During his first year in office, Argentine President Javier Milei gave 58 long interviews to media and delivered 84 official speeches, totaling 89 hours of audio and video, according to an investigation by La Nación.
After mapping all of Milei's speech in this period, the Argentine newspaper used artificial intelligence (AI) to transcribe and analyze the 739,000 words to identify patterns in the politician's language: from insults toward opponents, the State and the press, to phrases he repeats routinely. The result was the special report "Así nos habló Milei: 9 trucos del relato libertario para construir un nuevo sentido común" (This is how Milei spoke to us: 9 tricks of the libertarian narrative to build a new common sense).
“Milei became president without a political party, with very little institutional power and without a history,” Martín Rodríguez Yebra, executive editor and columnist at La Nación who signed the article, told LatAm Journalism Review (LJR). “His power rested particularly on his words. He is a leader very focused on communication, and we thought it would be very important to unravel the characteristics of his discourse: what makes it so unique, what his most common ‘tricks’ are, how he seeks to build an epic that allows him to overcome his political weaknesses. That's why we started collecting all of his speeches and interviews. We wanted to conduct an analysis based on objective, quantifiable data.”
The article reported nine main characteristics of Milei's speech, including the frequent use of insults — there were more than 4,000 disqualifying expressions present in 130 speeches and interviews — , the construction of the image of a savior in an inherited catastrophe and the recurring use of hyperbole.
“The most surprising thing is the number of insults and attacks on his rivals, the press and public figures who, while not his enemies, have expressed positions different from his,” Rodríguez Yebra said. “We knew this was common, but we didn't know we'd be able to identify more than 4,000 such attacks in just one year.”
The project also analyzed Milei's systematic attacks on the press. According to the Argentine newspaper's analysis, there were 410 pejorative mentions about journalism recorded in 69 speeches and interviews.
“Liar” was the word mentioned most in the ranking of disqualifications against journalism, according to the investigation.
The article also says that, over the course of a year, the president disqualified more than 60 journalists with offensive terms in a daily exercise in which he bombarded people with expressions such as "corrupt,” "miserable," "resentful,” "tyrants," "liars," "thugs" and "fakes."
“It was clear from the beginning that the systematic attack on the press was one of the essential characteristics of Milei's discourse,” Rodríguez Yebra said. “Never before had a president in Argentina been so verbally aggressive toward the media. Our goal was to determine patterns to understand this behavior: what kind of words he uses, when he is most aggressive and on what issues he reacts most aggressively. I believe the analysis allows us to better understand that this is a strategy, activated in certain moments of political necessity, and not a simple spontaneous reaction.”
La Nación team, from left to right: Florencia Rodríguez Altube, Martín Rodríguez Yebra, Martín Pascua, Gabriela Bouret, Pablo Loscri and Florencia Fernández Blanco (Credit: Disclosure/La Nación)
To decipher the president’s communication style, a multidisciplinary team of 14 professionals worked for about two months from the conception of the idea to the publication of the article. The team included a political analyst, an AI specialist, programmers, digital design team, video editors and a general editor.
“I think without AI, it would have been impossible to quantify the number of insults Milei uttered throughout his first year. Or to accurately determine how much hyperbole plagued his speeches,” Rodríguez Yebra said. “I think it's humanly impossible to work with dozens of hours of speech and be able to extract quantitative and precise information like this work has.”
The development process was divided into several well-defined stages, as explained by Martín Pascua, a developer at La Nación’s AI lab.
The first phase consisted of collecting the more than 100 videos of speeches and interviews Milei gave in his first year, and was led by the editorial team under the direction of Rodríguez Yebra. Then, the operations and DevOps team uploaded all that material to Google Cloud Platform (GCP).
“From there, we began evaluating the best technical strategy to process the videos,” Pascua told LJR. “We had prior experience from earlier projects, such as our live presidential debate analysis [October 2023], but this time we were working at a much larger scale, with longer files and a significantly higher volume. Also, nearly 12 months had passed between that project and this one, a long time in terms of generative AI evolution.”
Pascua said that before choosing the best AI models, the team ran controlled tests with a variety of options, including Google Speech, OpenAI’s Whisper API, Whisper running locally, and Gemini via Vertex AI. They ultimately chose Whisper Large v3 running locally for transcriptions and Gemini Pro for content processing and analysis.
“Our criteria balanced transcription quality, processing speed, and operational costs — especially considering we were working with over 100 speeches, many of them quite lengthy,” Pascua said. “Whisper Large v3 delivered the best results for our needs. It allowed us to avoid API usage costs while providing excellent transcription accuracy, particularly in speaker differentiation, which was crucial for discourse analysis. Running it on our own hardware also gave us full control over the process and the flexibility to tailor the environment to our needs.”
Gemini Pro, Pascua said, provided an excellent balance of quality, speed and cost. They used the tool to correct residual transcription errors, standardize the text and prepare it for structured analysis, explore different strategies for classifying and grouping relevant excerpts, and collaborate with the editorial team to iterate over different thematic angles.
After obtaining the transcripts of Milei's speech, Pascua said that the team built a Python script to clean up common errors from automatic speech recognition — like punctuation issues, speaker labeling and filler words. They also developed a small application using Streamlit, which allows the quick uploading and exploration of texts, to facilitate quality control.
Once all the transcripts were complete, the editorial team began identifying key thematic focuses. At the same time, the infographics and digital design team joined to begin transforming the findings into clear, impactful visual narratives.
“The design team worked with the video team on engaging ways to present the information,” Rodríguez Yebra said. “We tested several models until we arrived at the one we liked best, which allows us to see temporal evolutions of each variable of the speech and short video clips that summarize the idea.”
With all the key points of Milei's speech defined and validated, Pascua said the team did another round of processing to quantify its findings and deliver them in a structured format to the design team, which built the published visualizations.
“As a developer, this was one of the most enriching parts of the project: seeing how journalists work with large volumes of text, how they navigate, compare, and extract meaning was a true learning experience,” he said. “The team dynamic was one of the most rewarding aspects of the entire project. It was a true collaborative effort, with very different — but highly complementary — profiles working together with a shared goal. Personally, I think this kind of interaction — between developers, journalists, and designers — is one of the most enriching parts of working in a newsroom today.”
Despite the advanced use of technology, both professionals reinforced that human journalistic work was essential to guarantee the accuracy and contextualization of the information obtained by AI.
“Perhaps the most complex part of the work was verifying that the data AI extracted from Milei's speeches was accurate,” Rodríguez Yebra said. “We repeatedly reviewed the responses and ensured it [the AI] could interpret Milei's very specific language. We wanted to be very precise with the data. Afterward, we also had many discussions about how to present the information and what things we would leave out, for the sake of clarity.”
Pascua stressed that human review was an essential part of the process at every stage. The team implemented several mechanisms to ensure quality, including cleaning and correction scripts to fix common transcription issues and an internal review tool to make content easier to navigate. According to the developer, cross-team validation workflows were also created. The technical team reviewed the results structurally — including for formatting, completeness, processing integrity — and the editorial team reviewed them for meaning, tone, source accuracy and journalistic relevance.
“From the very first transcriptions to the final outputs, we worked in close coordination with the editorial team to ensure that the data remained consistent and true to the original content,” Pascua said. “This hybrid validation process — combining automation and human oversight — was key to delivering a robust and reliable result. It allowed us not only to correct technical errors, but also to fine-tune the editorial focus as we progressed through the material.”
La Nación’s project won the “Best Use of Artificial Intelligence in the Newsroom” award from the World Association of News Publishers (WAN-IFRA) in early May.
According to the award jury, the use of AI in La Nación's work allowed for a level of depth and precision that only the combination of advanced technology and journalistic rigor can provide.
“Its implementation clearly contributed to a more impactful and insightful form of journalism,” the jury said.
Rodríguez Yebra said he hopes the recognition of the project will serve as an incentive for editorial teams at La Nación and other outlets to make greater use of AI tools not only to analyze speech of authorities, but also for journalistic investigations in general.
“We are working to develop tools that allow us to perform analyses like this more simply and quickly,” he said.” These tools can be applied, for example, to a specific discourse and in real time. I think the possibilities are endless.”
For Pascua, from a technical perspective, the role of developers in newsrooms is evolving rapidly. The developer believes that it is no longer just about “supporting” the editorial team with tools, but rather being actively involved in the investigation itself, proposing new ways to collect, process and tell stories with data. In a world where most information is digital, massive and constantly changing, he said collaboration between developers and journalists is essential to conduct deep and impactful investigations.
“The ability to apply language models, scraping techniques, automation, interactive visualizations or real-time verification opens up storytelling possibilities that were unthinkable just a few years ago,” Pascua said. “That’s why I believe this role will not only continue to grow but also become increasingly specialized — with developers who understand editorial needs, journalists who are technically skilled, and hybrid teams that speak a shared language. In my opinion, that’s where some of the most promising futures of investigative journalism lie.”
