AI search results: Who is liable when the AI lies?

AI answers are now part of everyday life: Google shows AI Overviews and AI Mode, users ask ChatGPT or Perplexity, and many companies use their own chatbots. The answers are often accurate, but they can be factually wrong (“hallucinations”) or contain legally problematic content.

For companies, this raises the question of who answers for such errors. In this article, we show where responsibility lies, which three risks matter most in marketing, what the AI Act requires, and which measures keep the risks under control. This article is an assessment from an online marketing perspective and does not replace legal advice.

The operator answers for what an AI says

You cannot sue an AI. Responsibility lies with whoever operates it or distributes its answers. For companies, that means: whatever an AI system says on the company's behalf reflects back on the company.

The Hamburg Regional Court held the operator of Grok responsible

A case from fall 2025 shows that this is not theory. Asked by a user on X, the chatbot Grok claimed that the association Campact was funded with tax money. Campact then obtained a preliminary injunction against the company X.AI from the Hamburg Regional Court. According to Campact, the court classifies the statement as an untrue statement of fact and holds X.AI directly responsible for it. If the chatbot repeats the claim, a fine of up to €250,000 may be imposed.

The decision is an urgent ruling which, according to Campact, X.AI could still contest. But it shows the direction: the fact that a statement was generated automatically does not relieve the operator. For your company, this mainly concerns three types of errors:

• False information: The AI gives wrong product details, invents prices or test results.
• Reputational damage: The AI makes false statements about competitors or individuals.
• Discrimination: The AI treats user groups unequally because of biased training data.

Three risks are decisive for your marketing

Risk 1: The AI claims something false

Language models are not knowledge databases. They predict the most likely next word and fill gaps with statements that sound convincing but can be wrong. The risk has two sides:

1. Your own AI: Your shop chatbot gives wrong product details, invents warranty promises, or states a wrong price. In the most harmless case, this annoys customers; in the worst case, legal claims follow. Every false promise also costs margin if you honor it as a goodwill gesture.
2. Other people's AI: An AI system misrepresents your product or your company, for example in an AI answer in Google Search. As the Campact case shows, those affected can take action against the operator. Which claims exist in an individual case must be assessed by a lawyer.

In both cases, the false statement is out there, and you have to respond, as the cause or as the party affected. We therefore recommend regularly checking how AI systems present your company. We describe how in our article on Generative Engine Optimization.

Risk 2: The AI content infringes third-party rights

If an AI generates text, an image, or program code that is too similar to an existing protected work, this may constitute copyright infringement. This is possible if the original was part of the training data.

What matters for your company: whoever publishes such content on the website or uses it in a campaign distributes it, not the AI. Then there is the reverse question: whether texts an AI creates without significant human editing are themselves protected by copyright is doubtful. Do not rely on competitors not being allowed to copy such content.

Risk 3: The chatbot processes personal data

As soon as you integrate an AI chatbot or an AI search into your website, the General Data Protection Regulation (GDPR) applies. Users inevitably enter personal data, from the IP address to inquiries that may contain names, health data, or financial concerns. Under Art. 9 GDPR, health data is one of the special categories of personal data that are specially protected.

This results in specific tasks for operators:

• Transparency: You must be able to explain to users which data is processed and stored, and fulfill access or deletion requests. With many AI systems, this is hard to trace.
• Use of inputs: If the provider uses your users' inputs to train its own models, you need a legal basis for this and must inform your users.
• Data processing: If an external provider processes the data on your behalf, a data processing agreement under Art. 28 GDPR is required.
• Transfers to third countries: Many AI models run on servers in the US. A transfer there requires a basis under Chapter V of the GDPR, such as an adequacy decision or standard contractual clauses.

If the processing is likely to result in a high risk to the rights and freedoms of users, Art. 35 GDPR requires a data protection impact assessment in advance. This applies in particular to large-scale processing of special categories of data or to profiling on which decisions with significant effects for those concerned are based.

In marketing, the AI Act mainly requires transparency

In addition to the GDPR, the EU's AI Act applies. It regulates AI according to its risk: the higher a system's risk to fundamental rights or safety, the stricter the rules. According to the EU Commission, it provides no rules for minimal-risk systems such as spam filters.

Transparency obligations have applied since August 2026

For most marketing applications, the decisive rules are the transparency obligations that have applied since August 2, 2026:

1. Chatbots: People must be able to tell that they are talking to a machine.
2. AI-generated content: Providers of generative AI must ensure that generated content is recognizable as such. Deepfakes and texts that inform the public on matters of public interest must be labeled clearly and visibly.

In addition, the AI literacy obligations have applied since February 2, 2025. Clarify with your legal department which training and documentation this means for your team.

Strict obligations for high-risk AI apply from December 2027

Stricter rules apply as soon as AI makes decisions about people, for example when selecting applicants or checking creditworthiness. Such high-risk systems require, among other things, human oversight, logging, and risk management. With the AI Omnibus, which entered into force on July 27, 2026, the EU postponed these obligations: for stand-alone high-risk systems to December 2, 2027, and for AI in regulated products to August 2, 2028.

What the AI Act means for your company

Does the AI Act apply to all companies? The AI Act addresses companies that develop or use AI systems. Which obligations apply specifically depends on the system's risk level and on your role as provider or deployer. For small and mid-sized companies, it provides partly simplified requirements, for example for technical documentation.

Does every AI application have to document how its decisions are made? No. Extensive documentation and risk management obligations concern high-risk systems. A product recommendation in an online shop usually does not fall under them. For chatbots and AI content, the transparency obligations apply.

Which areas are particularly affected? High-risk fields of use include recruitment, education, lending, and critical infrastructure. If you use AI there, clarify the obligations early with your legal department.

Four measures to keep your use of AI under control

The risks are real, but no reason to do without AI. They require a clear process. The following measures are guidance from an online marketing perspective; have them reviewed by your legal department or a specialized law firm.

1. Human review: Every piece of AI content is reviewed by a person before publication.
2. Transparency and labeling: Chatbots are recognizable as AI, and labeling follows the requirements of the AI Act.
3. Contracts and consents: Data processing, data transfers to third countries, and consents are reviewed and documented.
4. Internal guidelines: Fixed rules define which tools are used, which data may go into them, and who approves.

We recommend seven points for the technical implementation

• Closed environment: AI applications are developed and tested in an isolated environment, such as a private cloud or your own network.
• Where data is processed: You decide deliberately whether data is processed on your own servers, in a private cloud, or by the provider.
• Security testing: Regular penetration tests uncover vulnerabilities.
• Roles and access: Access rights are assigned by role and reviewed regularly.
• Anonymization: Sensitive data is anonymized or pseudonymized before it reaches an AI model.
• Logging: AI activities, data access, and decisions are logged so they are traceable and auditable.
• Model control: Models are regularly reviewed, versioned, and documented.

Conclusion: responsibility for AI stays with the company

The legal risks of AI are real: false statements, third-party rights, data protection, and new obligations under the AI Act. But they are manageable when it is clear who reviews, who approves, and which data goes into which systems.

AI is a valuable tool, but not a responsible employee. Legal responsibility cannot be handed over to an algorithm. We advise you on how to use AI safely in marketing and on your website, and how your company is represented correctly in AI answers. More on our page about GEO.