Brazil's newsrooms are arguing over who actually wrote the column

01

WHAT ACTUALLY CHANGED

This is a continuing debate, not a fresh newsroom incident. On September 10, the Media Diversity Institute published a new report by Brazilian journalist Raphael Tsavkko Garcia about a dispute that began at Folha de S.Paulo in February. His interviews move the story beyond one columnist and toward a harder question for every publication: when does AI assistance become part of the authorship?

The original episode is unusually well documented. On February 7, Folha ombudsman Alexandra Moraes wrote that a subscriber had flagged one opinion column as likely AI-generated. Moraes tested more columns by the same author with two commercial detectors and said most produced extremely high AI scores. She also warned that detector output alone was not enough to judge someone's work, then asked the columnist and the newspaper directly.

Columnist Natalia Beauty confirmed that she used AI and defended the practice. She told the ombudsman that she chose the subject, developed the argument aloud, supplied her views and conclusions, and used the system to organize that material. On February 9, she published her own response under a headline that translates to, "My texts use artificial intelligence; my thinking does not." Her position was that authorship rests in the ideas, experience, judgment, and responsibility, not in manually typing every sentence.

Folha's ombudsman identified a policy gap rather than a clear rule violation. The newspaper's editorial manual permitted AI tools, required human review, and said the journalist remained responsible for the result. Moraes wrote that Folha did not prohibit AI-written text and had no rule requiring an AI label. Her criticism was that readers had not been told when the technology moved from an instrument in the workflow to a substantial part of the finished product.

Garcia's September report interviewed veteran journalist Paulo Markun, technology journalist Cora Rónai, and Rio de Janeiro State University media professor Erick Felinto. They did not agree on every boundary. They did converge on transparency and responsibility. The useful distinction in their discussion is not a magic percentage of machine-written sentences. It is whether the named author still chose the question, built the argument, selected the evidence, made the judgments, and can answer publicly for the result.

02

WHY THIS MATTERS

A byline is more than a label for whoever pressed publish. Readers use it as a promise that a particular person did the reporting or thinking, made the relevant choices, and accepts responsibility for the work. AI can assist that chain, just as editors, translators, researchers, and transcription tools already do. The problem begins when the chain changes materially and the reader is left to imagine the old one.

Opinion columns make the question sharper. People often read them for an individual voice, not merely clean sentences arranged around a thesis. If a system turns spoken notes into most of the prose, the writer may still own the argument. But the reader may reasonably want to know that the voice reached the page through a machine that can standardize tone, smooth uncertainty, and introduce errors with impressive manners.

Detection cannot carry the policy. Commercial AI detectors can produce false positives and do not reveal who chose the argument, checked a claim, or edited the final version. In this case, the scores prompted the ombudsman to ask questions, and the columnist then confirmed her process. That is a much stronger evidentiary sequence than treating a probability meter as a tiny robot judge.

Brazilian newsrooms have a practical chance to write rules before improvisation becomes habit. A useful policy can separate low-stakes assistance, such as transcription or spelling, from substantial generation, such as drafting the argument or producing publishable passages. It can also require source checking, preserved notes, human editing, disclosure at a clear threshold, and one accountable editor who can stop publication.

The signal travels beyond Brazil because the economic temptation is universal. Generative AI can make routine prose faster and cheaper. If news organizations use those savings without making responsibility visible, they risk weakening the trust that gives a byline commercial and civic value in the first place. The scarce product is no longer fluent text. It is accountable judgment with receipts.

03

WHERE IT COULD HELP

• Publish a plain-language disclosure when AI drafts or substantially restructures an article
• Define separate rules for transcription, research support, translation, editing, and generative drafting
• Preserve source notes, prompts, drafts, and human edits for sensitive work
• Require a named editor to verify claims and accept responsibility before publication
• Use detector scores only as a reason to investigate, never as proof by themselves