Europe & U.S.: EU Refines AI Act Enforcement, UK and Ukraine Form Defence AI Partnership, FTC Finalizes AI Marketing Case, California Advances AI Transparency Amendment

On August 24, 2026, the European Commission updated its The Enforcement Framework of the AI Act page, consolidating information on rules enforceable since August 2, 2026, the division of responsibilities among regulators, and currently available investigation and complaint tools. The framework covers obligations for general-purpose AI models (GPAI models) and the Article 50 transparency requirements for generative AI content. It is important to distinguish the nature of the update: the August 24 publication was an informational enforcement update by the Commission. It did not amend the AI Act itself or create a new set of substantive obligations on that date; the relevant GPAI obligations and Article 50 transparency rules had already entered their applicable phases.

The update first brings the question of “who supervises what” into a more consolidated view. The European AI Office (AI Office) directly supervises providers of GPAI models. It is also responsible in certain cases where the same provider or corporate group supplies both an underlying GPAI model and an AI system built on top of it, as well as relevant AI systems integrated into very large online platforms or very large online search engines designated under the Digital Services Act. Other AI systems are generally enforced by competent national market-surveillance authorities, while AI systems used by EU institutions are overseen by the European Data Protection Supervisor (EDPS). For GPAI, the AI Office can request information, conduct model evaluations, and require access to models; where necessary, it may require providers to implement risk-mitigation measures or even restrict a model’s availability. In cases of intentional or negligent non-compliance, the Commission may also adopt penalty decisions under the AI Act. The page also makes available or consolidates the AI Act complaints tool, whistleblower tool, and a dedicated complaints channel for downstream providers integrating third-party GPAI models.

Transparency for generated content now forms a separate responsibility chain running alongside GPAI model governance. Article 50 has applied since August 2, 2026: AI systems that directly interact with natural persons generally need to inform users that they are interacting with AI; providers of systems that generate synthetic audio, image, video, or text content must ensure outputs carry machine-readable markings that identify them as AI-generated or manipulated; and deployers using deepfake content, or publishing AI-generated text intended to inform the public on matters of public interest without qualifying human review or editorial control, face corresponding disclosure duties. A critical distinction is between model-level transparency and system/content-level transparency. GPAI rules mainly require model providers to supply technical documentation and information to regulators and downstream providers, publish a summary of training content, and address copyright compliance. Article 50, by contrast, focuses on how final AI systems disclose their AI identity to users and how generated content is marked and disclosed. For generative AI systems placed on the market before August 2, 2026, the machine-readable marking obligation under Article 50(2) is subject to a transition period, with relevant providers expected to complete adaptation by December 2, 2026.

At the enterprise-deployment level, the practical adjustment is to map these two responsibility sets separately onto model procurement and product-output workflows. Enterprises using third-party GPAI APIs should first determine whether, under the AI Act, they remain downstream AI-system providers or whether extensive fine-tuning, re-provisioning of a model, or similar conduct could make them GPAI model providers themselves. Even where an enterprise remains downstream, it should obtain enough information from its model supplier to explain model capabilities, limitations, and compliance conditions. For gateways, agents, or SaaS products that connect to multiple models, model substitution should not be evaluated only on price and performance. Enterprises also need to record which model version supplied which transparency information, whether it supports machine-readable marking of generated content, and who bears responsibility at the application layer for end-user notice and content disclosure. Enterprises offering chatbots, digital humans, content generation, or automated publishing to EU users should place AI-identity notices, machine-readable provenance markings, and visible disclosure labels into separate product-testing and release workflows rather than treating them as one generic “AI label.” The Commission’s transparency guidance also makes clear that visible disclosure for deepfakes cannot be replaced solely by an underlying machine-readable mark.

The August 24 update is best understood as a milestone in the AI Act’s shift from “obligations written into law” toward “who enforces them, how investigations work, and where enterprises can be complained about.” Obligations for GPAI model providers have applied since August 2, 2025, while the Commission’s related enforcement powers became operational from August 2, 2026; Article 50 transparency obligations also began to apply on that date. Meanwhile, the EU’s 2026 Digital Omnibus on AI adjusted the timetable for high-risk AI: requirements for standalone high-risk systems listed in Annex III were postponed to December 2, 2027, while high-risk AI systems embedded in regulated products were postponed to August 2, 2028. The EU is therefore not enforcing all AI Act provisions at once. Instead, a layered implementation pattern is emerging: GPAI rules, prohibited practices, and parts of the transparency regime are already in real enforcement, while the full high-risk AI compliance regime remains subject to later application dates. For enterprises operating across regions, this is why EU deployment readiness cannot be managed through a single “Is the AI Act in force?” checklist; compliance timing needs to be mapped separately by model type, system use, enterprise role, and specific provision.

On August 24, 2026, the governments of the United Kingdom and Ukraine signed the Joint Declaration of Intent between the United Kingdom of Great Britain and Northern Ireland and Ukraine on a UK-Ukraine Artificial Intelligence Partnership, a political declaration establishing a government, industry, and research cooperation framework around defence AI, autonomous systems, and network-enabled capabilities. The declaration includes co-developed models, secure data and compute pathways, and AI assurance within the scope of cooperation. It explicitly states that it does not create legally binding obligations.

The declaration divides cooperation into three interlocking pillars. Government-to-government collaboration covers co-developed models, secure data and compute pathways, and joint assurance; industry cooperation, coordinated through government-to-government arrangements, focuses on developing, testing, and deploying AI-enabled capabilities against agreed operational problem sets; academia and research focus on autonomy, AI assurance, cyber security, and synthetic data. The principles of cooperation further require respect for each party’s sovereignty over assets and data, trusted safeguards for data, intellectual property, and export controls in line with national law, and consideration of NATO interoperability. The framework adopts a “pilot-first” approach, prioritizing practical cooperation already under way, with further implementation arrangements to be developed separately. Here, AI assurance should not be understood merely as ex post audit. The UK Ministry of Defence’s existing JSP 936 framework places safety, reliability, effectiveness, cyber security, and appropriate human oversight within full-lifecycle governance of defence AI, meaning joint assurance is closer to an ongoing trust and verification mechanism accompanying development, testing, deployment, and operation.

The declaration itself does not create new compliance obligations for ordinary commercial enterprises. Its practical impact falls mainly on technology companies, research institutions, and suppliers participating in UK-Ukraine defence and national-security AI projects. The UK government has also confirmed that the UK will become the first international partner to gain access to Ukraine’s Avengers AI Labs; the platform uses real battlefield data to train AI models, and Ukraine’s Ministry of Defence has previously said that its core dataset contains around five million annotated frames. For enterprises entering such projects, a more specific set of deployment questions therefore arises alongside model performance: which party controls training data and model assets, who may access the data, through what environments data and compute are processed, how intellectual property in jointly developed outputs will be allocated, and whether cross-border sharing of models, chips, technical information, or training capabilities is constrained by export controls. The declaration does not yet publish technical standards or approval procedures for these matters, so it would be inappropriate to infer that it already mandates data localization, fixed compute regions, or a specific audit format. At this stage, the more accurate conclusion is that data location, compute pathways, access permissions, intellectual property, and export controls are being designed into project architecture from the outset rather than addressed separately after model development is complete.

This cooperation did not begin from zero. The One Hundred Year Partnership Agreement, signed in January 2025, had already placed defence-industrial cooperation, technology transfer, and intellectual property protection within a long-term bilateral framework, while its accompanying declaration explicitly called for cooperation in emerging technologies such as AI and drones and for maintaining safe and responsible AI development, governance, and regulation. In 2026, that trajectory began moving into more concrete infrastructure and deployment stages: in March, Ukraine opened real battlefield data to partners for training unmanned-system models and developed the UK-supported A1 defence AI centre; in June, the UK established the Rapid AI Delivery Taskforce to accelerate defence AI from strategy and pilots into real deployment while retaining existing assurance and oversight processes. Unlike the EU’s risk classification, transparency, and compliance duties for general market actors, the UK-Ukraine declaration addresses a different governance problem: in cross-border, high-sensitivity AI projects, the governance object is no longer only model behaviour, but also who owns the data, who controls compute, whether technology can move across borders, and whether collaborative outputs can be shared with third countries. It is closer to a “sovereign AI + defence supply-chain governance” deployment framework than to a generally applicable AI regulation.

On August 27, 2026, the U.S. Federal Trade Commission (FTC) finalized three administrative consent orders involving CMG Media Corporation, doing business as Cox Media Group, MindSift LLC, and 1010 Digital Works LLC, resolving allegations that the companies made false or misleading representations about an AI-powered marketing service known as “Active Listening.” The three companies were required to pay a combined $930,000. Following public comment, the Commission voted 2-0 to approve the final consent agreements.

A key fact in this matter needs to be separated from the headline: the FTC did not find that the three companies actually used smart devices to listen to consumers’ conversations. The complaints alleged that the companies marketed the AI-powered Active Listening service to small and medium-sized business customers as being able to listen to conversations near consumers’ smart devices, use algorithms to identify advertising-relevant content, and target ads to consumers in specific locations. The FTC’s investigation found that the service did not in fact collect or use voice data, did not use AI to detect relevant conversations, and did not perform geographic targeting as represented; the actual service largely resold email lists obtained from data brokers. The companies also represented that consumers had consented to this use of voice data, while the FTC alleged that such consent had not been obtained. The final orders therefore prohibit the companies from making misrepresentations about advertising-service functionality, collection and use of voice data, whether consumers have consented, and geographic-targeting capabilities. CMG is required to pay $880,000, while MindSift and 1010 Digital Works each pay $25,000.

The most direct enterprise AI deployment implication is not model training, but that product-capability evidence and the data-consent chain must be capable of substantiating marketing claims. If sales materials claim that a product can use AI to analyze voice, emotion, location, or user behavior, product, data, and compliance teams should first confirm whether the system actually has that capability, what data it uses, and whether statements that “users have consented” map to a real consent flow. The FTC emphasized in this matter that consumers do not thereby “opt in” to household voice-monitoring-style data use merely because they accepted general terms of service when downloading an app. The FTC further indicated that if Active Listening had operated as marketed, collecting and using consumer voice data without adequate consent could itself potentially violate Section 5 of the Federal Trade Commission Act. The evidence enterprises need to retain therefore goes beyond advertising-approval records. It should be capable of answering three questions: which models and data sources the product actually uses, what testing or technical documentation supports the capability claim, and where, for what specific purpose, and through what choice mechanism the user consented when personal data is involved. For brands procuring external AI marketing tools, vendor sales materials are not a substitute for technical and data due diligence.

The cases also extend responsibility upstream to parties that supplied marketing materials and sales narratives. The FTC separately alleged that MindSift and 1010 Digital Works provided CMG with the “means and instrumentalities” of deception through marketing materials, sales representations, and responses to customer questions. For the AI product ecosystem, this is more practically significant than simply penalizing the final seller. If model or data suppliers provide resellers or channel partners with unsubstantiated claims such as “AI can identify this type of data,” “consumers have authorized the use,” or “the system can precisely target users,” the risk may not remain solely with the downstream seller. Supplier-management processes should therefore require verifiable materials on technical capabilities, data sources, and authorization grounds rather than allowing marketing teams to reuse upstream vendor claims without substantiation.

In timeline terms, the FTC first announced the three proposed administrative settlements on May 21, 2026, setting out the $930,000 payment arrangement and restrictions on misrepresentations before opening the matters for public comment. The August 27 action finalized those previously proposed orders rather than initiating the cases for the first time. The process also illustrates the enforcement boundary. The Center for AI and Digital Policy had urged the FTC to expand the orders to include prohibitions on monitoring and surveillance services, explicit affirmative consent, data deletion and transfer restrictions, algorithmic disgorgement, and governance programs. The FTC responded that those suggestions concerned digital eavesdropping, interception of private communications, and biometric-information use that were not alleged in the complaints. Rather than describing the cases as “the U.S. banning AI listening,” the more accurate framing is that the FTC is using existing consumer-protection law to evaluate whether AI capability claims, actual data-processing practices, and representations about user consent are consistent with one another. In the U.S. market, which still lacks a single comprehensive federal AI law, enterprises can therefore enter existing consumer-protection enforcement directly when AI product marketing does not match actual system behavior.

On August 28, 2026, the California Legislature completed its final legislative vote on AB 2713, California AI Transparency Act: system provenance data. The Senate had passed the bill 40-0 on August 27, and the Assembly concurred in the Senate amendments the following day by a 78-0 vote, after which the bill moved into enrollment and presentment. As of August 28, the bill had passed the Legislature but could not yet be described as effective law; it still required the Governor’s signature to become law.

AB 2713 does not create a new generative AI labelling regime from scratch. Instead, it amends the existing California AI Transparency Act’s rules for how large online platforms handle content provenance data. Provenance data here means information embedded in, attached to, or otherwise associated with content that indicates the authenticity, origin, or modification history of digital content. System provenance data excludes personal information that can reasonably be linked to a specific user and instead principally reflects what device, system, or service generated the content and information about its authenticity. Under the latest Senate amendments dated August 21, large online platforms would be required to detect provenance data embedded in, attached to, or otherwise associated with content and notify users through the interface where that data or a digital signature indicates that content was generated or materially altered by a generative AI system or produced by a capture device. Users would also need to be able to inspect the relevant information through the platform interface, an external link, or a separate download of the provenance data. The bill would further extend the prohibition on removing provenance data from the stages of “upload or distribution” to the point at which users download content from the platform, while clarifying that the obligations do not require platforms to preserve, display, or allow the download of personal information, or to process provenance data incompatible with widely adopted standards. The relevant large-platform provisions are still scheduled to apply beginning January 1, 2027.

For platform companies, the main impact falls on three technical points—content ingestion, display, and export—rather than creating a simple additional “AI content label.” Platforms first need to detect provenance metadata inside files or otherwise associated with them during upload or distribution, then determine which elements qualify as system provenance data that can be shown to users and which may contain personal information. At the display layer, interfaces also need to let users see whether provenance information or digital signatures exist and what generation or capture system they relate to. The download chain is easier to overlook: if a platform automatically strips existing provenance information when compressing images, transcoding video, generating thumbnails, or repackaging files, AB 2713’s proposed rules may turn those backend processing steps into independent compliance checkpoints. The real question for enterprises is therefore not merely “do we have an AI label?” but whether provenance information can be detected, preserved, and appropriately presented to users throughout the upload-processing-distribution-download lifecycle. The bill’s personal-information boundary also requires platforms to distinguish traceability from privacy: being able to prove what system generated content does not mean the platform can disclose information that identifies the specific creator or device user. This is a deployment mapping derived from the amendments; the bill itself does not prescribe a single metadata format, digital-signature scheme, or platform UI design.

AB 2713 sits within a policy trajectory California has been building over the past two years around content provenance. SB 942, the California AI Transparency Act, enacted in 2024, first placed obligations on large generative AI providers, requiring machine-detectable latent disclosures for generated images, audio, and video and requiring detection tools. AB 853 in 2025 then expanded the governance chain to large online platforms and content-capture devices, with platform obligations scheduled to begin in 2027. AB 2713 addresses implementation details that emerged as this chain moved into real platform operations. One explicit background issue was copyright-holder concern that allowing users to download complete content together with provenance information could facilitate piracy. Legislative committees therefore adjusted the approach so users could download the provenance data itself without necessarily downloading the full work for verification, while also adding a prohibition on platforms actively stripping system provenance data or digital signatures when users download content. Compared with Article 50 of the EU AI Act, which focuses mainly on machine-readable marking and visible disclosure obligations for AI-system providers and deployers, California’s approach places greater emphasis on whether provenance information continues to travel with content once it enters social-media, file-sharing, or search platforms and remains inspectable by users. The common enterprise deployment question is therefore not a single “watermark,” but whether origin-side embedding, platform-side detection, distribution-side preservation, and user-side verification can form a continuous technical chain.