The probe is examining whether the chatbot implemented effective safeguards to stop users generating sexualized images.
X is expected to face a preliminary decision by the end of June in Brazil’s probe into Grok AI over sexualized deepfakes, as the country’s data protection authority reviews the company’s compliance with local rules. The investigation has taken on added importance after the company submitted largely self-declaratory evidence. The Brazilian probe is examining whether the chatbot implemented effective safeguards to stop users generating sexualized images.
A Brazilian investigation into Grok AI is expected to reach a preliminary decision in two months’ time, the data protection regulator responsible for the probe has told MLex.
Grok, an AI chatbot that operates both as a standalone tool and integrated within the social media platform X, has drawn scrutiny from regulators across the world this year for its use to create non-consensual, sexualized deepfake images using photos of real people.
One of the many authorities currently investigating whether X and Grok have complied with online safety and data privacy regulations is Brazil’s Agência Nacional de Proteção de Dados.
Jorge André Fontelles de Lima, the ANPD’s general coordinator of inspection, said a preliminary decision and assessment of the documents provided by the company is expected “within this semester.”
The ANPD probe has taken on added importance after X’s response relied largely on self-declaratory submissions, with limited evidence to back its claims. Brazilian authorities have previously raised concerns that the company failed to provide technical evidence.
Fontelles de Lima pointed out that the ANPD has hired additional staff since the investigation was opened, and that Grok's case is being handled as a priority.
“We issued a preventive measure requiring them to take sufficient and appropriate steps to prevent users from using their product to generate sexual images of individuals, especially women,”
he said.
The preventive measure, issued in February with the Federal Public Prosecutor’s Office and National Consumer Secretariat, ordered the company to implement safeguards across all versions of the tool to stop the creation of sexualized content involving identifiable individuals without consent, and to remove existing material.
“We also required them to submit documentation demonstrating the measures adopted and the effectiveness of those measures, because under the General Data Protection Law, it is the company’s responsibility to show it has taken appropriate steps, in line with the principle of accountability,” Fontelles de Lima said. “At this point, we are analyzing these documents to verify whether the platform has in fact implemented these safeguards and whether it can once again be considered suitable for the public.”
Fontelles de Lima also said the ANPD is looking into Grok’s compliance with its own terms of service.
“What drew our attention is that the possibility for users to produce this type of content was against the platform’s own terms of use,”
he said.
International outrage
The use of Grok to generate sexualized deepfakes caused outrage around the world, with regulators responding with different probes and a patchwork of enforcement approaches.
While the full scale of the issue is still unknown, the Center for Countering Digital Hate estimated that Grok generated around 3 million sexualized images in just 11 days, mostly of women, and including thousands that involved children.
In the UK, Ofcom and the Information Commissioner’s Office have opened parallel investigations examining online safety compliance and data protection issues linked to Grok’s image generation. Both are still ongoing.
Across the EU, authorities have taken a more procedural approach under the Digital Services Act, while in the US and Canada regulators have begun exploring investigations into the large-scale creation of non-consensual deepfakes. Other jurisdictions have taken more drastic action, with countries such as Indonesia and Malaysia temporarily blocking access to the tool.
Source: MLex
