Sérgio Lüdtke is a journalist and editor of Projeto Comprova, a coalition of 24 media organizations working collaboratively to investigate rumors about public policy in Brazil.
In 2018, Comprova reviewed suspicious content shared on social media and messaging apps about the presidential elections in Brazil.

On Sept. 6, 2018, a month before Brazil's presidential election, far-right candidate Jair Bolsonaro held a campaign event in downtown Juiz de Fora, a city of 560,000 people 200 kilometers from Rio de Janeiro.

It had been a week since Bolsonaro became the leader in the first-round polls for the Brazilian presidential election. He took the first position after the candidacy of former President Luiz Inácio Lula da Silva, the previously isolated leader in the polls, was barred by the Superior Electoral Court.

Bolsonaro, however, was losing in the runoff simulations to three of the four closest candidates in the polls.

Bolsonaro’s situation was worrisome, since he had only two 9-second daily blocks in the free electoral broadcasts on TV. Brazilian electoral rules require radio and TV stations to give free time to political parties to publicize their proposals. This time is distributed according to the number of seats won by each party in the last election of the House of Representatives. Bolsonaro’s lack of seats meant very little free airtime. As a result, he had to rely on his supporters on social networks and make direct contact with voters on the streets.

In Juiz de Fora, as in other cities he visited before, Bolsonaro participated in a march by being carried on the shoulders by his supporters. He was trailed by a crowd of admirers when the march was suddenly interrupted. In the middle of the crowd, a man reached out and stabbed the candidate. The knife left a deep wound in Bolsonaro’s abdomen — and opened a Pandora’s Box on social networks.

Rumors and conspiracy theories spread, with some accusing Adélio Bispo de Oliveira, the man who stabbed Bolsonaro, of being linked to the party of former President Dilma Rousseff, who was removed from office in 2016. Fake photos showed the attacker standing next to Lula. That Bispo had been affiliated with the left-wing Partido Socialismo e Liberdade (PSOL), and his lawyers’ refusal to say who was paying their fees only served to feed the conspiratorial claims.

At the same time, videos and messages that tried to undercut Bolsonaro gained traction on social media platforms. Some of the malicious content claimed the stabbing was staged, that Bolsonaro had actually been in hospital to treat cancer, and that the photos published showing the surgery had been forged.

The stabbing gave Bolsonaro a reason to withdraw from campaign activities, but earned him a better position in the polls. (Eventually, of course, Bolsonaro won the election.)

On Sept. 19, nearly two weeks after the attack, Eleições sem Fake, a WhatsApp group monitoring program created by the University of Minas Gerais, identified an audio recording that was making the rounds. The audio was shared by 16 of nearly 300 groups monitored by the project; some were Bolsonaro supporters.

That same day our organization, Comprova, began to receive, also by WhatsApp, requests from readers to verify the integrity of the recording.

In the audio, which was just over one minute long, an angry man with a voice that resembles that of Bolsonaro argued with someone appearing to be his son, Eduardo, and complains about being kept in the hospital. On the recording, the man said he can no longer stand “this theater,” suggesting that it was all an act.

That day, Bolsonaro was still a patient to the Semi-Intensive Care Unit at Albert Einstein Hospital in São Paulo. The medical report said he had no fever, was receiving intravenous nutrition, and had recovered bowel function.

Comprova could not find the original source of the recording. The audio primarily spread through WhatsApp at a time when files could still be shared in up to 20 conversations. This enabled it to spread rapidly, and soon make its way to other social networks. It became impossible to track it back to the original source. (WhatsApp has since restricted the number of groups you can forward a message to.)

Unable to identify the author(s) of the recording, Comprova focused on a more conventional investigation, and requested the help of an expert report from Instituto Brasileiro de Perícia (the Brazilian Institute of Forensics). Experts compared the viral recording with Bolsonaro’s voice in an April 2018 interview and concluded that the voice of the candidate was not the voice on the recording bring shared on social networks.

The experts made a qualitative analysis of the voice, speech and language markers of the man who spoke in the recording. Then they compared these parameters in each voice and speech sample. In this analysis, they investigated vowel and consonant patterns, speech rhythm and speed, intonation patterns, voice quality and habits presented by the speaker, as well as the use of specific words and grammatical rules.

For example, the below image shows a frequency analysis of “formants,” the name of the pitches produced by vibrations of the vocal tract, the cavity where the sound produced at the larynx is filtered. The air inside the vocal tract vibrates at different pitches, depending on its size and shape of the opening. The image shows a frequency analysis of the formants using the vowels “a,” “e” and “o.” The green vowels correspond to the audio sample we obtained on WhatsApp, and the blue vowels correspond to a sample taken from an interview given by Bolsonaro a few days before the attack on him.

Additional analysis found that the speaker in the WhatsApp audio was found to have a typical accent from the countryside of the state of São Paulo. But this did not appear in Bolsonaro’s speech patterns. Differences in resonance, articulation, speech rate and phonetic deviation were detected in the compared samples.

Comprova consulted a second expert. This professional also concluded that the voice in the recording differed from Bolsonaro’s for several reasons. He said the tone of the voice appeared to be a little more acute than Bolsonaro’s. He noted the pace of speech was also faster than another video recorded by the candidate at the hospital.

Another element that reinforced the conclusion that the audio was fake is the poor quality of the recording. According to experienced experts, this is a typical phony trick: Lowering the resolution of audios, videos and photos make analyzing them more difficult.

In terms of Bolsonaro’s response, his sons, Flavio and Carlos, posted on social media to say the audio was “fake news.”

If this audio went viral today, it would probably be harder to believe that the voice belonged to Bolsonaro. Before the election, with only 18 seconds a day on TV and his missing the campaign debates due to hospitalization and treatment, the current president’s voice was not so well known. That created an opportunity for a faked audio recording to fool many.

More than a year later, however, it is still difficult to understand why groups in favor of Bolsonaro or campaigning for his candidacy shared this audio, which, if proved authentic, could have destroyed his candidacy. We will never fully know why these groups so eagerly shared this content. Even so, it’s a powerful reinforcement of the fact that a piece of content that makes an explosive claim will spread rapidly across social media.