AI Act: What changed in 2026 for ChatGPT, Copilot, and chatbots

The short answer

Which obligations apply depends on where the AI is doing its work:

• If employees use ChatGPT, Copilot, or similar tools, the business has to take measures that support their AI literacy. Since July 2026, nobody has to guarantee a particular level. No certificate is needed, and Germany has no fine for this.
• If a chatbot answers on the website or an AI agent writes to customers on its own, it has to be clear by the first interaction at the latest that an AI is answering. This has applied since August 2, 2026, including to chatbots that have been running for a while.
• Published AI-generated text only needs a label in exceptional cases. AI images, videos, and audio that look or sound deceptively real always need one.
• AI that pre-sorts job applications or rates performance usually counts as high-risk AI, with obligations from December 2, 2027. Using AI to read employees’ emotions has been prohibited since February 2025, except for medical or safety reasons.

What applies when: the EU made changes in July

The AI Act has been in force since August 2024 and applies in stages. Many guides date from before this summer and are out of date on AI literacy, the deadlines for high-risk AI, and supervision in Germany.

On July 27, 2026, an amendment to the AI Act took effect: the Digital Omnibus on AI. It weakened the AI literacy obligation and pushed back the high-risk deadlines by at least a year. It left the transparency obligations for chatbots, AI-generated text, and deepfakes unchanged in substance. The one exception is the machine-readable marking that makers of text and image generators have to build in: older systems got until December.

Two days later, Germany’s act implementing the AI Act, the KI-MIG, took effect. Since then, the Federal Network Agency (Bundesnetzagentur) has been the supervisory authority, point of contact, and complaints office.

Provider or deployer: building a chatbot brings more obligations

The AI Act assigns obligations by role. The provider owes the notice in the chat window. The deployer owes the labeling of deepfakes and certain text. A business can be both at once.

A business with licenses for ChatGPT or Microsoft 365 Copilot is a deployer of those tools. That’s the AI Act’s term for anyone who uses an AI system under their own authority in a professional context.

If a business builds its own chatbot on a language model, for example through the OpenAI API or as an agent in Copilot Studio, and puts it on its website, it’s the provider of that chatbot. The Commission’s guidelines name the in-house chatbot explicitly as an example. A provider is anyone who develops an AI system, or has one developed, and puts it into service under their own name.

So the same holds when an agency or contractor built the chatbot. The contractor doesn’t put it into service under its own name; the business does. A business that embeds a vendor’s finished chat product unchanged remains a deployer, and the vendor is the provider.

According to the Commission, a business that modifies someone else’s AI system, for example with its own training data, and puts it into service under its own name becomes the provider of the new system. The guidelines don’t say where the line falls for simple setup, such as adding the company’s own FAQ text. In that case, I would make sure the notice is there myself. It costs one line.

AI literacy: brief people and keep a record, no certificate needed

The AI literacy obligation in Article 4 has applied since February 2025 to every business that uses AI professionally. Originally, providers and deployers had to take measures “to ensure, to their best extent, a sufficient level of AI literacy” among their staff. Since July 27, 2026, the measures only have to “support the development” of that literacy, and the new text says explicitly that nobody has to guarantee a specific level.

According to the European Commission’s questions and answers on AI literacy (as of October 20, 2026), that means in practice: no certificate, no duty to test employees’ knowledge, and no dedicated AI officer. Its example is employees who use ChatGPT to write ad copy or translate text. They should know the typical risks, such as the tool making up facts and stating them convincingly. As a record, the Commission suggests an internal log of trainings.

Neither the AI Act nor Germany’s KI-MIG provides a fine for a lack of AI literacy; the KI-MIG’s penalty provisions don’t mention Article 4. Still, in its guidance paper from June 2025, the Bundesnetzagentur points out that a lack of AI literacy can be seen as a breach of the duty of care, especially if it causes damage.

So I would take the obligation seriously, not least because it targets the real risk: AI mistakes that land unchecked in a quote or a customer email. It doesn’t take much:

• one page of rules: which tools are approved, which data doesn’t belong in them, that a person checks every result before it goes out, and who to ask with questions
• a half-hour briefing with examples from the team’s own work, recorded with the date, content, and participants, as the Bundesnetzagentur also recommends
• the same for new colleagues when they start, plus a refresher when new tools come in

How Katio Tech sets up rules and onboarding material for a team is described under AI in business.

The AI Act doesn’t regulate whether customer data may go into a tool. The GDPR does, for example through a data processing agreement with the tool’s vendor.

Chatbots and AI agents: the notice belongs in the chat window

Article 50(1) requires AI systems that interact directly with people to be designed so those people learn they’re dealing with an AI. It has applied since August 2, 2026, without a transition period, including to chatbots that were already running.

According to the Commission’s guidelines, this covers chatbots in customer service, complaints handling, and online stores, voice assistants on the hotline, and AI agents that write emails, book appointments, or place orders on their own. Agents also have to say on whose behalf they’re acting.

Spam filters, translation tools, and search features that don’t generate anything aren’t covered. A customer service team that has an AI draft replies, checks them, and sends them itself doesn’t need a label either: the customer is dealing with a person.

No notice is needed when it’s obvious that an AI is answering, but the Commission reads this exception narrowly. An internal assistant for trained employees who know they’re using AI needs no notice. A chatbot on the website that customers write to does.

The law doesn’t prescribe what the notice looks like. It has to be clear and accessible, so it also has to work with a screen reader. And it has to appear by the first interaction at the latest. The guidelines give two examples: a first message in which the chatbot says it’s an AI, and a notice right next to the input field. A greeting such as “You’re chatting with an AI assistant, not a person” meets that. According to the Commission, these aren’t enough:

• a notice only in the terms and conditions or the privacy policy
• a general statement such as “This website uses AI”
• the label “assistant” without the word AI
• a name and photo that suggest a human

If someone asks whether a person is answering, the chatbot has to answer truthfully.

Responsibility lies with the provider: the business itself if it built the chatbot or had it built. With a purchased product, the vendor has to build in the notice; it’s still worth checking that it’s actually switched on.

According to Microsoft, agents built in Copilot Studio greet users by default with the sentence “Just so you are aware, I sometimes use AI to answer your questions.” Anyone who customizes the greeting should keep the notice. For an agent that generates every answer with AI, though, “sometimes” is rather vague. I would say it as plainly as the greeting above.

AI-generated text and images: only two cases need a label

The first case is deepfakes. An AI photo that looks like a real shot of the company’s own workshop and staff has to be labeled as artificially generated. Under the AI Act, deepfakes are AI-generated images, audio, or video that resemble real people, places, objects, or events and would falsely appear authentic. A clearly drawn illustration isn’t a deepfake.

The second case is text published to inform the public on matters of public interest. According to the Commission, ad copy and product descriptions aren’t included, as long as they make no claims about health, safety, or sustainability. A listed company’s investor report is included, and an expert article on a company blog can be.

Even then, no label is needed if a person has checked the content of the text, including the facts, and a person or the company visibly holds editorial responsibility, for example through the legal notice. A pass with the spell checker or a quick sign-off isn’t enough. And if the AI rewrites the text again after approval, the exception no longer applies.

Content published before August 2, 2026, doesn’t need a label after the fact. Text the AI wrote earlier but that is only published after that date does fall under the new rules.

The invisible, machine-readable marking of AI content is up to makers such as OpenAI, Microsoft, or Google, not the businesses using their tools. A voluntary code of practice from the Commission, published in June 2026, shows what labels can look like, including an EU icon for AI content.

Job applications and performance reviews: high-risk from December 2027

AI in HR usually counts as high-risk AI under Annex III: when it screens or filters applications or evaluates candidates, when it influences decisions on promotions or terminations, assigns tasks based on employees’ behavior, or monitors and evaluates their performance.

From December 2, 2027, deployers of such systems have their own obligations under Article 26. They have to use the system according to the provider’s instructions for use, assign oversight to a person with the necessary competence and authority, monitor its operation, and keep the automatically generated logs for at least six months. Before using it in the workplace, they have to inform workers’ representatives, such as the works council, and the employees affected.

A business introducing applicant management with AI pre-screening today should ask the vendor how the product will meet these requirements from 2027.

AI that infers employees’ emotions at work, for example from their voices in a call center, has been prohibited since February 2, 2025, except for medical or safety reasons.

Supervision and fines: anyone can file a complaint

In Germany, the Bundesnetzagentur is responsible; for regulated financial companies, BaFin is. Anyone can file a complaint about a violation with the Bundesnetzagentur’s complaints office, including about a chatbot without a notice.

Violations of the transparency obligations and of the obligations of deployers of high-risk AI can cost up to €15 million or 3% of the previous year’s worldwide annual revenue. For prohibited practices, it’s up to €35 million or 7%. As a rule, the higher amount applies; for small and medium-sized businesses, the lower amount applies to all fines. For a business with €3 million in annual revenue, the ceiling for a violation of the transparency obligations is therefore €90,000. When setting the amount, authorities have to take into account the interests of small businesses and their economic survival.

For a first assessment of whether and how the AI Act affects a particular system, the Bundesnetzagentur offers the KI-Compliance Kompass in its AI Service Desk (in German). The result is guidance, not a binding answer.

What to do now

First, list where AI is at work in the business: the tools employees use, chatbots and agents, and AI features in specialized software, especially in HR. Then, in this order:

• Give chatbots and agents that deal with customers a notice that an AI is answering. That has been mandatory since August.
• Write down rules for the AI tools, brief employees, and record the briefing with the date.
• Decide who checks the content of published text, and label deceptively real AI images.
• Prepare AI in HR for the high-risk obligations by December 2027.

For the business at the start, that means the chatbot gets one line in its greeting. Everyone who uses ChatGPT or Copilot gets one page of rules and a half-hour briefing. The tools keep running as before.