Rio's 'Homegrown' AI Wasn't So Homegrown — Sovereign AI Meets the Benchmark Trust Problem

Governments everywhere are racing to build “our own AI.” So when Rio de Janeiro’s city government unveiled a homegrown language model, it fit the pattern perfectly — until allegations surfaced that the model was really a patchwork of existing open weights. This isn’t a one-off embarrassment. It lands squarely on the trust problem hiding underneath the entire sovereign AI movement.

Let me be upfront. The Nex-N2 story has not yet been heavily scrutinized by the community or widely reported. Public discussion over the past 30 days has been thin. So this piece is less about declaring guilt and more about a recurring question: why do these controversies keep erupting, and what should we actually be checking when a government says it built its own model.

The Sovereign AI Rush, and the Trap Inside It

Sovereign AI is the idea that a country — or a city — should build AI on its own data and its own infrastructure, rather than leaning on US Big Tech. The motivations are real: data sovereignty, local jobs, national security. The justifications write themselves.

The problem is that “we built our own AI” covers an enormous technical spectrum. At one end, you pretrain a model from scratch. At the other, you fine-tune someone else’s open model. And somewhere in between sits model merging — mathematically blending the weights of several public models into something that looks brand new.

That last category is exactly where the Rio allegations land. The claim is that Nex-N2 isn’t the product of original research, but a merge of already-public models with a fresh name stamped on top.

Why Model Merging Becomes a Problem

To be clear: model merging is not fraud. It’s a legitimate, widely used technique in the open-source world. You combine the weights of models with different strengths and aim for better overall performance. Nothing wrong with that.

The problem is how you describe it. If you announce “we built this from the ground up” while quietly merging someone else’s open weights, that’s no longer a technical question — it’s a question of honesty. Because the moment the method is misstated, every claim about the budget, the research team, and the development timeline starts to wobble too.

Then come the licenses. If the original models require attribution or restrict commercial redistribution, a merged model that scrubs its sources is a legal liability. For a taxpayer-funded city project, that risk is sharper still.

How Benchmark Scores Get Inflated

The recurring co-star in these controversies is benchmark gaming. A benchmark is essentially an exam that scores a model’s ability. And there are several ways to juice that score.

The most common is leaking the exam. Slip the test data used for evaluation into the training set, and the model walks into the exam having memorized the answers. The score looks great; it tells you nothing about real capability. This is called contamination.

Merging invites a similar trick. Cherry-pick models that have been tuned to ace specific benchmarks, blend them, and the scoreboard lights up. But when an actual resident asks the chatbot a real question, the answers often fall apart. That’s the gap between the score and the lived experience.

Government-Built AI: What to Actually Verify

When a private startup oversells, the market eventually corrects it. Government projects are different. Public money is on the line, and so is policy trust. That demands a stricter bar.

First, disclose the training lineage. What data, on top of which base model. Second, reproducibility — a third party should be able to reproduce the published benchmark scores under the same conditions. Third, real-world evaluation. What matters isn’t the scoreboard, it’s performance in actual user scenarios.

The value of the Rio case is that it forces these three questions into the open. Whatever the truth turns out to be, we’ve reached a point where interrogating the phrase “homegrown” is simply good hygiene.

Sovereign AI is an unstoppable trend. But declaring “we built it” is not the same as proving you can. The next time a government brags about its independently developed AI, what’s the first question you’ll ask? Checking the sources and the reproducibility before the scoreboard — that may be the new media literacy of the sovereign AI era.