Analyzing the News: Lula's fourth term

Eleven outlets covered Lula's comeback win in Brazil. Fíor read two of them as AI.

Fíor is a natural language classifier that predicts whether a piece of text was more likely written by a person or by AI.

On 25 October, Luiz Inácio Lula da Silva won a fourth term as Brazil's president, beating Senator Flávio Bolsonaro, the son of former president Jair Bolsonaro, in a runoff. With all votes counted, Lula had about 51.8% to Bolsonaro's 48.2%, a lead of roughly 4.4 million votes. Three weeks earlier, Bolsonaro had beaten him in the first round by more than two million votes.

Bolsonaro refused to concede. He says he will seek to annul the result because voting was extended by two hours. Both campaigns had asked for extensions in different areas, Lula's over transport problems and Bolsonaro's over heavy rain, and electoral officials say the extension did not change the outcome. Lula now faces a strongly conservative Congress.

The outlets were Reuters, AP, AFP, the Guardian, the BBC, TIME, PBS NewsHour, Sky News, The Independent, Euronews and Al Jazeera. That is three wire services, British and American newspapers and broadcasters, Euronews and Al Jazeera. PBS NewsHour was scanned from the transcript of a television segment, so much of it is spoken.

Two more pages were left out. The Washington Post's page again carried an AI-written summary of reader comments, and the scan picked it up with the article, so it is not counted, for the same reason as in the Meta penalty piece. CBS News ran AP's own reporting, so counting it would count AP twice. It still showed something useful. The same AP reporting read confidently as human on AP's site, and only narrowly as human on CBS's page. What surrounds an article on the page can move the result.

The result

Nine read as human: Reuters, AP, AFP, the Guardian, the BBC, TIME, PBS NewsHour, Sky News and Al Jazeera.

Two read as AI: The Independent and Euronews.

The BBC read as human, but not confidently. The Independent read as AI, also not confidently. For both The Independent and Euronews, Brath was unsure how the article reads. None of the pages drew a writing-habit flag.

This is the eighth story scanned with Brath, Fíor's new model. Euronews has now read as AI in six of its seven appearances in the series. The Independent has read as AI in both of its appearances since the new model arrived. The Guardian read as human again, and has now read as AI in six of the ten stories where it ran its own reporting.

The two that read as AI

Both put the result in the widest frame. The Independent's headline called it a blow to Trump that ends Latin America's right-wing streak. Euronews set it within a wider political shift across South America.

That framing is the obvious pattern. It does not hold up. TIME and Al Jazeera framed the result the same way, around the region and American influence, and both read clearly as human.

Euronews is the more persistent case. Something in how it writes or assembles its copy keeps reading this way to Fíor, story after story. Fíor can show the pattern. It cannot say what causes it.

Why election coverage reads like this

Election-night stories follow a known shape: the winner, the margin, the loser's response and what comes next. Every outlet had the same count and the same statements from both camps, and much of the writing was done against the clock.

The pieces that added their own reporting read as human. The Guardian went to Lula's supporters. The BBC looked at the Congress he now has to deal with. AFP followed the market reaction. PBS NewsHour told the story on air.

Some of these pieces may have had AI help, with a first draft, a rewrite or a final polish. Or election coverage has become formulaic enough that a reporter and a model would write it the same way. Fíor cannot tell those apart on its own.

Bolsonaro wants the result thrown out over two extra hours of voting, in a race decided by 4.4 million votes. Fíor's results come with margins of their own, and two of this week's calls were marked not confident. A close call deserves to be labelled as one, whether it comes from an election or a classifier.

What Fíor is

Fíor is a machine learning classifier, currently in beta and available as a browser extension. It gives a prediction, not a verdict. It cannot see who sat at the keyboard, and nothing here is a claim about any journalist or publication. Reading as AI and being written by AI are not the same thing.

Treat it as a signal, not proof, and check the pages you read yourself.

Try the browser extension at Fíor.

Analyzing the News is a series where we run one major story through Fíor and see what comes back.