinguishing authentic from synthetic material and relations with AI developers.
Such a record explains a newsroom’s general practice. A note beside an individual article explains how the technology was used in that piece.
During digitalisation mentoring in the National Union of Journalists of Ukraine’s ‘Strengthening the Resilience of Frontline Media as a Tool Against Disinformation’ project, I worked with newsrooms short of staff and resources. The project was supported by the Lithuanian Embassy in Ukraine. In feedback for the NUJU, Tetiana Kaushan, editor of Putyvlski Vidomosti, highlighted the usefulness of automatic transcription, which I had discussed in my sessions.
For a small newsroom, transcription can free time for another interview or work on a document. Tetiana Kaushan’s feedback concerns that convenience. It cannot establish journalism quality or a programme’s ability to write a column.
The newsroom decides where the saved hour goes. It can spend it on further verification or demand several more articles from a journalist. If drafts multiply while reading time stays unchanged, ‘a person checks everything’ may mean little more than a quick glance before publication.
In a Reuters Institute online survey across nine countries, published on 29 October 2026, 45% of respondents felt comfortable with news produced mostly by a person with AI assistance. For news produced mostly by AI with some human oversight, the figure was 19%. Among those who had used AI for news in the preceding week, 72% reported trusting its answers. Ukraine was not included.
The first two figures describe attitudes towards how news is produced; the third describes trust within a particular user group. They demonstrate neither general acceptance of machine-produced news nor wholesale rejection of AI. Nor do they prove that a technology disclosure alone increases trust in a newsroom.
Some AI-related work is entirely invisible to readers. If an unpublished interview recording goes to an external service, we must establish how it stores data and who can access it. A note beneath the finished article does not explain what happened to the recording beforehand. Source confidentiality must be protected during preparation.
With images, the significance of production methods is clearer. A verified photograph of destruction taken at the scene can serve as evidence. A generated image shows what destruction might look like. Both may have a similar emotional effect, but a generated illustration does not become evidence of an event simply because it carries a provenance label.
The Photoshop reference at the start of this column would not tell us whether the photographer adjusted exposure or added a detail that was never before the camera. Assessing material requires understanding the intervention. The same applies to text: a system may add an explanation or remove a contradiction the author has not yet investigated.
At the NUJU webinar ‘Can AI Be Made to Cite the Media?’, I discussed newsroom visibility in generative search. AUP prepared an article based on the webinar. Against a background of interchangeable generated text, original reporting matters even more to me: knowing local life, speaking to witnesses and obtaining documents. That work makes facts publicly available that might otherwise never appear.
Whether a newsroom earns income from the information it gathers is another question. Pew Research Center analysed the search activity of 900 US adults in March 2025. Visits to Google pages with an AI summary produced clicks on ordinary search results in 8% of cases, compared with 15% without one.
The research describes US users’ behaviour. It does not measure Ukrainian newsrooms’ losses or prove that the AI summary alone caused the difference. Yet the risk for a story’s author is clear: a search engine may use their facts while the reader settles for the summary and never visits the newsroom’s site.
Automation can make text cheaper to prepare. Travelling to a community, finding a document and interviewing someone still cost time and money. Lost visits and related revenue may leave a newsroom unable to fund its next report.
There is also the question of experience. Journalists learn to construct articles through their own attempts, editorial revisions and conversations with sources. If a beginner immediately receives a finished draft, some of that work changes. Will they learn to notice gaps while working mainly within a structure already supplied? A tool’s convenience does not answer that question, which matters to journalism education.
Approving a finished text can conclude meaningful authorship or replace it. A byline alone does not tell us which.
As a reader, I want to know what would be missing from this text without its author.
The Academy of Ukrainian Press delivers practical AI training for education and communication. Its ‘EDU-AI: Methods of Using AI for Communication Educators’ webinar series covers learning materials, verification and professional ethics. On 4–5 November 2025, journalism, PR and related lecturers planned courses with AI assistants, examined false information and deepfakes, and discussed using the technology in their disciplines. The event was supported by the Friedrich Naumann Foundation for Freedom.
In ‘FreeMind Ukraine: Youth and Communities Against Manipulation’, implemented by AUP and Freie Universität Berlin, work with generative tools is combined with fact-checking, manipulation analysis and digital security. During training in Rivne on 13–14 October 2026, students checked generated links, studies and figures and discussed the limits of digital assistance. They then developed university initiatives, identifying a problem, an audience and the change they wanted to achieve.
Translation note: This is a literary translation into British English, produced and proofread using a language model. Read the Ukrainian original.