Article 50 Went Live on 2 August and the EDPB Just Rewrote What "Anonymous" Means

Week of 3 August 2026 | Compiled by Team Paringa

Two things landed in the same fortnight: an enforceable AI transparency regime and a new European standard for when data stops being personal data. Together they change what DPOs and CISOs have to be able to prove.

SECTION 1 - What Happened This Week

🔴 EU AI Act Article 50 became enforceable on 2 August, and it was never deferred

Most enterprises spent the first half of 2026 absorbing the message that EU AI Act enforcement had been pushed back. That message was half right and expensively misleading. The Digital Omnibus package, which reached provisional political agreement on 7 May 2026 and was approved by the Council on 29 June, moved the compliance deadline for standalone high-risk Annex III systems from 2 August 2026 to 2 December 2027, and gave product-embedded Annex I systems until August 2028. Article 50 was never in that package. Its transparency obligations attach to a functional category of system, not a risk tier, and they applied on schedule from 2 August 2026.

Four duties bite. Providers of systems that interact directly with people, meaning chatbots and voice assistants, must make clear the user is talking to a machine. Providers of systems generating synthetic audio, image, video or text must mark outputs in a machine-readable, detectable format. Deployers of emotion-recognition or biometric-categorisation systems must inform the people exposed to them. Deployers publishing deepfakes or AI-generated text on matters of public interest must disclose it. Disclosure must be clear and distinguishable at the latest at first interaction. Non-compliance carries fines up to 15 million euro or 3 percent of worldwide annual turnover, whichever is higher, and the burden of proving timely disclosure sits with the organisation, not the regulator.

The same date switched on the European Commission's supervision and enforcement powers over general-purpose AI model providers. GPAI obligations came into force on 2 August 2025 with a one-year adjustment period. That period has now expired. The AI Office can request documentation, run technical evaluations of models, demand risk-mitigation measures, restrict or withdraw a model from the EU market, and fine. Providers of GPAI models placed on the market before 2 August 2025 have until 2 August 2027.

The one piece of relief inside Article 50 is narrow: generative systems already on the market before 2 August 2026 have until 2 December 2026 to meet the machine-readable marking sub-obligation. Systems entering the market on or after 2 August get no runway at all. And the technical foundation is shakier than the statute implies. ETH Zurich research presented at ICML 2024 showed watermarking schemes previously considered safe could be scrubbed and spoofed at above 80 percent success rate for under 50 dollars per attack. A follow-on ICML 2025 attack defeats high-entropy-token watermarking. A publicly released tool now defeats Google's own SynthID detector on images its current Gemini models produced.

Sources: Cloud Security Alliance, The Register, EU Artificial Intelligence Act

Anyone who told their board that EU AI Act compliance was deferred to December 2027 now has a correction to make. Article 50 is the obligation most likely to touch an ordinary enterprise deployment, because every customer-facing chatbot and every AI writing assistant is in scope regardless of risk tier. Start with an inventory mapped to the four Article 50 categories, not to Annex III, because a compliance programme built around high-risk classification will miss most of what is now enforceable. The watermarking research matters commercially: a control with publicly documented bypasses cannot be shipped once and forgotten, so build a decision log that ties each technology choice to the state of the art on the day it was made. That log is the defensible artefact when a market surveillance authority asks.

🟡 The EDPB replaced the 2014 anonymisation test, and contracts no longer count

On 7 July the EDPB published draft Guidelines 02/2026 on anonymisation, the first substantive update to European anonymisation guidance in over a decade. The previous reference point was Article 29 Working Party Opinion 05/2014, which predates the GDPR itself, the current case law and the whole of generative AI. Alongside it the Board adopted Guidelines 03/2026 on web scraping in the context of generative AI. Both are open for public consultation until 30 October 2026.

The new test is three criteria. No record isolation, meaning no unique combination of attributes points to one individual. No linkage, meaning records cannot be matched to other data about the same person. No inference, meaning no specific and meaningful personal information can be deduced. Pass all three and data can be treated as anonymous and falls outside GDPR. Fail one and further analysis is required. Crucially, the framework is relative: following the CJEU ruling in EDPS v SRB (C-413/23), the same dataset can be anonymous for one organisation and personal data for another, depending on what each can realistically access.

Two findings should change practice. First, contracts do not do the work. A statutory prohibition on re-identification can remove certain means from the "reasonably likely" test, but a contractual prohibition cannot. The EDPB is explicit that contracts complement robust technical measures rather than replace them, a harder line than the ICO's more pragmatic UK position. Second, the guidelines say AI models and synthetic datasets are not automatically anonymous. If you can prompt a model and elicit personal information about a specific individual, the no-inference criterion fails. The Board also notes that agentic AI will keep driving down the time and cost of re-identification, so what counts as "means reasonably likely to be used" is a moving target.

Sources: EDPB, Lewis Silkin

This is the most consequential privacy document of the year for anyone putting enterprise data near a model. The practical read: if your AI programme rests on a data-sharing agreement and a promise not to re-identify, the EDPB has just told you that is not enough. Technical controls at the data layer are what survives the three-criterion test, which is exactly the case for tokenising or pseudonymising PII before it reaches a model, a vendor or a third-party API. Note the distinction carefully though - pseudonymised data remains personal data under Article 4(5) and GDPR still applies in full, so tokenisation reduces exposure and blast radius rather than removing the obligation. The 30 October consultation is a genuine opportunity for regulated firms to push back on the contracts position before it hardens.

🟡 The Cloud and AI Development Act pushes data residency from preference to procurement rule

The EU published the Cloud and AI Development Act in the Official Journal on 15 July 2026, the legislative counterpart to the sovereignty agenda that has been building since the Commission's April procurement moves and the May reporting that Brussels was weighing restrictions on US cloud platforms processing sensitive government data. The Act is designed to expand European cloud and datacentre capacity and to establish governance structures including a European Cloud and AI Agency alongside national competent authorities. Application is phased, with the first tier of cloud sovereignty requirements arriving in 2028 and the highest tier following in 2029.

The commercial signal arrived alongside it. On 23 July Atos launched Atos Sovereign Cloud, an application orchestration and modernisation platform aimed squarely at governments, defence, healthcare and critical infrastructure operators, combining enterprise cloud capability with sovereign controls over data management, operational independence and cyber resilience. That is a vendor reading the same demand curve that the Act codifies.

For DPOs and CISOs at UK and EU enterprises the practical effect is that data residency questions move out of the architecture discussion and into procurement and audit. The phase-in dates look distant, but sovereignty tiering will start showing up in tender requirements long before 2028, particularly in financial services, healthcare and the public sector supply chain.

Sources: European Parliament Legislative Train, AIwire, CNBC

Residency is the expensive way to solve a data problem and it is rarely the complete answer. Moving a workload into an EU region does nothing about what leaves that region inside a prompt or an API call to a US-hosted model. The cheaper and more durable control is to separate identity from data so that the sensitive fields never travel in the first place, which lets you keep using the best model or the best SaaS tool while the regulated data stays where it is governed. Firms that solve this at the data layer will find sovereignty tendering a formality. Firms that solve it by rebuilding infrastructure will pay twice.

SECTION 2 - What DPOs, CISOs and Data Leaders Are Dealing With

The pattern across all three stories is the same: the regulatory question has moved from policy to proof. Article 50 puts the burden of demonstrating disclosure on the organisation. The EDPB anonymisation framework asks you to defend a technical claim about identifiability rather than point to a contract clause. The Cloud and AI Development Act will eventually ask you to evidence where data sits and who can reach it. None of these are answered by a policy document or a supplier attestation.

That is compounded by what is actually happening inside enterprises. The Belgian supervisory authority's investigation into AI conversational applications found developers prioritising product velocity and scalability over GDPR, with specific concern about free-text interactions containing sensitive personal data, model retraining and onward sharing with external providers. That is not a chatbot vendor problem, it is a description of most enterprise AI pilots. The controls that hold up under the new anonymisation test are technical, sit close to the data, and work regardless of which model or vendor is downstream.

Three Things Worth Watching Next Week

1. First Article 50 enforcement signals - National market surveillance authorities can act from 2 August. Watch which regulator moves first and whether it targets a mainstream enterprise deployment or a deepfake case, because that sets the practical bar.

2. EDPB consultation responses - Submissions on the anonymisation and web scraping guidelines run to 30 October. Expect industry pushback on the contracts position and on the treatment of AI models as potentially non-anonymous.

3. GPAI documentation requests - The Commission's AI Office can now demand documentation and run technical model evaluations. The first requests will reveal how deep the evidentiary expectations go for anyone building on top of a GPAI model.

Questions? hello@paringa.ai

Paringa works with enterprises on AI security, data privacy and governance through a partner ecosystem, including data-layer privacy controls that tokenise and pseudonymise PII before it reaches a model or a third party. We help you prove the control, not just document the policy. | 3 August 2026 | © Paringa Ltd