EU AI Act August 2026: What QA Teams Need to Test

On 2 August 2026, the transparency rules in Article 50 of the EU AI Act started to apply. If your product includes a chatbot, an AI agent, an avatar or a feature that generates images, audio, video or text, these rules may now affect you, even if your company is outside the EU.

For QA teams, the change is practical. Requirements such as “people must be told they are talking to an AI” or “AI-generated content must carry a machine-readable mark” become things that can be tested, broken by a release and caught by a regression suite.

This essay explains what took effect, what the AI Omnibus changed in July 2026, and how to turn the obligations into a test plan.

Not legal advice. This essay is written for testers and engineers. It summarises official European Commission material as of mid-August 2026, but it is not legal advice. Whether and how the AI Act applies to your product depends on your role, your system and your market. Talk to qualified legal counsel before making compliance decisions.

What happened on 2 August 2026

The European Commission’s news article of 2 August 2026 summarises it: “new rules on the transparency of AI systems take effect.” The goal is to help people “recognise when they are interacting with AI or are exposed to AI-generated content.”

The AI Act itself entered into force on 1 August 2024 and applies in stages. Article 50 is one of those stages.

The two families of obligations

For providers (those who develop an AI system and place it on the market under their name):

• Systems that interact directly with people, “such as chatbots, AI agents, and avatars,” must be designed so people are informed they are interacting with AI, unless that is obvious.
• Outputs of generative AI systems must be “marked with effective, reliable, robust and interoperable machine-readable marks” so they can be detected as AI-generated or manipulated.

For deployers (those using an AI system under their authority in a professional capacity):

• Inform people exposed to emotion recognition or biometric categorisation systems.
• Clearly label deepfakes.
• Clearly label AI-generated or manipulated text published to inform the public on matters of public interest, where there has been no human review or editorial control.

These summaries come from the Commission’s FAQ on Article 50 and its quick facts page.

Who enforces it, and the penalties

Enforcement sits with national market surveillance authorities, the European AI Office for systems under its supervision, and the European Data Protection Supervisor when EU institutions are involved. The Commission lists fines of up to €15 million or 3% of global annual turnover for companies, with proportionality for SMEs and small mid-cap companies.

What the AI Omnibus changed

A week before Article 50 applied, the AI Omnibus entered into force on 27 July 2026. According to the Commission’s announcement, it changed several timelines and obligations. The changes most relevant to QA teams:

Change	What it means
High-risk AI systems in Annex III	Rules apply from 2 December 2027
High-risk AI embedded in products (Annex I)	Rules apply from 2 August 2028
Article 50(2) marking grace period	Until December 2026 for generative AI systems placed on the market before 2 August 2026
Deepfakes generated before 2 August 2026	No mandatory retroactive labelling, but encouraged
Regulatory sandboxes	Expanded access, plus an EU-level sandbox
Bias detection	Processing special categories of personal data allowed to detect and correct bias
AI literacy	Requirement for companies simplified

The grace period applies only to the machine-readable marking obligation. The obligation to tell people they are talking to an AI applies from 2 August 2026.

Why this is a QA problem, not just a legal one

Compliance documents describe what a system should do. Tests show what it actually does in each release. Article 50 obligations have the properties that make them good test targets:

• They are observable. A disclosure appears or it doesn’t. A mark is present in the output or it isn’t.
• They are easy to break. A redesign hides the disclosure. A new image pipeline strips metadata. A new model version changes response formats.
• They span teams. Product writes copy, design places it, engineering implements it, and a vendor may supply the model.

Testers are well placed to own the evidence that the obligations hold, release after release.

Testing AI interaction disclosure

The Commission’s guidance says the disclosure obligation applies when four cumulative criteria are met: the system is an AI system, it is designed for genuine two-way exchange, the interaction is direct, and it is with natural persons. Background or machine-to-machine systems fall outside it.

The guidance also says people “must be notified when they are interacting with an AI system from the start of the first interaction in a clear and distinguishable manner and in accordance with accessibility requirements.”

That sentence turns into test cases.

Test cases for disclosure

1. First interaction: the disclosure is visible before or at the first message, not after several turns.
2. Every entry point: web widget, mobile app, email assistant, voice channel and embedded partner widgets all show it.
3. Clear and distinguishable: it isn’t hidden in a tooltip, collapsed panel or footer.
4. Accessible: screen readers announce it, it has sufficient contrast, and it works with keyboard navigation and zoom.
5. Localisation: each supported language includes it.
6. Handover: when a conversation moves between an AI and a human agent, the user can tell which one they are talking to.
7. Voice and avatars: audio channels have an audible disclosure, and avatars have a visible one.
8. Regression: redesigns, A/B tests and new themes don’t remove it.

A Playwright check can cover the first and fourth cases:

import { test, expect } from '@playwright/test';
import AxeBuilder from '@axe-core/playwright';

test('chat widget discloses AI before the first reply', async ({ page }) => {
  await page.goto('/support');
  await page.getByRole('button', { name: 'Chat with us' }).click();

  const disclosure = page.getByTestId('ai-disclosure');
  await expect(disclosure).toBeVisible();
  await expect(disclosure).toContainText(/AI assistant/i);

  const results = await new AxeBuilder({ page })
    .include('[data-testid="chat-panel"]')
    .analyze();
  expect(results.violations).toEqual([]);
});

Accessibility tooling such as axe-core won’t tell you whether a disclosure is “clear,” but it will catch contrast and labelling issues that make it inaccessible.

The “obvious” exception

Disclosure isn’t needed when it is obvious the person is interacting with AI. The Commission says the exception should be “interpreted in a restrictive manner.” From a testing point of view, the safe approach is to treat disclosure as required unless your legal team has documented why it isn’t, and to record that decision next to the test.

Testing machine-readable marking of generated content

If you provide a generative AI system that produces synthetic audio, images, video or text, outputs must be marked in a machine-readable format and be detectable as AI-generated.

The guidance excludes some outputs, including short sequences of numbers, symbols or letters, source code, machine-to-machine outputs never exposed to humans, and some closed-loop production uses. Assistive functions for standard editing are also exempt.

The Code of Practice on Transparency of AI-generated Content is a voluntary tool providers can rely on to demonstrate compliance. If your company adopts it, the specific techniques it describes become your test oracle.

Test cases for marking

• Presence: every output type, whether image, audio, video or text, carries the mark your team has chosen.
• Survival: the mark survives your own pipeline, including resizing, compression, format conversion, CDN optimisation and download.
• Detection: your detection tool or API identifies marked outputs and doesn’t flag unmarked human content.
• Coverage: every feature that generates content is included, not just the flagship one.
• Exemptions: features your team classifies as exempt, such as standard editing, are documented and reviewed.
• Vendors: if you use a third-party model, check what marking it applies and whether your processing preserves it.

The survival test matters most. A common way to break marking is an innocent image optimisation step that strips metadata. Put a check after every transformation in the pipeline.

Testing deployer obligations

Deployers have their own tests to run.

Deepfake labelling

Deepfakes must be disclosed “upon first exposure at the latest” with labels that are understandable without special tools. The guidance is explicit that deployers “cannot simply rely on the machine-readable marking” applied by the provider. A visible or audible label is needed. The EU has published optional icons for this.

Test that the label appears in every place the content appears: feeds, previews, thumbnails, shares and embeds.

AI-generated text on public interest topics

Text published to inform the public on matters of public interest, such as politics, public health or consumer safety, must be labelled when it hasn’t had human review or editorial control. If your workflow relies on human review to fall outside this rule, test that the review step can’t be bypassed and that it is logged.

Emotion recognition and biometric categorisation

If you deploy these systems, people exposed to them must be informed. Test that the notice appears before exposure, in all channels.

Provider or deployer: who tests what

Many companies are both. You might deploy a vendor’s model inside your own chatbot and so be a deployer for some features, while building and selling a generative feature under your own name and so be a provider for others. The test plan should follow the role per feature, not per company.

Role for this feature	Main Article 50 checks
Provider of a conversational AI system	Disclosure from the first interaction, on every channel, accessible
Provider of a generative AI system	Machine-readable marks on outputs, detectability, survival through processing
Deployer of emotion recognition or biometric categorisation	Notice to exposed people
Deployer publishing deepfakes	Visible or audible label on first exposure
Deployer publishing public interest text	Label, or proof of human review or editorial control

Write the role into each feature’s test plan so reviewers can see why a test exists.

Agents need disclosure tests too

The Commission’s guidance names “AI agents” alongside chatbots and avatars. That matters because agents often act outside a chat window. An agent might send an email, post a comment, book a meeting or message a customer on your behalf.

Where an agent interacts directly with people, the four criteria above apply, so test the channels agents use:

• Outbound messages: emails or chat messages written and sent by an agent identify it as AI where the criteria apply.
• Voice calls: an audible disclosure at the start of the call.
• Multi-step conversations: the disclosure isn’t lost when a thread continues across days or channels.
• Machine-to-machine steps: document which steps never reach a person and so fall outside the obligation.

These cases are easy to miss because they don’t appear in the product’s main UI. An inventory of every place an agent can speak is the first deliverable.

Building an evidence trail

Regulators and auditors will ask how you know the obligations are met. A test suite can produce the evidence:

1. Map each obligation to tests, with a short requirement ID such as A50-1-disclosure.
2. Run them in CI on every release that touches AI features.
3. Store results with build IDs, dates and screenshots.
4. Review exemptions on a schedule and record decisions.
5. Track vendor changes, since a new model version can change outputs.
6. Report coverage of AI features to product and legal teams.

This is ordinary test management. The difference is that the audience now includes compliance.

High-risk systems: what to prepare for now

High-risk obligations now apply from 2 December 2027 for Annex III systems and 2 August 2028 for those embedded in regulated products. That delay is not a reason to wait. If your system might be high-risk, the testing capabilities it will need, such as documented accuracy and robustness testing, bias evaluation and logging, take time to build.

The Omnibus also allows processing special categories of personal data to detect and correct bias, under the conditions in the regulation. Discuss with your legal and privacy teams what this means for test data, rather than assuming.

Practical evaluation tools help here. We compare options in Promptfoo vs DeepEval vs Langfuse, and the broader approach is in How to test AI agents.

A 30-day plan for QA teams

Week 1: inventory. List every AI feature, its channel, whether your company is provider or deployer, and the content types it generates.

Week 2: classify. With legal input, decide which Article 50 obligations apply to each feature, and document exemptions.

Week 3: automate. Write disclosure and labelling tests in your E2E suite, and marking tests in your content pipeline.

Week 4: report. Add the tests to CI, set up a dashboard and agree how failures block releases.

Common mistakes to avoid

• Testing only the main chat widget and missing mobile, email or partner integrations.
• Assuming the vendor handles marking without checking that your pipeline preserves it.
• Hiding disclosures in terms of service or tooltips.
• Ignoring accessibility, which the guidance explicitly mentions.
• Treating the Omnibus delay as a pause on all AI Act work. Article 50 still applies from 2 August 2026, and the marking grace period is short.
• Not involving legal. Test design should follow documented legal interpretations, not guesses.

Frequently asked questions

When did the EU AI Act transparency rules start to apply?

Article 50’s transparency rules apply from 2 August 2026, according to the European Commission. Generative AI systems placed on the market before that date have a grace period until December 2026 for the machine-readable marking obligation only.

Does the EU AI Act apply to companies outside the EU?

According to the Commission’s FAQ, providers outside the EU are also subject to the Act if the output of their AI system is used in the EU. Check with legal counsel for your situation.

What do QA teams need to test under Article 50?

Typical tests cover AI interaction disclosure on every channel, accessibility of that disclosure, machine-readable marks on generated content and their survival through your pipeline, and visible labels for deepfakes where you are a deployer.

When do high-risk AI obligations apply after the Omnibus?

The Commission says rules for Annex III high-risk systems apply from 2 December 2027, and for high-risk AI embedded in products under Annex I from 2 August 2028.

What are the penalties for breaking the transparency rules?

The Commission lists fines of up to €15 million or 3% of global annual turnover for companies, up to €750,000 for EU institutions, with proportionality for SMEs and small mid-caps.

Is this essay legal advice?

No. It is a testing-focused summary of official material. Consult qualified legal counsel for compliance decisions.

Further reading on CheckOps

• How to test AI agents
• Promptfoo vs DeepEval vs Langfuse
• LLM testing, evals and vibe-coding quality
• MCP, OWASP LLM Top 10 and LLM testing
• Best AI testing tools 2026