New European Union requirements intended to protect the public from manipulation involving artificial intelligence (AI) are to be rolled out in Estonia. The regulations aim to help people identify when they are encountering AI-generated content or interacting with an AI system, and include measures to combat "deepfake" images. Fines of up to €15,000 may be issued in the case of violations.
The** new EU Guidelines **on transparency obligations for providers and deployers of certain AI systems took effect this month and are aimed at harmonizing requirements across sectors and defining when users must be informed about AI use or content.
"It may not be obvious when AI presents itself in the name of a person, or in a human-like manner, without a clear indication. The user may mistakenly come to believe they are communicating with a real person, or the AI may be integrated into a service in a way that is not visible to the user," said Aivar Pau, strategic communications adviser for digital affairs at the Ministry of Justice and Digital Affairs.
In practice, the presence of AI may be "obvious" or self-evident in some cases: For example, when a user enters an environment that is clearly presented as an AI chatbot, or an AI character, or when the name and design of the design of the service clearly indicate AI is used. In these cases, separate labeling is not required under the new EU rules.
Pau noted that individuals must also be informed when they encounter AI systems capable of processing biometric data, identifying or inferring a person's emotions or intentions, or categorizing them by gender, age, appearance, ethnic origin, personal preferences or by other definitions.
The regulations are the first of their kind when it comes to cross-sector AI, Pau noted; previous regulations were sector- or field-specific.
The justice ministry said it takes a positive view of the transparency obligations, as they ensure individuals are informed and reduces the risk of manipulation.
As well as AI systems being required to indicate that the user is not interacting with a human being, specific labeling is also required for deepfakes: Content created or altered using AI and which could mistakenly be perceived as authentic.
Notification is also not required when the use of AI is permitted by law for the detection, prevention, investigation or prosecution of criminal offense, Pau noted, adding: "The condition here is that safeguards must be implemented to protect the rights and freedoms of third parties," Pau added, "except when people use these systems to report crimes."
Exceptions will also apply to edited texts – for instance readers do not have to be separately informed about the use of AI if this was used in preparing a published text, provided authors, editors or others responsible have reviewed that text. In such cases, a "natural person," meaning a legally defined person or persons, bears editorial responsibility for publishing the AI-generated content, the ministry said.
Any required AI information must also be provided in a format accessible to people with disabilities.
The government will only be able to formally delegate responsible authorities for the AI labeling, monitoring, supervision and penalties once the Riigikogu has transposed the EU regulations into domestic law.
To indicate the use of AI, the European Commission has developed three visual labels: A basic icon, a label for content generated entirely by AI and a label indicating part-use of AI. Under the new rules, these labels must remain attached to the text throughout the entire creative process. The labels are available on the European Commission's ** digital policy hub**.
The new requirements apply both to providers of AI systems and to their users. Violations can result in fines of up to €15,000 or 3 percent of a company's total worldwide turnover in the preceding financial year.
Previously, the labeling of deepfakes was only done on a voluntary basis per standards developed by the tech industry – standards not applied uniformly worldwide.
-- Follow ERR News on Facebook and Twitter and never miss an update!
Editor: Andrew Whyte