Translation errors in EU AI Act could misclassify model training scale

A discrepancy in translation

A technical discrepancy has been identified in the official translations of the European Union's AI Act. Specifically, the Portuguese and Croatian versions of the text misstate how the computation threshold for training AI models is measured. While the majority of the Act's versions are consistent, these two specific translations introduce a time-based metric that does not exist in the others.

As arxo.io reported, the Portuguese version includes the phrase "por segundo" and the Croatian version includes "u sekundi." Both of these phrases translate to "per second" in English. This addition changes the requirement from a measurement of total work to a measurement of speed, which creates a significant inconsistency across the EU's legal framework.

The discrepancy was found in Article 51(2) of the regulation. This article sets a specific threshold of 10^25 floating point operations for training. Floating point operations are the standard way to measure the amount of math a computer performs to train a large model. By adding a time unit, the two translations create a different legal standard than the other 22 official versions.

The math of the error

The error is significant because it changes the fundamental way a model's scale is calculated. The other 22 versions of the Act measure the cumulative amount of computation. This means they look at the total number of operations performed during the entire training process, regardless of how long that process takes to complete.

In the Portuguese and Croatian versions, the inclusion of "per second" means the threshold applies only to the rate of training. This creates a scenario where a model could be trained very slowly over a long period. For example, if a model requires 2x10^25 operations and is trained over one million seconds, its rate would only be 2x10^19 operations per second.

In this hypothetical case, the total operations would exceed the 10^25 limit, but the rate per second would not. Because the Portuguese and Croatian texts specify the rate, such a model might not trigger the same regulatory requirements as it would under the other 22 versions. This highlights how a simple translation error can lead to different legal outcomes for the same piece of technology.

Evidence from the text

The evidence for the correct interpretation comes from other parts of the AI Act. Recital 111 and Annex XIII, point (c), both support a measurement of the total number of operations. These sections do not include a time unit, which suggests that the cumulative amount of computation is the intended standard for the entire regulation.

The discrepancy was identified by comparing the official versions published in the Official Journal on July 12, 2024. The Portuguese and Croatian versions are the only ones that deviate from the standard. The other versions, including the English version, correctly describe the cumulative amount of computation used for training without mentioning a time unit.

The finding was confirmed by comparing the official text across all languages. The discrepancy is clear because the added words in the two translations do not fit the description of a cumulative quantity. A cumulative quantity represents a total sum, whereas a "per second" metric represents a rate of flow. This distinction is critical for determining which models fall under specific EU oversight.

Proposed legal corrections

To resolve this issue, a correction has been proposed to align the Portuguese and Croatian versions with the rest of the EU. The proposal is to delete the specific words that introduce the time element. This would ensure that all versions of the AI Act measure the same threshold in the same way.

The proposed correction involves removing "por segundo" from the Portuguese text and "u sekundi" from the Croatian text in Article 51(2). By removing these phrases, the translations will correctly reflect the cumulative amount of computation. This change ensures that the law applies consistently to all companies operating within the European Union, regardless of which language version is being consulted.

The report was sent to several key institutions to initiate the correction process. These include the Council of the European Union, the European Parliament, and the European Commission. Specifically, the report was directed to the Directorate for Quality of Legislation and the European AI Office, which is responsible for applying the regulation and preparing guidelines.

Impact on AI developers

For developers of large-scale AI models, this discrepancy could have created a confusing legal landscape. If a company were operating in Portugal or Croatia, they might have been able to argue that their model did not meet the threshold by training it at a very slow speed. This would allow them to potentially avoid certain transparency and safety requirements.

However, the identification of this error by researchers and the subsequent reporting to the EU authorities aim to prevent such a loophole. The goal of the AI Act is to ensure that models of a certain size and complexity are subject to oversight. If the law can be bypassed by simply slowing down the hardware, the intended safety goals of the regulation would be undermined.

The correction process is the next step in ensuring the AI Act functions as intended. Once the corrigendum is processed, the legal certainty for developers will be restored. It ensures that the 10^25 threshold is a hard limit on the total computational resources used, making it a more reliable metric for regulatory compliance.