Commissioner-initiated complaints concerning X Corp.’s and X.AI LLC’s compliance with PIPEDA

PIPEDA Findings #2026-004

June 11, 2026

---

Overview

1. In late December 2025 and early January 2026, multiple news organizations reported that the artificial intelligence (“AI”) chatbot Grok had generated, and publicly disclosed, millions of sexually explicit deepfakes of real and identifiable individuals. Following the growing concern, on January 15, 2026, the Privacy Commissioner of Canada (“OPC”) initiated two complaints against X Corp., the operator of the platform X, and X.AI LLC (“xAI”), the developer of the chatbot Grok (together “the respondents”), pursuant to subsection 11(2) of the Personal Information Protection and Electronic Documents Act (“PIPEDA”).
2. Specifically, the investigation sought to determine i) whether X Corp. and xAI had obtained valid consent from individuals for the collection, use, and disclosure of their personal information to create explicit, sexualized deepfakes; and, ii) whether a reasonable person would consider the collection, use and disclosure of personal information for the purpose of an image generation service that is capable of producing explicit, sexualized deepfakes to be appropriate in the circumstances.
3. The OPC found that the two organizations had not obtained valid consent to collect, use, and disclose individuals’ personal information for the purpose of generating sexualized deepfakes, and that a reasonable person would consider this practice to be inappropriate in the given circumstances. Furthermore, while the respondents represented that they had performed a number of actions to address the concerns that Grok was being used to generate sexualized deepfakes (including improving their technical and organizational safeguards), the OPC found that their response to the widespread generation of such content was insufficient.
4. Therefore, with a view to bringing X Corp. and xAI into compliance with PIPEDA, the OPC shared its preliminary findings and made a number of recommendations to the respondents.
5. In their response to those preliminary findings and recommendations, X Corp. and xAI disagreed with our conclusions. They indicated that Child Sexual Abuse Material (“CSAM”) and Non-Consensual Intimate Imagery (“NCII”) have no place on their platforms and asserted that they are taking all reasonable steps necessary to prevent their occurrence and taking immediate action when such content comes to their attention.
6. While the OPC recognizes that X Corp. and xAI have now deployed a layered set of safeguards to reduce the risk that their tool will be misused to produce sexualized deepfakes, the OPC finds that the respondents have not, to date, demonstrated the effectiveness of these safeguards in fully mitigating this issue. In these circumstances, the OPC cannot conclude that the complaint is resolved at the time of writing this report. Consequently, the OPC considers this matter to be well-founded (and not resolved at the date of this report).
7. While the OPC is encouraged by the respondents’ commitments to implementing a number of privacy-protective measures in response to our recommendations, we will continue to monitor the implementation of these commitments to ensure that the serious issues highlighted in this report are fully addressed.

Background

Corporate structure

8. xAI is an AI company which is registered in the United States and the developer of the generative AI chatbot “Grok”.
9. X Corp. is a technology company responsible for the provision of the social media platform X to Canadian users. It is registered in the United States and has a registered company in Canada (X Internet Canada ULC).
10. Both X Corp. and xAI have holding companies (X Holding Corp. and XAI Corp. respectively), which on March 28, 2025, became wholly owned “sister” subsidiaries^{Footnote 1} under X.AI Holdings Corp. On February 2, 2026, X.AI Holdings Corp. and Space Exploration Technologies Corp. (SpaceX) effected a transaction pursuant to which SpaceX became the direct parent company of X.AI Holdings Corp. The xAI and X Corp. companies’ corporate structures have remained unchanged.

Grok

11. Grok was released in November 2023. It is a conversational generative AI service that can generate content in response to various types of user prompts^{Footnote 2}. It can answer questions, brainstorm ideas, write code, and create animations or images.
12. Grok is powered by a foundational Large Language Model (“LLM”). LLMs are extremely large, complex machine learning systems capable of routinely generating elaborate, plausible-sounding – but not necessarily factually accurate – content in response to queries on virtually any topic. Grok combines multiple specialized models or components to process different types of data, such as text, images, audio and video, integrating them into a single system.
13. It can be accessed via:
    1. “Grok on X”, a chatbot functionality integrated within the X platform and provided by X Corp. Grok on X enables users with an X account to engage with the chatbot within the X platform;
    2. the @Grok account, operated by xAI. The xAI company represented that “in many ways […] it is like any other user account on X”. Specifically, users can interact with @Grok within a thread on the X platform and @Grok responds on the thread; and/or;
    3. the standalone Grok app and website (“Standalone Grok”) operated by xAI. The Standalone Grok may be accessed independently and provides certain additional functionalities, such as configuration options and interaction features, which are not available via Grok on X and @Grok. An xAI account is needed to use the standalone app, although users may use the single-sign-on option to link their X and xAI accounts.
14. Grok on X, @Grok and Standalone Grok are therefore distinct offerings that differ in terms of technical integration, available features and user experience.
15. Whereas the content generated using @Grok on a public thread^{Footnote 3} is public by default, the content generated through Standalone Grok and Grok on X can only be accessed by others if the user shares a specific URL with them.

Timelines regarding the generation of non-consensual sexual content

16. Grok was initially upgraded with image generation capability in August 2024. At the time, the image generation feature was powered by a third-party AI model, and was only available on Grok on X.
17. In December 2024, xAI updated Grok’s image generation capability with a new in-house AI model named Aurora, which it made available on both Grok on X and Standalone Grok.
18. In March 2025, xAI rolled out image generation capabilities on the @Grok account. Specifically, users could now instruct the @Grok account on the X platform to generate images from a text prompt. Approximately one month later, the function was expanded so that users could also instruct the @Grok account to edit images.
19. In July 2025, xAI replaced the Aurora model with an enhanced image generation and editing model called Grok Imagine with improved image editing capabilities, particularly with respect to ensuring that the facial representation adheres to the original source image (i.e., that the original individual’s facial likeness is preserved in the edited image output by the model). Grok Imagine was made available within Standalone Grok, Grok on X and the @Grok account.
20. On December 23-24, 2025, xAI rolled out a feature that facilitated users’ ability to edit images on the X platform. Specifically, users could now click the “Edit Image” button on images in X, including images of other users. Clicking this button would redirect the user to the Standalone Grok experience at grok.com, where the image editing functionality would be provided. The user could then download the content generated by Grok (i.e., image generated or edited) or post it on the X platform.
21. In the days that followed, as the feature became more widely used, X Corp. and xAI represented that “adversarial users” began purposely circumventing safeguards and generating content that violated their policies. Specifically, these users began using seemingly innocuous terms like “paint” or “floss” to prompt @Grok to generate explicit, sexualized deepfakes of identifiable individuals – including CSAM or NCII.
22. As described in more detail in relevant sections of this report, xAI and X Corp. stated that they began investigating the generation of non-consensual sexual imagery on December 28, 2025, when they became aware that Grok was being used to generate such content. On that same day, the @Grok account posted an apology regarding the generation and sharing of an “AI image of two young girls (estimated ages 12-16) in sexualized attire based on a user’s prompt.”^{Footnote 4}
23. Given the significant risks of harm associated with sexualized deepfakes, as explained below, and the fact that vulnerable populations were particularly targeted (e.g., children and women), the OPC was satisfied that there were reasonable grounds to investigate X Corp. and xAI.
24. Therefore, on January 15, 2026, the OPC initiated two complaints against X Corp. and xAI, pursuant to subsection 11(2) of PIPEDA.

Impact of sexualized deepfakes

25. In response to the OPC’s request for an estimation of the breakdown of harmful or non-consensual deepfakes that Grok had generated of both minors and adults:
    1. X Corp. represented that as of March 6, 2026, based on reports from users in Canada, it removed 126 posts that it had identified as containing CSAM or NCII material on the X platform.
    2. xAI submitted that it was unable to provide a similar breakdown requested for Standalone Grok, explaining that it did not have a reliable method to do so.
26. According to certain news reports, by January 8, 2026, Grok had, worldwide, generated over six thousand sexualized deepfakes every hour^{Footnote 5}, and an estimated 1.8 million sexualized images since December 29, 2025.^{Footnote 6} The nonprofit Center for Countering Digital Hate estimated that between December 29, 2025, and January 8, 2026, Grok generated approximately 3 million sexualized deepfakes, including more than 23,000 images of children.^{Footnote 7} The nonprofit entity named AI Forensics analyzed over 20,000 images generated by Grok between December 25, 2025, and January 1, 2026. It found that over half of the images contained individuals in minimal attire, and that 2% of the images (400) depicted individuals that appeared to be 18 years old or younger. By January 13-14, 2026, close to 10% of the sampled images generated by Grok contained individuals in minimal attire.^{Footnote 8}
27. The OPC does not take a position as to the precise volume of NCII and CSAM on the respondents’ platforms but, based on the data available in the public domain and the fact that the respondents did not deny this fact, we conclude that these images were present.

Methodology and Scope

28. The investigation sought to examine whether:
    1. xAI and X Corp. obtained valid consent for the collection, use and disclosure of personal information for the purpose of generating sexualized deepfakes, pursuant to Principle 4.3 of PIPEDA; and,
    2. a reasonable person would consider the collection, use and disclosure of personal information for the purpose of an image generation service that is capable of generating sexualized deepfakes, to be appropriate in the circumstances.
29. Given the specific and considerable risks associated with sexualized deepfakes (detailed below), our investigation focused on these rather than deepfakes in general.^{Footnote 9}
30. Over the course of the investigation, the OPC considered information from a variety of sources, including (i) written representations that the respondents provided to the OPC and (ii) information that the OPC gathered and analyzed from publicly available sources concerning the issues relevant to the investigation.
31. While the OPC assessed the compliance of each respondent individually, our findings in this report were equally applicable to each respondent. Therefore, in the specific circumstances of this investigation – which in our view involved a pressing matter that warranted proceeding expeditiously – the OPC combined the analysis of the issues with respect to each respondent and issued a single report.

Analysis

Jurisdiction

32. Paragraph 4(1)(a) of PIPEDA states that the Act applies to “personal information that the organization collects, uses or discloses in the course of commercial activities.” PIPEDA defines commercial activity in part as “any particular transaction, act or conduct or any regular course of conduct that is of a commercial character.”
33. PIPEDA applies to organizations outside of Canada where a “real and substantial connection” to Canada exists. In our view, the circumstances in this matter demonstrate that a real and substantial connection to Canada exists. In coming to this conclusion, the OPC considered the following factors:
    1. The respondents offer their services in Canada and have monthly active users, including paid subscribers in Canada. Specifically, X Corp. generates revenue through advertisers and subscription-based services. The xAI company also generates revenue through subscription-based services, as well as from providing developers with access to the Grok Application Programming Interface (“API”) for a fee. Importantly, over the course of their commercial activity, both organizations collect, use, and disclose personal information of individuals who use Grok on both the X platform and Standalone Grok.
    2. Users located in Canada can visit the respondents’ websites and use their services. The X Corp. company represented that as of April 2025, it had approximately 8.8 million monetizable daily active users^{Footnote 10} in Canada. On the other hand, as of February 2026, xAI represented that it had approximately 1.3 million monetizable daily active users in Canada. In addition, X Corp. can collect, use and disclose personal information of Canadians that appears on the X platform or Standalone Grok even if those Canadians are not themselves users of the X platform or Standalone Grok. As a result, we find that the respondents collect, use and disclose personal information related to individuals in Canada.
    3. Both respondents’ Terms of Service and Privacy Policies are applicable to users in Canada and address consent to the collection, use, and disclosure of personal information, and associated rights of access and correction.
    4. The OPC notes that the respondents’ activities take place exclusively through a website or app. As referenced in paragraph 54 of A.T. v. Globe24h.com^{Footnote 11}, a physical presence in Canada is not required to establish a real and substantial connection when considering websites under PIPEDA, as telecommunications occur “both here and there.”
    5. The respondents’ operations require the transmission and receipt of personal information between Canada and the United States, both when collecting information and disclosing it through Grok on both the X platform and Standalone Grok.
34. Given that both X Corp. and xAI are engaged in “commercial activities” within the meaning of PIPEDA, and collect, use, and disclose personal information related to individuals in Canada, the OPC determined that there is a real and substantial connection to Canada. Therefore, the OPC has jurisdiction to investigate and make recommendations related to the respondents’ handling of personal information within its federal jurisdiction of responsibility.

Issue 1: Did the respondents obtain valid consent to produce sexualized deepfakes of individuals?

35. As described below, the OPC finds that X Corp. and xAI did not obtain valid consent from affected individuals, for their collection, use and disclosure of personal information to generate sexualized deepfakes.

Deepfakes and personal information

36. Principle 4.3 of Schedule 1 of PIPEDA requires, among other things, “knowledge and consent […] for the collection, use and disclosure of personal information”.
37. PIPEDA defines personal information as “information about an identifiable individual.” This would include an image of an identifiable individual generated by an AI service such as Grok, regardless of whether what is depicted in that image is real or authentic.
38. While there are many types of synthetic media (i.e., artificially generated images, video, text, or audio content), the Canadian Digital Regulators Forum, of which the OPC is a member, defines deepfakes as synthetic media that is “digitally altered or generated by AI, to depict real or fictional people or things.”^{Footnote 12} In the case of real people, an individual’s traits and characteristics depicted in a deepfake may not actually belong to the depicted individual (or be real) – namely, where these characteristics have been artificially generated by a tool. However, because a deepfake appears to be about a specific individual, even if the information presented by the deepfake is inaccurate, we find that it nevertheless constitutes the individual’s personal information.^{Footnote 13}
39. As a result, the OPC finds that the deepfakes of individuals generated by Grok, including those that contained CSAM and NCII, constitute the personal information of those individuals.

Form of consent

40. In this section, the OPC assesses the form of consent that is required for the collection, use and disclosure of personal information to generate sexualized deepfakes.
41. Principle 4.3.4 of PIPEDA provides that the form of consent sought by an organization may vary, depending on the circumstances and the type of information. In determining the form of consent to use, organizations shall take into account the sensitivity of the information.
42. In addition, Principle 4.3.5 states that when obtaining consent, the reasonable expectations of the individual are also relevant. Section 6.1 of PIPEDA further provides that for consent to be valid pursuant to Principle 4.3, an individual to whom the organization’s activities are directed should reasonably be expected to understand the nature, purpose and consequences of the collection, use or disclosure of the personal information to which they are consenting.
43. Finally, the OPC’s Guidelines for obtaining meaningful consent provide that organizations must generally obtain express consent when: (i) the information being collected, used or disclosed is sensitive; (ii) the collection, use or disclosure is outside of the reasonable expectations of the individual; and/or (iii) the collection, use or disclosure creates a meaningful residual risk of significant harm.

Sensitivity

44. The OPC’s Interpretation Bulletin on Sensitive Information specifies what factors are relevant when determining whether personal information is sensitive. For example, personal information that pertains to an individual’s sex life or sexual orientation is generally considered sensitive and requires a higher degree of protection.
45. In addition, the joint resolution on the Best Interests of Young People, of which the OPC is a signatory, notes that young people’s personal information is particularly sensitive.
46. Sexualized deepfakes of individuals present them in sexually explicit situations. They are deeply invasive, showing individuals in vulnerable states and disclosing physical and behavioural characteristics that are deeply personal and sensitive.
47. As a result, the OPC considers sexualized deepfakes to contain sensitive personal information of individuals depicted.

Reasonable expectations

48. The OPC finds that individuals would not reasonably expect innocuous images of themselves, posted on a social media platform, blog post, or other similar websites, to be processed using an AI tool to generate a sexualized deepfake of themselves. This is especially true where those deepfakes pertain to minors.
49. Indeed, individuals are most likely to have posted these images on the Internet for a completely unrelated purpose (for example, sharing a vacation photo with friends and family, or for professional reasons). Furthermore, individuals may not even have posted the images themselves.
50. As a result, the OPC finds that these individuals would not expect a third-party individual with whom they likely did not have any relationship to use the respondents’ AI tools to transform their images into a sexualized deepfake and disseminate it.

Risk of harm

51. The OPC finds that the generation and publication of such sexualized deepfakes, especially where it is without the knowledge and consent of the depicted individuals, can create a real risk of significant harm to these individuals, including reputational, psychological or financial (e.g., loss of employment opportunity, sextortion) harm.^{Footnote 14}
52. For all these reasons, the OPC finds that express consent must be obtained for the collection, use, and disclosure of personal information to generate sexualized deepfakes.
53. The respondents represented that they do not have a process in place to obtain individuals’ consent for the collection, use and disclosure of personal information for the generation of deepfakes via @Grok, Grok on X or Standalone Grok. They stated that they are not responsible for obtaining consent from individuals that appear in sexualized deepfakes because “users bear primary responsibility for the content that they generate on their respective platforms.”
54. More specifically, X Corp. stated that “where an X user instructs the @Grok account to generate images that are subsequently posted on the X platform, any collection, use, or disclosure of personal information for the purpose of generating and sharing that content is initiated and determined by the user.”
55. Similarly, xAI submitted that “the generation of synthetic images by Grok would be properly characterized as a collection, use and/or disclosure of information by the individual user who had entered the prompt for the creation of the image.”
56. Both respondents further argued that the use of these tools is for “personal” and “artistic” purposes by the user, and that they would therefore not be required under PIPEDA to obtain consent for the creation of deepfakes.
57. The OPC disagrees for the following reasons:
    1. As outlined above, PIPEDA applies to organizations that collect, use and disclose personal information in the course of commercial activities, including xAI and X Corp.
    2. xAI’s privacy policy states that, with a view to providing, analyzing, and maintaining its service, it may collect personal information directly from users, whether it is the user’s personal information or not. User content includes “prompts and other content [users] input, such as files, images, audio, voice, video and other materials” and Grok’s outputs (such as responses that Grok generates).
    3. X Corp.’s privacy policy states that with a view “to provid[ing] […] products and services, […] helping keep X more secure and respectful for everyone, and more relevant […]”, it collects personal information including, “posts […] interactions with other users’ content, such as reposts, likes, bookmarks, shares, downloads, replies”.^{Footnote 15}
    4. While users provide the prompts and images on the basis of which Grok generate sexualized deepfakes, it is ultimately X Corp. and xAI that collect this information and enable their use for the purpose of generating this content; they do so via a tool that they developed and deployed, and that they host and make accessible to users of their platforms in the course of a commercial activity.
58. As a result, the OPC finds that xAI and X Corp. are accountable for the content generated on their platforms and are therefore required to ensure that valid consent has been obtained for the collection, use, and disclosure of personal information to generate sexualized deepfakes.
59. As mentioned above, X Corp. and xAI represented that they do not verify consent of the depicted individuals. Consequently, the OPC finds that xAI and X Corp. contravened Principles 4.3 and 4.3.4 of PIPEDA.

Issue 2: Would a reasonable person consider it appropriate for the respondents to produce sexualized deepfakes of individuals?

60. Subsection 5(3) of PIPEDA states that “[a]n organization may collect, use or disclose personal information only for purposes that a reasonable person would consider are appropriate in the circumstances.”
61. The OPC’s Guidance on inappropriate data practices: Interpretation and application of subsection 5(3) elaborates that, in Turner v. Telus Communications Inc, the Federal Court set out factors for evaluating whether an organization’s purpose is in compliance with subsection 5(3). Accordingly, the OPC considered the following pertinent factors in analyzing the practices of both respondents: i) the degree of sensitivity associated with the personal information; ii) whether the organizations have a legitimate need or bona fide business interest in the practice; (iii) whether there is a less privacy invasive means of achieving the same ends at comparable cost and with comparable benefits; and, iv) the proportionality of the loss of privacy against the benefits of the practice.
62. As explained below, the OPC finds that the respondents have a bona fide business interest in providing their users with an image generation and editing tool. However, as pertains to the generation of sexualized deepfakes, given that these involve sensitive personal information and result in both a loss of privacy and a risk of significant harm to individuals, the OPC finds that the generation of sexualized deepfakes causes privacy and other harms that significantly outweigh the benefits to the organizations and that there are less privacy invasive means to achieve the respondents’ purposes at comparable cost and with comparable benefits.

Degree of sensitivity associated with the personal Information

63. As discussed at paragraph 44 and following, the OPC generally considers sexualized deepfakes to contain sensitive personal information.

Whether the organizations have a legitimate need or bona fide business interest

64. The xAI company represented that its stated purpose is to develop and provide AI services to end users, enterprise customers and other developers. In turn, X Corp. represented that its stated purpose is to provide a platform to connect users, advertisers, creators and businesses. Both organizations represented that they neither had, nor have, any intention of generating sexualized deepfakes of identifiable individuals.
65. The OPC accepts that the deployment of an image and video generation technology can serve the general purpose of both respondents and, therefore, represents a legitimate need/bona fide business interest for them. However, the OPC finds that this does not extend to an image generation tool that allows for non-consensual sexualized deepfakes of identifiable individuals.

Whether there is a less privacy invasive means of achieving the same ends at comparable cost and with comparable benefits

66. The OPC finds that with forethought and anticipation, the respondents could have developed and deployed a less privacy invasive AI image generation and editing tool, at a comparable cost and with comparable benefits (for users and the respondents).
67. In particular, by embedding robust safeguards at the outset (before launching the technology) rather than attempting to address the misuse of their tool retroactively, the respondents could likely have prevented the tool from generating harmful, sexualized deepfakes, while still providing their users with an image generation and editing feature.
68. As explained at paragraph 82, while the respondents conducted a Privacy Impact Assessment (“PIA”) with respect to certain features, they failed to identify some key risks, including the fact that multiple users may prompt Grok to generate sexualized deepfakes.
69. By better identifying and mitigating risks early in the development lifecycle, the respondents would have limited the magnitude of the harms caused to depicted individuals and maintained trust by demonstrating accountability.

Proportionality of the loss of privacy against the benefits

70. In weighing the loss of privacy against the benefits of deploying an AI generation and editing tool capable of producing sexualized deepfakes of identifiable individuals, the OPC took into account: i) the significant risk of harm associated with the collection, use, and disclosure of such personal information, ii) the fact that the respondents do not have a process to obtain consent for this practice, and more generally the challenges of obtaining consent in the circumstances; and, iii) the absence of robust safeguards and an effective response to the non-consensual generation and disclosure of sexualized deepfakes.
71. As explained below, the OPC finds that the loss of privacy and the risk of significant harm to individuals significantly outweigh the benefits of this service.

Risk of harm

72. As mentioned at paragraph 51, the OPC finds that the risks of harm associated with the collection, use and disclosure of personal information to generate sexualized deepfakes are very significant, especially where consent is not obtained from the depicted individuals. Risks include, but are not limited to, reputational harm, sextortion, cyberbullying, grooming^{Footnote 16}, and loss or denial of employment.

Absence of a consent process/Impracticality of obtaining consent

73. The respondents stated that they did not intend to collect, use and disclose personal information for the purpose of generating non-consensual sexualized deepfakes and that they have safeguards in place to prevent the generation of sexualized deepfakes altogether (as detailed below).
74. That said, the OPC notes that the respondents’ terms of use do not seem to formally prohibit individuals from using Grok to generate consensual sexualized deepfakes, and that some individuals may in theory consent to the collection, use, and disclosure of sexualized deepfakes that depict them.
75. The OPC has previously found that in the context of pornographic images or videos, websites and platforms should have rigorous measures in place to ensure that user-generated content is only posted with the knowledge and consent of all depicted individuals.^{Footnote 17}
76. As explained at paragraph 53, the respondents both acknowledged that they do not have a process in place to ensure that depicted individuals have consented to the collection, use and disclosure of their personal information to generate sexualized deepfakes.
77. The OPC recognizes that such a consent verification process may be difficult or even impractical to implement in the context of a general-purpose social media platform such as X or an online, freely accessible chatbot (Standalone Grok).
78. However, operational challenges in monitoring compliance do not justify reducing obligations under PIPEDA, particularly when the organization creates those challenges itself and where there are significant harms to individuals.^{Footnote 18} In this case, the OPC finds that the respondents made informed decisions in how they structured their business, developed and deployed AI technology on their respective platforms and in how these decisions impacted individuals’ privacy. Therefore, they cannot claim that it would be impossible to limit the harms associated with a tool that they created and released.
79. Given the above, the OPC finds that a reasonable person would consider that allowing the generation and sharing of sexualized deepfakes without having first verified consent of depicted individuals is inappropriate, and that the resulting harms and loss of privacy clearly outweigh potential benefits to the respondents.

Absence of effective safeguards and response

80. In this context where consent is not, and likely cannot, be verified, the OPC finds that the respondents should have mitigated privacy harms by implementing robust technical and organizational safeguards to effectively prevent Grok from generating sexualized deepfakes, particularly in the context of CSAM and NCII.
81. The respondents stated that prior to the concerns raised regarding the generation of non-consensual sexualized deepfakes, they had, and continue to have, a number of organizational and technical safeguards in place to prevent this content from being generated and posted.
    1. Organizational measures:
       1. The respondents both represented that they conducted PIAs, as discussed later in more detail.
       2. X Corp. further submitted that other organizational measures it has in place include: a suite of policies that support the mitigation of risks associated with posted content on the X platform (including AI-generated content) and deter violative content and conduct; reporting and grievance redressal mechanisms; and an enforcement framework for the review and removal of violative content.
       3. xAI also submitted that its Terms of Service require users to comply with its Acceptable Use Policy, which explicitly prohibits the sexualization and exploitation of children and “[d]epicting likeness of persons in a pornographic manner”; and, like X Corp., whenever violative content is detected, it is taken down and enforcement measures are taken against the user that created the materials, including reporting the user to law enforcement where appropriate.
    2. Technical measures:
       1. xAI represented that it has implemented input and output filtering (e.g., using classifiers^{Footnote 19} that check for specific key words such as “child”), content moderation, and use of the US National Center for Missing and Exploited Children (“NCMEC”) database and other CSAM hash lists^{Footnote 20}, amongst others.
82. However, with respect to the PIAs, the OPC notes that:
    1. X Corp. only conducted a PIA with respect to Aurora (the AI image and editing generation model with which Grok was updated in December 2024). The PIA was only updated to include Grok Imagine in March 2026, after widespread reports of non-consensual intimate imagery and CSAM being disseminated by Grok. It was only in the updated PIA that X Corp. included an “Incident Response Protocol”, which notes its incident response protocol that was triggered in late December 2025. The PIA does not specify what the incident response protocol involves; rather, it states that in response to the revelations in late December 2025, “X took swift action to remove violative content in relation to the incident, and additional technical measures were introduced to improve the guardrails in place.” In addition, in the PIA, X Corp. claims that “this incident [in late December 2025] evidences the effectiveness of this incident response protocol in addressing any gaps in the existing guardrails.” As discussed further at paragraph 88, the OPC disagrees with the respondents’ assertion that their incident response protocol was effective in addressing the issue.
    2. xAI conducted a PIA about Grok generally, which did not include an assessment of @Grok specifically.
       1. xAI explained this discrepancy by noting that “no necessity arose for a separate assessment of the provision of the existing image generation capabilities via @Grok” because the @Grok account does not provide users with “any new data processing as such”, but merely with “an alternative means of access to both existing services simultaneously,” namely, “the generation of content in response to the user’s prompt (tagging @Grok) and the automatic posting of that content on the X platform.”
       2. While the OPC recognizes that @Grok does not use any new data processing, the OPC is of the view that a separate PIA should have been prepared for @Grok given that the privacy impact of generating synthetic media using @Grok is distinct from doing so by using Standalone Grok. As noted above, in the case of @Grok, the content generated by Grok is shared with anyone and everyone who has access to the thread in which the content is posted. In addition, the OPC notes that the concerns raised following the generation of non-consensual sexualized deepfakes occurred primarily on the X platform using @Grok, demonstrating the heightened risk and harm that occurred when the “Edit Image” button was introduced to images on the X platform.
    3. In its PIA, xAI considered, but did not accurately reflect, the risks to security, safety and privacy that could result if Grok were used with malicious intent. Specifically, the PIA notes that an “LLM could […] be used to generate photorealistic images or other types of content for the purpose of sexual exploitation/ blackmail/ financial exploitation, misleading recipients and spreading misinformation.” However, xAI noted that to mitigate the risks, besides safeguards, it limits the number of queries that an “Unauthenticated Grok User, as opposed to an Authenticated Grok User who logs into Grok with some form of credentials […] can submit within a period of time without logging in.” As stated in the PIA, this is intended to “reduce[s] the likelihood that bad actors could successfully use Grok for malicious purposes and increase[s] the probability that, to the extent bad actors do attempt to use Grok for such purposes, their activity is associated with a verifiable Grok User account that xAI can disallow their future use of Grok.” The OPC notes that xAI did not consider the risk that a large number of users would generate sexualized deepfakes where there were insufficient safeguards to prevent it, and even considered the periodic generation of sexualized deepfakes to constitute a “low” risk, given their existing safeguards.
    4. X Corp. claimed that it conducted a security assessment at the time of integrating the Grok model into Grok on X, as indicated in the PIA.
83. The OPC notes that the widespread generation of non-consensual sexualized deepfakes (with millions of images reported to have been generated by Grok) demonstrates that the respondents’ safeguards were insufficient. Specifically, the OPC finds that the respondents developed and deployed the image generation and editing feature without having properly assessed the risks associated with the tool and implemented robust safeguards, which resulted in a failure to effectively mitigate these risks.
84. Therefore, noting that the generation of sexualized deepfakes involves sensitive personal information, the OPC finds that, while the respondents have a bona fide business interest in providing their users with an image generation and editing tool, there were less privacy invasive ways to do so. In addition, the OPC finds that the loss of privacy, as well as the harm resulting from this practice, significantly outweigh the benefits to the organizations. This assessment is based on the significant risk of harm resulting from the generation of sexualized deepfakes of identifiable individuals, the absence of a consent process (or its impracticality) and the respondents’ failure to implement effective safeguards. Consequently, the OPC finds that xAI and X Corp. contravened subsection 5(3) of PIPEDA.

Initial response to the concerns following the generation of sexualized deepfakes

85. X Corp. represented that on December 28, 2025, shortly after it became aware of the publicly raised concerns with respect to the generation of non-consensual sexualized deepfakes, it initiated its crisis protocol and began working with xAI to address the issue.
86. More specifically:
    1. X Corp. and xAI stated that they immediately started conducting “sweeps” to detect potentially violative content, which “resulted in the review of additional categories of Grok-generated content and prompts to Grok”. They further indicated that they conducted additional sweeps in early January, and took action against violative content and users generating such content.
    2. xAI reinforced its technical measures and introduced additional measures. In particular:
       1. On January 8, 2026, xAI prevented the @Grok account from responding to prompts or circumvention attempts that would generate sexualized deepfakes of identifiable individuals. These same preventative measures were applied to Grok on X and Standalone Grok on January 14, 2026, across all users, including Premium users of the platform X.
       2. On January 8, 2026, xAI, working with X Corp., prevented users without an active X Premium subscription from accessing image generation via the @Grok account.^{Footnote 21}
       3. xAI refined the classifiers (see definition in footnote 19) used for content moderation.
    3. X Corp. updated the guidance and training for its Safety team to ensure that staff is better equipped to detect malicious actors on the platform. These updates also included a revision to its internal policies and stricter guidelines (issued on January 16 and 22, 2026) to ensure that the company is generally better positioned to address unwanted sexual content. X Corp. did not elaborate on the details of these revisions.
    4. X Corp. also explained that it suspended, and where relevant, reported to law enforcement authorities, the accounts of users who engaged in repeated or egregious policy-violating behaviour, such as prompting the @Grok account to generate and share certain types of content.
87. In submissions provided to the OPC in March 2026, X Corp. indicated that the above-noted actions had resulted in a decrease by approximately half of “unwanted sexual content violations by Grok”.
88. For the following reasons, the OPC found X Corp. and xAI’s initial response to be insufficient:
    1. While X Corp. stated that the changes resulted in a decrease by approximately half of “unwanted sexual content violations by Grok”, it was still possible to generate and share non-consensual sexualized deepfakes on the X platform. In addition, it is unclear what total was cut in half, and therefore how many sexualized deepfakes continued to be generated using Grok. As mentioned at paragraph 26, certain news reports and studies estimated that Grok generated several millions of sexualized deepfakes. Moreover, xAI did not provide any metrics to concretely demonstrate the effectiveness of its actions with respect to Standalone Grok.
    2. While the respondents indicated that they had initiated their crisis protocol on December 28, 2025, it took them over ten days to introduce the technical measures described in paragraph 86(b) to attempt to prevent users from violating their policies. The OPC considers this to be an insufficient and untimely response to the potentially millions of non-consensual sexualized deepfakes that were generated by Grok.

Recommendations

89. Given all the above, the OPC found that X Corp. and xAI contravened section 5(3) and principle 4.3 of Schedule 1 of PIPEDA.
90. Therefore, with a view to bringing X Corp. and xAI into compliance with PIPEDA, the OPC recommended that the respondents:
    1. Immediately suspend the functionality of Grok Imagine on all of X Corp.’s and xAI’s products until they can demonstrate to the OPC that their safeguards are sufficient to prevent the generation of sexualized deepfakes.
    2. Within six months of receiving this report, develop and implement a formal process for anticipating and mitigating privacy issues (beyond non-consensual sexualized deepfakes) associated with image generation tools and other novel products.
    3. Provide the OPC with an annual third-party audit – by a qualified independent third-party – of Standalone Grok, @Grok, and Grok on X’s safeguards, including statistics on their effectiveness, until the issue of sexualized deepfakes is fully resolved.
    4. Within six months of receiving this report, provide the OPC with documentation demonstrating that X Corp. and xAI are proactively and regularly monitoring, in a privacy protective manner, for sexualized deepfakes that may be generated in circumvention of their safeguards, and not only in response to an incident.
    5. Inform the OPC within three months of receiving this report of the progress made with respect to the following initiatives, which the OPC understands are being explored by xAI:
       1. Enhancing its user interface, such as through prominent warnings about potential limitations and risks of AI outputs, to promote informed use, consistent with its transparency commitments.
       2. Improving its reporting and assessments mechanisms (including by expanding audits, privacy impact assessments, and pre-release evaluations).

Response to the recommendations

91. In response to our preliminary findings and recommendations, the respondents expressed their disagreement with our conclusions.
92. More specifically, they represented that NCII and CSAM have no place on the X platform and Standalone Grok and asserted that they are taking all reasonable steps necessary to prevent their occurrence and taking immediate action when such content comes to their attention.
93. They also described additional measures that they have recently implemented to further prevent Grok from generating violative content.
94. Specifically, xAI submitted that it has:
    1. introduced new and improved technical safeguards, including to better identify and reject user requests to remove or alter clothing, update its real-time monitoring and better detect images of minors;
    2. banned the generation of images and videos by non-paid subscribers of Standalone Grok in the Imagine tab of the platform and limited the amount of content that can be generated by non-paid subscribers in the Chat tab of the Standalone Grok platform;^{Footnote 22}
    3. started carrying out further proactive checks of social media sites to detect and remove content that may violate its moderation policies (which is in line with the recommendation in paragraph 90(d)); and
    4. started developing and testing a more sophisticated solution that would take into consideration the context of user prompts to better detect violative content.
95. X Corp. further indicated that it has removed infringing and illegal content from the X platform, and that it has carried out further proactive checks of the X platform to detect potentially infringing content.
96. In these circumstances, the respondents indicated that suspending the operation of Grok Imagine (as recommended in paragraph 90(a)) is not warranted.
97. The OPC recognizes that the respondents have now deployed, or are deploying, safeguards designed to reduce the risk that their tool will be misused to produce sexualized deepfakes.
98. While this is encouraging, the OPC notes that the respondents have not, to date, demonstrated the effectiveness of these safeguards in preventing and mitigating this issue. Instead, the respondents have represented that the implementation of these safeguards “will improve xAI’s ability to detect and address any issues that malicious actors are able to exploit to circumvent Grok’s safeguards and violate xAI policies” and “aims to improve its prevention capability continually over time.” [our emphasis added]
99. In these circumstances, at the time of writing this report, the OPC cannot conclude that the complaint is resolved.
100. That said, the respondents confirmed that they would continue to work with the OPC to resolve the remaining concerns and committed to implementing the recommendations in paragraph 90(b), (c), (d) and (e).
101. Additionally, with respect to the recommendation in paragraph 90(b), the respondents indicated that they have in place formal processes for anticipating and mitigating privacy issues (beyond non-consensual sexualized deepfakes) associated with image generation tools and other novel products, but that as those processes are being further refined, they would provide an update on their development and implementation in six months.
102. Finally, the respondents committed to providing the OPC with quarterly reports to explain and demonstrate, with detailed submissions and corroborating evidence (including statistics and metrics): (i) the specific measures that they have implemented to improve the efficacy of their safeguards; (ii) the effectiveness of the mechanisms that they have in place to review and remove violative content and address reports and grievances; and (iii) any progress made towards compliance with the other recommendations.

Conclusion

103. Given all the above, the OPC considers the matter to be well-founded.
104. While the OPC is encouraged by the respondents’ commitments, we will continue to monitor the implementation of these commitments to ensure that the serious issues highlighted in this report are fully addressed.