Nine Myths About AI in Journalism - From Someone Who’s Seen Them Up Close

By Sebastián Auyanet

Over the past year, I’ve had the privilege of coaching newsrooms across Latin America through the Google/Fathm AI Prototyping Sprint and several other programs — from large legacy outlets in Brazil and Mexico, to small independent digital newsrooms in Venezuela, Colombia, and Perú. I’ve watched teams build remarkable things. I’ve also watched well-intentioned projects stall, collapse, or never make it past a demo.

What follows isn’t theory. It’s a list of the myths, false narratives, and overpromises I keep encountering — and what I’ve actually seen happen instead.

1. “AI will replace journalists”

This is the myth that generates the most fear — and the least evidence.

In every newsroom I’ve worked with, the most successful AI implementations did the opposite: they freed journalists to do more human work. At one large Ecuadorian outlet, a tool that cut audio transcription time from two hours to fifteen minutes meant reporters could spend that recovered time on source development and story analysis. At an Argentinean multimedia group, a tool that converts live radio audio into publishable text didn’t eliminate the journalist — it meant a reporter covering a protest didn’t have to choose between being on the scene and filing on time.

The real risk isn’t replacement. It’s deskilling — what happens when journalists stop developing certain capabilities because AI handles them, and then the AI fails, changes, or gets cut from the budget. Managing that risk requires intentional adoption strategies, not fear of the technology itself.

2. “Implementation is fast and cheap”

This one gets newsrooms into trouble more than any other.

At one of the largest newsrooms in South Brazil — a strong cross-functional team with genuine leadership buy-in — compliance approval for a simple internal tool took two months. Not because anyone was obstructing the project, but because that’s how long it takes to move through the bureaucracy of a 2,000-person media company.

At the other end of the spectrum, smaller newsrooms face a different version of the same problem. When your entire technical capacity lives in one person, a sick day can freeze a project for weeks. I watched a major Colombian outlet’s project grind to a halt for exactly this reason — not because the tool wasn’t working, but because the one person who knew how to run it was out. Added to that, once you get past the prototype stage, the costs of APIs, tokens, and ongoing maintenance add up in ways that weren’t in anyone’s original budget, which might completely change the scope and the original goals of the project.

The organizations that navigate these situations nicely are the ones that plan for friction from day one — not the ones that assume the path will be smooth.

3. “Our archive is gold — AI will unlock it”

This might be the most seductive myth in the industry, and one of the most consistently disappointing in practice.

The pitch sounds compelling: decades of journalism, now queryable. The reality I’ve encountered is that most archives are disorganized, stored in incompatible formats, and lack the metadata that would make them useful for AI systems. The archive isn’t gold — it’s ore, and usually the refinery doesn’t exist yet.

The broader market has validated this the hard way. General-purpose chatbots built on large news archives — including high-profile experiments at major outlets — failed to gain traction.

What does work are narrow, focused RAG (Retrieval-Augmented Generation) systems built around a specific beat or topic. One Argentinean regional outlet built a RAG system trained on thousands of articles about a single energy region — and it became a genuinely useful tool for journalists covering that beat, reducing the knowledge gap for non-specialists and helping generate contextual drafts. The lesson: the narrower the scope, the more useful the tool.

4. “AI can automate social media distribution without editorial oversight”

Several projects I coached started with this ambition. None of them ended there.

The consensus that emerged — across coaches, newsrooms, and program coordinators — is that automating social media publishing without a human approval step carries real and underestimated risk. Tone errors, missing context, sensitivity failures — these aren’t edge cases. They’re the normal output of a system that doesn’t understand what it doesn’t know.

One Brazilian business paper built a strong MVP for social media content curation: the system reads recently published articles, selects pieces based on editorial guidelines, and generates platform-specific posts. But the workflow they landed on — and the one that made the tool actually usable — kept an editor in the loop to review and approve before anything went live. AI drafts. Humans decide.

5. “LLMs work equally well across all languages and contexts”

This assumption is false everywhere. In Latin America, it’s structurally damaging.

Let me show you an example. Guaraní is spoken by approximately 7 million people across Paraguay, Bolivia, Brazil, and Argentina. Those 7 million people cannot meaningfully use ChatGPT, Gemini, or any mainstream LLM. The same is true, to varying degrees, for Quechua, Aymara, and dozens of other indigenous languages across the region.

This isn’t a minor localization issue. It’s a structural bias baked into the training data and investment priorities of the dominant AI companies. El Surti, a media outlet based in Paraguay received funds to begin training a Guaraní language model — running community “mingas” (collective work events) where speakers record and validate audio using Mozilla’s Common Voice platform. The fact that this requires a dedicated grant-funded project to address a gap that affects millions of people tells you everything about how the default assumptions of the AI industry are built.

For newsrooms serving multilingual or indigenous-language communities, “just use the available tools” is not a neutral recommendation.

6. “With AI, editors no longer need editorial intuition — data decides”

In every project I’ve seen, the editor retains the final call. Not as a formality — as a genuine necessity.

But there’s a subtler version of this myth worth examining: the idea that because AI recommendations are data-driven, they’re neutral. They’re not.

One Mexican outlet built an editorial assistant that suggests story topics and recommends which journalist to assign based on historical performance data. During the showcase, someone raised a pointed question: wouldn’t a system trained on past performance systematically favor journalists with longer track records, and disadvantage newer voices? The team’s answer was honest — the editor maintains the final decision, and the metrics are meant to support intuition, not replace it.

That’s the right answer. But it requires newsrooms to be conscious of the bias risk, not just assume that data-driven means fair. A recommendation engine that reinforces the existing editorial diet isn’t neutral — it’s conservative by design.

7. “AI adoption is primarily a technical problem”

It isn’t. Not even close.

The biggest blockers I’ve encountered across dozens of newsrooms weren’t technical. They were organizational: leadership that hadn’t bought in, editorial teams that felt threatened, middle managers who didn’t see the point, and cultures that hadn’t developed the psychological safety to experiment and fail.

The projects that moved fastest shared one characteristic: a convinced editor or editorial supervisor who drove adoption from the inside. At one Ecuadorian outlet, a previously resistant editorial supervisor became the tool’s most vocal advocate — visiting the newsroom twice a day to request assistance — after seeing the time savings firsthand. That shift in one person changed the entire trajectory of the project.

The best tech stack in the world won’t save a project that doesn’t have that internal champion.

8. “A prototype equals implementation”

This is the gap that swallows the most promising projects.

Many AI-in-journalism programs — including well-funded, well-designed ones — end with a working prototype and a successful demo. What they don’t end with is a plan for what happens next: who maintains the tool, who updates the prompts when the underlying model changes, who trains new staff, who decides when to retire it.

I’ve seen this play out repeatedly. A newsroom builds something genuinely useful during a six-week sprint. Six months later, the tool is dormant — not because it failed, but because no one planned for the post-consultancy reality. The person who built it moved on. The API costs weren’t in the budget. The prompts stopped working when the model was updated and no one knew how to fix them.

Sustainability has to be designed from day one, not treated as someone else’s problem after the program ends.

9. “Newsrooms without AI policies are just behind the curve”

This framing makes it sound like a timing problem. It’s actually a governance problem — and a more serious one.

Without clear institutional criteria for what AI can and can’t do editorially, tools get adopted chaotically and individually. One journalist uses AI to draft social posts. Another uses it to summarize sources. A third uses it to generate story ideas. None of them are coordinating, none of them are documenting, and the newsroom has no way to audit, correct, or learn from any of it.

The risk isn’t just inconsistency. It’s that the newsroom loses the ability to make intentional decisions about its own editorial identity. AI policy isn’t a bureaucratic checkbox — it’s how a newsroom decides, collectively, what kind of journalism it wants to make.

What I’ve actually seen work

None of this is an argument against AI in journalism. It’s actually the complete opposite. The tools I’ve watched newsrooms build over the past year are genuinely impressive — and more importantly, genuinely useful. An editorial agenda assistant in Brazil. A live audio transcription tool in Argentina. A beat-specific RAG system in Patagonia. An editorial recommendation engine in Mexico. An audio transcription assistant in Ecuador. These are real tools, used by real journalists, solving real problems.

What they have in common isn't technology. It’s that they started with a clearly defined problem, kept humans heavily involved around, planned for the sustainability of any experiment, and had at least one person inside the newsroom who believed in them enough to push through the friction.

That’s not a myth. That’s just how change works.

Sebastián Auyanet is a media consultant and AI coach who has worked with newsrooms across Latin America through programs including the Fathm’s AI Catalyst for Wan-Ifra, the Fathm’s AI Prototyping Sprint for Google and several other programs for organizations based in Latin America and the US.