Imagine listening to a podcast about a story you wrote yourself and slowly realizing that artificial intelligence is telling a different version of your own story.
That is exactly what happened when some journalists tested NotebookLM, a Google tool that turns written articles into AI-generated podcast episodes. The idea is compelling: make news more accessible for people who learn better by listening than by reading. But what reporters actually found in practice was a curious mix of impressive, unsettling, and at times genuinely concerning results.
The AI sounded way too human in the first few minutes, then started making up details, dramatizing calm stories, ignoring entire sections of articles, and even changing the tone of pieces that had a very specific sense of humor. All of this raises a question that goes way beyond the technology itself: when a tool decides what matters in a news story, it brings along its own criteria, limitations, and of course the biases baked into the model. And that rabbit hole goes deeper than you might think. 🎙️
When the AI decides to tell the story its own way
NotebookLM was built to work as a kind of smart reading and content organization assistant. Among its features is the ability to take a text, whether it is a newspaper article, an investigative report, or a short note, and turn it into a podcast episode with two virtual hosts chatting about the topic. The voice is smooth, the pacing is dynamic, and for the first few minutes you almost forget you are listening to a machine. It is technically impressive, no way around it.
Worth noting that this is not the first experiment of its kind. Google had recently put to the test converting written stories into TikTok-style videos, and this new round of NotebookLM podcast tests only confirms a pattern. Journalists described the results with words like surprising, bland, and explainer-like. In other words, the technology delivers a slick package, but the content inside does not always match what was originally reported.
The problem starts when the tool goes beyond simply narrating what was written. Journalists who tested the system reported situations where the artificial intelligence added information that was not in the original text, created contexts that did not exist in the story, and in some cases completely changed the tone of the piece. A story written with irony turned into a heavy, serious analysis. A balanced report on a sensitive topic got a layer of drama the reporter never intended. And the wildest part: the AI did all of this with a narrative confidence that made it really hard to tell where the original content ended and the invention began.
This is not an isolated bug. It is a structural feature of how large language models work. They do not read a text the same way a human does. They identify patterns, associate concepts, and reconstruct a version of the content based on what they learned during training. When that process hits a gap, the model fills it in with whatever makes the most statistical sense, even if that means inventing a detail or reinterpreting an editorial intention the journalist spent hours carefully building.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
How the AI decides what makes it into the podcast
According to Parsa Hejabi, a PhD student in computer science at the University of Southern California’s morality and language lab, audio can be a really useful tool. But transforming an article into a different format using artificial intelligence can introduce bias, emotion, or communication failures. And that happens because the whole process is what he calls a pipeline, a sequence of stages where each phase has its own way of failing.
In general terms, the AI starts by extracting the main information from the source material. Then it filters the data and prioritizes the points it considers most relevant. In the third stage, the narrative construction kicks in, when selected excerpts become a script with transitions, explanations, and conclusions. Finally, text-to-speech technology produces the voice of the host or narrator.
The catch is that each stage can generate distortions. The extraction phase can miss important information. The prioritization phase can give too much weight to one perspective or highlight the wrong details. And the narrative construction phase can invent new connections or toss in irrelevant metaphors. As Hejabi put it, the voice, music, pacing, or visual elements can introduce an emotional charge that simply did not exist in the original material. And that extra dimension is precisely what opens the door to bias and loss of accuracy.
What journalists actually found in practice
The reports from journalists who tested NotebookLM illustrate these problems really well. Gabriela Flores, a freelance reporter for The Wave, said she was startled at first by how human the voice sounded, with pauses and little laughs sprinkled throughout the conversation. But it only took a few minutes for her to realize it was a machine. The podcast did a good job explaining the first half of her story about the New York City mayor’s approach to two major infrastructure projects, but ended up focusing on only one of them and adopting the perspective of the protesters.
Worse: the AI claimed that a new train line would connect a geographically isolated area, which is simply not true since the Queens neighborhood in question is actually well connected. The tool also used exaggerated expressions, like describing the mayor’s decision as pouring concrete over a time capsule, inventing motivations that were nowhere in the article. For Flores, it is a slippery and dangerous situation to let AI make its own guesses about someone’s intentions.
Miranda Dunlap, an education reporter for Open Campus, had a similar experience. Her story about a community affected by flooding took on a tone of constant shock throughout the 18-minute audio, along with tangents she never wrote, like one about disaster insurance being a scam. The AI also added information about flood mechanics and urban development that came from outside the original text. Her biggest frustration, she said, is that the time saved is an illusion since she would need to go through everything again to verify whether each claim is actually correct.
Emanuel Maiburg from 404 Media mainly criticized how long the podcast took to get to the point. In written journalism, the reader gets the gist of the story right from the headline, but the audio did not arrive at the main subject until four minutes of empty commentary and roundabout explanations had passed. He believes this happens because the tool is trained on popular podcasts full of chitchat, which does not make sense from an informational efficiency standpoint. And his story, which had a very specific sense of humor about a Vatican mascot, turned into a piece that ridiculed the church, completely changing the original tone.
Not everything is a problem: when the AI gets it right
Interestingly, not every experience was negative. Jessica Craig, an epidemiologist and health reporter, was impressed by how the podcast summarized a complex scientific topic about how parents make vaccination decisions for their children. According to her, the tool did not oversimplify or try to push the listener toward a specific conclusion, respecting the different perspectives she had intentionally included in her story.
Craig even praised some metaphors the AI created to explain technical concepts, like comparing the immune system to a sleepy security guard that can be woken up by an alarm present in vaccines. It was a creative and educational way to translate something complicated. Still, she admitted that hearing someone who does not exist speaking with such familiarity about her own work felt strange and hard to process.
Accuracy and bias: two sides of the same coin
The question of accuracy in content generated by artificial intelligence is already a long-running debate in the tech world, but it takes on a whole new dimension when we talk about journalism. In other fields, a fact invented by an AI can be caught and corrected before it causes harm. In journalism, wrong information delivered in podcast format, with a pleasant voice and an engaging narrative, has enormous persuasive power. People tend to trust what they hear more than what they read, especially when the audio sounds natural and well produced.
But factual inaccuracy is only half the problem. The other half is bias. Every language model is trained on a massive amount of data, and that data reflects the prejudices, preferences, and gaps of the real world. When the AI decides that one part of a report is less important than another, it is not making that call based on conscious human editorial criteria. It is doing it based on patterns it learned, which often prioritize dramatic, conflict-driven, or emotionally charged content simply because that type of content shows up more in the training data.
The practical result is that the AI-generated narrative can distort the reality of a story in subtle but significant ways. A piece about a successful public policy can turn into a story about red tape and roadblocks, because conflict sells better than success. A carefully reported interview with multiple perspectives can get boiled down to a single viewpoint, because the model identified that one as the dominant perspective. These slip-ups are not intentional, but they are systematic. And systematic is much harder to detect and fix than accidental. 🤖
What this means for people who get their news from podcasts
For the average listener who might not even know that episode was generated by artificial intelligence, the experience can be pretty misleading. The voice sounds human, the content seems organized and well-researched, and there is no clear indication that parts of that conversation were literally made up by the model. This puts the consumer in a vulnerable position, especially at a time when misinformation is already a massive problem in the digital media ecosystem.
Transparency around the use of AI in journalism is a point still being negotiated among news outlets, platforms, and regulators around the world. Some publications have already adopted mandatory disclosure policies whenever a text or audio has gone through automated generation. Others still operate in a gray area where AI is used as a support tool but the line between assistance and authorship keeps getting thinner. The automatically generated podcast is a clear example of where that line can disappear entirely if there are no well-defined standards in place.
Beyond that, there is the direct impact on journalists themselves. When an AI takes your work and re-presents it in a different form, with details you did not write and a tone you did not choose, it raises serious questions about authorship, editorial integrity, and accountability. If someone listens to that podcast and makes a decision based on information the AI invented, who is responsible? The journalist whose text served as the source? The company that built the tool? The platform that hosted the content? These questions still do not have clear answers, and while the debate stalls, the technology keeps being used at scale. 🎧
Automated narrative and the limits of what AI actually understands
One of the most fascinating and at the same time most problematic aspects of this discussion is what artificial intelligence simply cannot capture in a narrative. Quality journalism is built in layers: what is written, what is between the lines, the cultural context, the timing of publication, the intent behind the reporter’s word choices. A good piece of journalism carries editorial decisions in every comma. The AI, no matter how sophisticated it is, does not have access to those layers. It sees the surface and tries to reconstruct what it imagines lies underneath.
This becomes even more obvious in stories that rely on humor, irony, or sarcasm. These devices are deeply cultural and contextual. A joke that lands with one audience can sound offensive or just flat to another. An AI trained primarily in English, for example, has real trouble capturing the nuances of Brazilian humor, which is extremely regional, historically layered, and packed with references that do not exist in other languages. When NotebookLM takes a humorous report and turns it into a serious podcast episode, it is not carelessness — it is a structural limitation of the model.
And narrative accuracy also takes a hit when the model tries to simplify complex content for a more dynamic audio format. Technical explanations, statistical data, and arguments with multiple variables frequently get flattened into simple sentences that lose a lot of the original depth. For a listener who already knows the topic, it can feel superficial. For someone encountering that subject for the first time, it can create an incomplete or even misleading understanding of reality, without any way for them to notice. This might be the most silent bias of all: the kind of simplification that looks like clarity but hides complexity. 🔍
What listeners should keep in mind
At the end of the day, the most important piece of advice from researchers who study these systems is simple and worth its weight in gold. According to Hejabi, it is essential to remember the first stage of the process, the one where the model chooses what it considers important. Those points may not be as relevant as they seem, because every model carries its own weights, values, and priorities that vary across companies, fields of knowledge, and products.
He also emphasizes that generated content can sound plausible overall but still contain inaccuracies. That is why verification is the most critical step any user should adopt, no matter how much they trust the system or how pressed for time they are. Blindly accepting AI output is exactly what should not happen. In other words: use the tool as a starting point, but never as the final word. Critical curiosity is still the best defense we have against misinformation, even when it shows up with a friendly voice and a polished script. 😉
