Artificial intelligence in Moldova’s press: speed cannot replace editorial responsibility

Artificial intelligence can help a newsroom search documents, transcribe interviews, translate and explain information. But it cannot take over its duty to verify, protect sources and take responsibility for what it publishes. In the Republic of Moldova, this distinction is no longer just a recommendation for the future: the Journalists’ Code of Ethics already contains a chapter devoted to the use of AI.

The important question is not whether the press should use artificial intelligence. The question is under what conditions its use produces better journalism, not just more content.

A newsroom can save time with the help of an automated tool. But time saved is not, in itself, evidence of editorial progress. It benefits the public only if it is turned into additional research, clearer explanations, greater accessibility and accuracy.

One word must remain at the centre of this discussion: responsibility. The tool can generate the text. The media organisation decides whether to put it in front of readers.

Moldova already has rules. The challenge is applying them

In the Republic of Moldova, the debate about AI in journalism is not starting from a blank page.

Amendments and additions to the Code of Ethicsapproved at the Media Forum on 3 December 2024 include explicit rules on artificial intelligence. They require accuracy, independence, confidentiality and copyright to be respected, as well as the public to be informed when media content is generated wholly or partly with AI.

The decisive provision is point 3.2: “When using AI, editorial decision-making remains at all times the essential and exclusive responsibility of journalists and media organisations.”

This is a professional self-regulatory standard; it should not be presented as a special law on artificial intelligence. Its importance is clear, however: the use of technology does not move editorial decision-making outside the newsroom. The Code assigns responsibility for applying it to every journalist and every newsroom, while the Press Council is responsible for monitoring and examining complaints.

Therefore, the explanation “that’s what the algorithm wrote” is not an editorial justification. It describes a process that was not adequately controlled.

A convincing text is not necessarily a true one

One important risk of generative AI is the difference between fluency and accuracy. A text can be coherent, well-structured and confidently worded, yet contain incorrect information. Google’s guidance on using generative content warns about these inaccuracies and calls for manual verification before publication, including of headlines, descriptions and structured data.

For a newsroom, the practical consequence is simple: quality of expression cannot take the place of evidence.

Suppose a tool summarises a draft resolution and presents it as a decision that has already been adopted. Or it turns a maximum eligible amount into assistance guaranteed to all applicants. These are hypothetical examples, not incidents attributed to any publication. But they show where verification needs to focus: the status of the document, the conditions, exceptions and date of application.

It is not enough for an editor to ask: “Does it sound good?” They must ask:

• Is the document authentic and up to date?

• Are we talking about a proposal, an approval or a measure that is already in force?

• Does the figure have the same unit, period and meaning as in the source?

• Does the quote exist in the form reproduced?

• Has a condition that changes the meaning of the information been removed?

These questions are not an exotic procedure for the AI era. They are ordinary journalistic checks that must also be applied to texts produced with new tools.

An official source does not replace journalism

Using original documents is essential. But the existence of an official source does not exempt a newsroom from critical analysis.

A press release shows that an institution made an announcement. It does not automatically prove that the promised results will be achieved, that the estimates are complete or that all affected people have been heard.

That is why a good article must distinguish between:

• what the document establishes;

• what the institution states;

• what the data allow us to conclude;

• what remains unknown.

The Code of Ethics calls for collected information to be questioned, relevant positions to be presented and, as a rule, information to be verified using at least 2 independent sources. This is not about mechanically collecting 2 links, but about genuinely verifying the information.

For AI, this distinction needs to be turned into a working rule: a press release should not be automatically rewritten into an account that adopts the issuer’s perspective.

“The ministry estimates” cannot become “it will happen”. “The company claims” cannot disappear from the sentence just because the summary needs to be shorter.

Automated verification is useful, but it is not independent confirmation

Moldova Comunică states in its policy that AI-generated drafts are automatically checked against sources, including for figures, names, dates and quotes, and that news articles are published only after approval by a human editor. The policy specifies that automated verification does not replace the editor’s judgement.

This distinction is important. An automated check can help identify discrepancies. It should not, however, be treated as a final verdict.

A system can check whether an amount appears in a document without determining whether the amount represents the annual budget, the cost of the entire project or just one funder’s contribution. It can confirm that a statement exists without noticing that a correction has since been published.

The Council of Europe recommends that editorial oversight cover both the results and the processes that produced them. Its guidelines for AI in journalism stress that translating professional values into computer rules cannot replace editorial control.

A second automated check is an additional filter. Merely repeating the process does not turn it into a second source.

Transparency should explain the process, not just display a label

The label “made with AI” tells readers that technology was involved. It does not necessarily tell them what it did.

Did it transcribe an interview? Translate a text? Suggest the headline? Write most of the article? Generate the image or voice?

The differences matter. The Council of Europe recommends transparency about when and how AI is used if this may influence the interpretation of content or the rights of the public and the people concerned. The Moldovan Code of Ethics requires explicit disclosure when content is generated wholly or partly with AI.

A useful disclosure should explain, in proportion to the technology’s role:

1. Which part of the content was produced with AI.

2. What human verification took place.

3. What relevant sources and limitations apply.

4. How an error can be reported.

Transparency is not an exemption from responsibility. Readers should not understand the label to mean: “We used AI, so check for yourself whether it’s true”.

Access to sources gives the public the opportunity to verify. The obligation to verify before publication remains with the newsroom.

The same standard should apply to a publication’s own content

A publication that discusses the responsible use of AI must also examine its own procedures.

Moldova Comunică states that it uses AI for drafts, translations, summaries, answers to questions, audio readings and illustrations. Its policy makes a relevant distinction: news articles are approved by an editor, but the Daily Summary is generated automatically, without a separate human review, from already approved articles and public data taken from official sources. The headline and introduction are also generated with AI.

This transparency is necessary. But it does not close the discussion about risk.

The fact that the original article was checked does not guarantee that every subsequent rewording preserves all its conditions. A summary may omit an exception, while a new headline may make a more categorical conclusion than the source.

Therefore, this article proposes that automated products have controls proportionate to the risk, rather than being considered safe merely because of the origin of the information.

For structured data, the newsroom can establish checks for the date of the latest update, units of measurement and source availability. For sensitive information, it can introduce escalation to an editor or exclude it from summaries published without human review.

If a source cannot be accessed, the system should say so, rather than fill the gap with plausible wording.

These are recommendations for developing procedures, not claims that such errors have already occurred on comunica.md.

Romanian, Russian and English are not just three versions of the same page

For a multilingual publication, machine translation is a genuine opportunity to improve accessibility. But every published version also represents a distinct editorial relationship with the reader.

Moldova Comunică's policy marks automated translations and states that the original version remains the reference. It also acknowledges that translation may lose nuance.

This clarification should not make readers of translated versions readers with a lower level of editorial protection.

In a text about justice, “suspect”, “accused” and “convicted” are not interchangeable terms. In an economic article, “income”, “profit” and “turnover” do not mean the same thing. And the difference between “could benefit” and “will benefit” can change a family's expectations.

The solution is not to abandon translation, but to apply appropriate controls: glossaries, checks of sensitive terms and coordinated correction of all versions when an error is discovered.

The quality of the original version should be the starting point, not the limit of the editorial obligation.

Photography, voice and source protection require distinct rules

Not all uses of AI carry the same risk.

An abstract illustration about technology does not serve the same purpose as an image that appears to document an event. A labelled audio reading is not equivalent to a synthetic voice presented as a person's statement.

The Code of Ethics calls for authentic images and recordings to be given priority in coverage of real events and for a clear distinction between them and content created or modified with AI. It also calls for avoiding generated content that imitates real recordings or realistically depicts real people in a way that misleads the public.

In the standards updated on 23 July 2026, the Associated Press continues to prohibit the use of generative AI to create, modify or enhance news photography. The AP example is its own institutional standard, not a legal rule that automatically applies to all newsrooms, but it demonstrates the importance of distinguishing between documentation and illustration.

What a journalist puts into the tool is just as important.

An unpublished document, a confidential recording or a source's identity should not be uploaded to an external service just because it can quickly produce a summary. Data protection and confidentiality are part of the responsible evaluation of a tool, not secondary technical details.

The newsroom must know what data it transmits, who can access it and what contractual terms apply. If these questions do not have satisfactory answers, convenience does not justify exposing the source.

SEO is no excuse for producing content without value

Artificial intelligence can suggest clear headlines, descriptions and easy-to-follow structures. But search engine optimization should not be confused with mass-producing pages that rephrase the same information.

Google states that the use of AI is not prohibited in itself. What matters is the usefulness and quality of the content; generating many pages without added value may violate the scaled content abuse policy.

For a newsroom, added value should mean verifiable things: the original document, local context, an original explanation, a question put to an institution, an observed limitation or a fair comparison.

The same discipline is useful when content may be picked up in AI-generated answers. Google says that basic SEO principles remain relevant for its own AI features: technical accessibility, compliance with policies, and useful, reliable, people-first content. These recommendations do not guarantee inclusion and should not be extrapolated into a promise about every AI agent.

A publication should write clearly to be understood, not make claims more categorical so they are easier to quote.

What a credible editorial minimum looks like

Standards should not remain on a page that nobody consults while working.

Drawing on the Code of Ethics and the Council of Europe's recommendations, I propose the following operational minimum for a newsroom that uses AI:

• Research: AI identifies leads; the journalist consults the original documents and checks that they are up to date.

• Writing: every central claim must be traceable to identifiable evidence.

• Quotes: no statement is published without checking its wording and context.

• Approval: the editor has time, access to sources and the authority to stop publication.

• Translations: sensitive terms are checked, and corrections are carried across versions.

• Automated products: there are topic limits, monitoring, incident reporting and the ability to stop them quickly.

• Transparency: the public is told what role AI played and what editorial oversight was applied.

• Corrections: the error is corrected in the summary, audio, translations or posts as well, not just in the original article.

Success should not be measured only by the number of pieces and minutes saved. A newsroom should also track errors identified before publication, subsequent corrections, response times and cases where automation was justifiably stopped.

A good procedure is not one that allows any text to be published. It is one that also allows publication to be refused.

AI should leave more time for journalism

In its updated standards, the Associated Press allows AI to be used for tasks such as preliminary research, summarizing documents, transcription, translation and suggesting headlines. It specifies, however, that journalists review and edit the results before publication, and that AI does not replace research, sources, verification or editorial judgment.

This is a useful benchmark, not an invitation to copy mechanically. Each newsroom must develop its procedures according to its products, resources and risks.

The fundamental choice remains the same: do we use the time saved to publish more or to verify better?

If automation allows a reporter to consult an additional document, ask for explanations and speak with the people affected, it serves journalism. If it eliminates precisely these steps, it serves only text production.

The public does not need a newsroom to prove how quickly it can generate a news story. It needs to know why it should believe it.