EU AI Act Article 51(2): Two Language Versions Say "Per Second" — A Translation Error in the 10^25 Threshold

An independent register called Arxo says two official language versions of the EU AI Act set the model-risk compute threshold on the wrong quantity. In the Portuguese and Croatian texts of Article 51(2), the 10^25 figure is "measured in floating point operations per second," while the other 22 versions say the cumulative amount of computation used for training, according to the Arxo finding, published October 31, 2026.

It was shared on Hacker News with 5 points and no comments at the time of writing. The finding is small, checkable and quirky, which makes it a good example of how a single phrase in a regulation can carry weight for the whole AI industry.

TL;DR: the discrepancy

What does Article 51(2) actually say?

Article 51 of the AI Act decides when a general-purpose AI model counts as having "systemic risk", which triggers extra obligations such as evaluations, incident reporting and cybersecurity measures. Paragraph 2 sets a presumption: a model is presumed to have high-impact capabilities when the compute used to train it exceeds 10^25 floating-point operations.

The English text, quoted in the register, refers to "the cumulative amount of computation used for its training." That means a total: the sum of all operations over the whole training run. Public discussion of the threshold, including most analysis of which frontier models cross it, uses this meaning. For background on how the Act's obligations are being applied, see our post on what changes after EU AI Act enforcement.

What is different in the Portuguese and Croatian texts?

Arxo quotes the Portuguese phrase "medido em operações de vírgula flutuante por segundo" and the Croatian "iskazana brojem matematičkih operacija nad brojevima s pomičnim zarezom u sekundi", and gives its own English translations, "measured in floating point operations per second" and "expressed as the number of floating point operations per second." Those translations are the register's, marked as such.

Its argument is that "per second" does not fit with the rest of the same paragraph, which still speaks of cumulative computation, and that Recital 111 and Annex XIII, point (c), support a total count. In effect, the added words turn a quantity (total operations) into a rate (operations per unit time).

Why the difference is huge on paper

Arxo gives a worked example. Suppose a training run proceeds at a constant rate, and in total performs 2×10^25 operations over a million seconds. The total is above 10^25, so the cumulative test is met. But the rate is 2×10^19 operations per second, far below 10^25. Under the "per second" reading the model would not cross the line.

How big is the gap? Arxo's own example already shows a factor of a million between the total and the rate for a million-second run. A training run lasts weeks or months, so the total is many orders of magnitude larger than the per-second rate. For a literal rate reading to catch a model, a cluster would need to sustain 10^25 operations every second. We have no published hardware figure to cite for that, but it is many orders of magnitude beyond the cumulative totals discussed for frontier runs. The practical upshot is that a rate reading would regulate effectively no models, while the cumulative reading is the one that frontier-scale training runs are discussed against.

Why compute thresholds are written this way

The Act uses training compute as a proxy because it is measurable before release and hard to fake. Regulators cannot test every capability in advance, so they pick a number that correlates with scale and presume systemic risk above it. The Commission can update the number by delegated act, and providers can argue a model below the line still poses systemic risk, or that one above it does not.

That makes the unit the heart of the rule. "Cumulative operations" is a bookkeeping total a lab can report from its training logs. "Operations per second" describes a machine, not a model; two labs training the same model on different clusters would report different rates and the same total. A rule written in terms of rate would punish speed, not scale, which is the opposite of the legislators' intent as expressed in the recital.

This is also why the mistake reads as a translation slip rather than a policy choice. The added words conflict with the cumulative wording in the same sentence, and nothing in the other 22 versions suggests a rate was intended.

How translation errors get fixed in EU law

EU legislation is drafted in one language and translated into the others by lawyer-linguists before publication in the Official Journal. When a discrepancy is found later, institutions can adopt a corrigendum, which amends the affected language versions without reopening the law. The process takes time because it needs agreement between the Council and Parliament, and it must be published.

Arxo says it reported the Portuguese and Croatian discrepancy on October 31 to the Council's General Secretariat, the Parliament's legislative acts directorate and the Commission's AI Office. We have no information on how fast they will respond, and nothing suggests the AI Office plans to treat the threshold differently in the meantime.

Does it change anything in practice?

Almost certainly not immediately, for three reasons.

1. Regulators apply the cumulative reading. The European AI Office works from the Commission's guidance and the English text. Nobody serious is arguing that the threshold is a per-second rate.
2. The error is confined to two languages. The other 22 versions agree with each other.
3. A corrigendum procedure exists for exactly this type of slip, run by the lawyer-linguists of the Council and Parliament.

The residual risk is legal. All official language versions of an EU regulation are equally authentic, and a company in Portugal or Croatia that wanted to contest a systemic-risk designation could in theory point at its national text. Courts would likely read the provision in light of the recitals and annexes, as the register suggests, but the ambiguity is avoidable.

How to check it yourself

1. Open Regulation (EU) 2024/1689 on EUR-Lex and select the English version. Find Article 51(2) and read the threshold sentence.
2. Switch the language selector to Portuguese and then Croatian and compare the same paragraph.
3. Compare with Recital 111 and Annex XIII, point (c), which the register says support a total number of operations.
4. If you are a lawyer or translator, check the Official Journal PDF, not an unofficial site, because the register's claim concerns the official text.

The register publishes SHA-256 hashes of the files it used, so you can confirm you are comparing the same documents. We checked the claims against the register's page and its quotations; we did not independently diff the official Portuguese and Croatian files, so treat the finding as Arxo's until the institutions respond. Unofficial copies, such as artificialintelligenceact.eu, are English-only and cannot settle this.

What this means for AI developers and policy watchers

Developers. Keep tracking cumulative training compute, in total floating-point operations, not a rate. If you train in the EU, document how you count: which precision, which phases, and whether fine-tuning is included. Our write-up of the EU AI Office's formal probes shows that the Office is already engaging with specific labs.

Policy watchers. This is a reminder that a 24-language regime has a long tail of drafting risk. The threshold also sits at the center of a wider debate about whether a fixed compute number is the right proxy for risk, which several recent posts, including the EU's comments on agents, touch on.

Journalists and researchers. Compute thresholds are cited constantly. A quick check of the legal text in your reader's language is now worth the minute it takes.

What is verified and what is not

Verified from the register's page: the quoted Portuguese and Croatian phrases, the English quotation, the proposed correction, the stated recipients, and that the register reports the official text unchanged on October 31, 2026. Arxo also states that AI models helped read and compare texts and review its draft letters, which readers should weigh. Not verified by us: the contents of the 22 other versions, any official reply, and whether a corrigendum is under way. We will add an update if the institutions respond.

Related reading

• What changes after EU AI Act enforcement
• EU AI Office formal probes into OpenAI and Anthropic
• EU's Virkkunen on the AI Act and rogue agents
• China AI labs and safety disclosure
• Anthropic's compute commitments: claimed vs verified
• Hacker News discussion

Details reflect public information as of November 1, 2026.