Publish a plenary session on the portal of transparència It may seem like a simple process. But before making it public, it must be reviewed and any personal data that should not be published must be anonymized. This requires a decision, on a case-by-case basis, as to which information can remain visible and which needs to be protected. And this decision does not always depend on the type of data, but on the context in which it appears.
This task can involve a significant manual review burden and, at the same time, a risk of error that could end in the accidental publication of personal data. For this reason, at the AOC we have explored how generative artificial intelligence can help anonymize in an agile and reliable way, interpreting the content of the documents and automatically proposing which information needs to be anonymized, always with the final supervision of a person.
Anonymization that understands context
The main differential value of the developed solution is its ability to interpret the context of the document, beyond detecting data based on predefined patterns or entity recognition systems.
This allows, for example, to distinguish between the name of an elected official, which must be kept public in a plenary session, and that of a natural person who appears in another context and who may require anonymization. Similarly, a date may be part of institutional information that must be published or be personal data that must be protected.
Pattern-based or entity recognition solutions remain useful for identifying certain types of data. However, they cannot always determine whether a particular piece of data should be anonymized because this decision may depend on its function within the document. The AOC proposal seeks to reduce this intervention by combining quality in detection, contextual intelligence and automation, without losing human control over the final decision.
AI with human supervision
The process is simple: the user uploads the minutes, selects the anonymization criteria and starts the analysis. The system interprets the document and generates an anonymized proposal. The user then reviews the masks, modifies them or adds new ones if deemed necessary and downloads the document ready for publication.
This human supervision model allows the automation capacity of AI to be exploited without replacing the legal criteria of the administrations. The platform also incorporates whitelists to avoid over-anonymization and traceability and explainability mechanisms that allow the criteria applied to be reviewed.
An overall accuracy of 96%
The proof of concept has been validated with 52 documents provided by 9 municipalities, in addition to several internal tests by the AOC. The results show a 96% overall accuracy, that is, that the vast majority of the data that the system proposes to anonymize is correctly identifiedThe F1 indicator, which combines accuracy and coverage (the ability to detect all data that needs to be anonymized), exceeds 90%.
These results confirm the technical and functional feasibility of the proposal and the potential of combining generative AI and human supervision to reduce the workload and the risk of accidental disclosure of personal data.
Towards a future shared service
The test has also allowed us to identify areas for improvement, especially regarding detection coverage in complex documents and the scalability of the solution.
The PoC has now been completed and has validated the potential of the approach. The continuation of the initiative is currently being contracted, through a pilot with a wider scope that will allow the solution to continue to be improved and progress towards a possible shared service for local administrations.
