What Europe should demand from AI providers: a number for every language
The AI Act speaks extensively about transparency. For newsrooms in Central and Eastern Europe, transparency should start with something simple: how much the language they work in actually costs.
Andrei NistorJurnalist, fondator aistiri.eu · Story Agents · RomaniaTranslated from Romanian with AI assistance and reviewed by our editors.
As of August 2, 2026, the European Union will apply and enforce a new part of the AI Act obligations, including rules for labeling AI-generated content. On paper, Europe has the most ambitious artificial intelligence regulatory framework in the world.
In practice, things are moving slower. At the beginning of August, Romania had not yet adopted the national law specifying who controls and who sanctions. ANCOM was proposed as the market surveillance authority, but, as the institution itself explained, it will only be able to verify and sanction after the law enters into force. For providers of general-purpose models, the European AI Office has direct powers regardless.
I am not writing this to add another complaint about delays. I am writing it because, in the entire discussion about the AI Act, a dimension that is essential for the press in our region is almost completely missing: language.
What I showed in the two papers
In the first paper, I argued that language models do not function the same way in all languages and proposed a benchmark for this: Linguistic Media Equity. One of its dimensions is equity in governance: communities have the right to verify how these systems behave in their own language.
In the second paper, I measured part of the problem. The same news story requires 1.68 times more tokens in Romanian than in English, 2.01 times in Hungarian, and 1.83 times in Slovak, using one of the most widely used tokenizers. None of the eight tokenizers tested comes close to English.
All 24 official languages of the Union are equal in treaties. In the invoices of AI providers, they are not.
Four things Europe can demand
I don't think a new law is needed. I think it is necessary for the existing rules to be read through the eyes of an editorial office in Cluj, Košice, or Debrecen. Specifically, I would ask for four things.
- The first token per EU language. Any provider selling a model in the Union should publish how many tokens, on average, each official language requires compared to English, for texts of the same type. It is a single number. It can be calculated, verified, and compared between providers.
- Performance per language, not just the average. Providers often report aggregate results. For the press, the factual error rate in the language of publication matters. If a model makes more mistakes in Romanian, newsrooms need to know this before using it.
- The right to audit. Newsrooms, universities, and independent organizations should be able to test systems in their own language and publish the results, without depending on the goodwill of the provider.
- Investment in verification, not just automation. The part that is most often missing in small newsrooms is the human who verifies. Public money for media in smaller languages should go toward verification and media literacy, not just toward new tools.
Why it is a European problem, not a Romanian one
This gap does not stop at borders. Romanian is an official language in both Romania and the Republic of Moldova. Hungarian is spoken in Hungary, Slovakia, and Romania. A Hungarian newsroom in Transylvania pays the same token fee as one in Budapest. This makes the solution, inevitably, a coordinated one at the European level: the same measurements, the same reporting rules, the same market.
There is also an argument that Brussels understands very well: sovereignty. If a newsroom in Central Europe systematically pays more for the same tools, and those tools understand their context less well, the dependence on a few large providers becomes even more expensive. Models built in Europe reduce this gap, and my figures show it. But I am not closing the case.
A modest beginning
I am not asking for a revolution. I am asking for a published number for each language and the right to verify it. The data and code from my study are open, on Zenodo, at https://doi.org/10.5281/zenodo.23552863. Anyone can repeat the measurement on their own newsroom's texts.
If Europe wants the AI Act to mean something for the press in its smaller languages, it can start here. With a single number, which anyone can verify.
Sources
- Ziarul BURSA, „AI Act, a European law without «teeth» in Romania: authorities cannot yet apply sanctions”, August 4, 2026: https://www.bursa.ro/ai-act-lege-europeana-fara-8222dinti8221-in-romania-autoritatile-nu-pot-aplica-sanctiuni-inca-06767950
- Edupedu.ro, „Romania remains without a designated authority for the AI Act”, August 11, 2026: https://www.edupedu.ro/romania-ramane-fara-autoritate-desemnata-pentru-ai-act-in-timp-ce-alte-state-ue-au-deja-sanctiuni-operationale/
- A. Nistor, „The Token Tax in Central and Eastern European Newsrooms”, date and code: https://doi.org/10.5281/zenodo.23552863
Published under CC BY-ND 4.0: you may republish this article for free, unchanged, with credit to the author and aistiri.eu.
Republish this article
This article is free to republish under a Creative Commons licence. The rules:
- Keep the text unchanged: no edits, cuts or translations. You can use either the English or the Romanian version.
- Credit the author and aistiri.eu, with a link to the original article.
- If your site allows it, point the canonical link to the original article.