Study guide · Law, Data Protection and AI
Law, data protection and AI in a website project
A website processes data about people, shows images with third-party rights, runs on free software and uses AI more and more often. On this page you follow one client project through all of these topics. Afterwards you can spot the typical risks, you know the matching tools in TYPO3, and you know when a legal expert needs to get involved. The page gives you basic knowledge for consulting and is not legal advice itself.
The project and your role
The example on this page is the Seetal holiday region. A tourism association is having its website rebuilt with TYPO3. The plans include a contact form, a login for hosts, a YouTube video, a map and lots of photos. Marketing wants an AI to write excursion tips and wants to offer an AI chatbot. For testing, the agency needs a copy of the database.
As a consultant you keep the big picture. You spot issues, explain technical options and document risks. Whether something is legally permitted is not yours to decide in a binding way. That is the job of lawyers and data protection officers.
The GDPR in a nutshell
The General Data Protection Regulation, GDPR for short, governs how personal data is handled in the European Union. It defines when such data may be processed, which rights data subjects have and which duties apply to controllers. It also applies to services aimed at people in the EU. Copyright and software licences are separate areas of law.
Personal data is any information that can identify a person directly or indirectly. The Tutorial for Editors explicitly lists phone numbers, email addresses, IP addresses, nicknames, pictures and car licence plates. In the Seetal project, the host number and a visitor's IP address are therefore personal data too.
Article 9 lists special categories that are specially protected. They include health data, racial or ethnic origin, political opinions, religious or philosophical beliefs and trade union membership. Genetic data, biometric data used to uniquely identify a person and data about sex life or sexual orientation also count. They may only be processed under strict conditions. A business email address or a postcode is personal data, but not specially protected.
In English texts you will meet two technical terms. PII stands for Personally Identifiable Information. PHI stands for Protected Health Information.
The principles: Article 5 sets out the rules for every kind of processing:
- Lawfulness, fairness and transparency.
- Purpose limitation, so data is only used for defined purposes.
- Data minimization, so you only collect as much data as the purpose needs.
- Accuracy of the data.
- Storage limitation, so data is only kept as long as the purpose requires.
- Integrity and confidentiality, so data is protected against loss and unauthorized access.
- Accountability, so you can prove that you comply.
Collecting data "just in case" contradicts purpose limitation and data minimization.
The legal bases: Under Article 6, every processing needs one of six legal bases. They are consent, a contract, a legal obligation, vital interests and a task in the public interest. The sixth is the legitimate interest of the controller after weighing the interests involved. The fact that collection is technically possible, or that marketing would like it, is not enough.
Consent, transparency and data subject rights
Consent is only valid if it meets these conditions:
- It is freely given and applies to a specific case.
- The person was informed beforehand.
- It is unambiguous, that is a clear, affirmative action.
- It can be withdrawn at any time, as easily as it was given.
A pre-ticked box is not enough, and neither is simply continuing to browse.
Technically necessary cookies, for example for a login, usually need no consent. For non-essential cookies and services, consent is the norm, so for statistics, marketing and embedded third-party content. A consent management tool asks visitors, stores their choice and only loads services after they agree.
Transparency: Data subjects must be informed when their data is collected. On websites this happens through an easily reachable privacy policy. It names the controller, the data, the purposes and the legal basis. It also states how long data is stored and which rights people have.
Data subject rights: Every person can request access to the data stored about them. They also have rights to rectification, erasure, restriction, data portability and to object to certain processing. Seeing other people's data is not one of these rights. So you should know where TYPO3 stores the data of a frontend user or the entries of a form.
External content: videos, maps, buttons and fonts
The Seetal team wants to embed a YouTube video, a Google map and a like button. All three come from the provider's server. As soon as they load, the provider receives data, at least the IP address, and often sets cookies. This happens before the visitor can decide anything. That is why prior consent is usually needed. An iframe changes nothing, because it also loads from the third-party server. A notice like "We use cookies" does not replace consent.
For a privacy-friendly setup you suggest:
- A two-click solution first shows a placeholder and only loads the external content after an active click.
- As an alternative, a consent management tool only loads the services after consent.
- The privacy policy informs about all embedded services.
- Like buttons are replaced by simple share links without third-party scripts.
TYPO3 helps a little. The Core embeds YouTube through youtube-nocookie.com by default and adds the "do not track" parameter for Vimeo. This reduces the data transfer but does not prevent it.
Fonts and other resources: Every file from a third-party server means a connection to that server, including the IP address. This applies to fonts, scripts and images. In 2022, for example, a German court objected to a website that loaded Google Fonts dynamically without consent. The simple fix is to host such resources locally on your own server.
Duties in the project: contracts, third countries and data breaches
The tourism association is the controller. The hosting company and the agency process data on its behalf. That requires a data processing agreement.
Marketing wants an analytics tool that transfers data to a country outside the EU. Here you raise three questions:
- On what basis does the transfer to the third country take place, for example an adequacy decision or standard contractual clauses?
- Is there a data processing agreement with the provider?
- Does the website obtain visitors' consent?
You do not assess the answers yourself. You involve the data protection officer or legal counsel.
Data breaches: Suppose the names and email addresses of newsletter subscribers are stolen. Articles 33 and 34 set out what to do next.
- The controller reports the breach to the supervisory authority without undue delay, where feasible within 72 hours of becoming aware of it. This is only unnecessary if the breach is unlikely to result in a risk.
- If there is a high risk for the people affected, the controller notifies them as well.
- The controller documents the incident and the measures taken.
The website operator remains responsible, even if a service provider is involved. A processor such as the hosting company must inform the controller without undue delay. You assess the risk together with the data protection officer or legal counsel.
Fines: For serious violations, Article 83 provides for up to 20 million euros or 4 percent of worldwide annual turnover, whichever is higher. For other violations the limits are 10 million euros or 2 percent. The actual amount always depends on the individual case.
Managing data consciously in TYPO3
The old website holds form entries, frontend users and system logs going back ten years. This is where data governance helps. For each type of data you define what it is for, who has access, how long it stays and how it is deleted. Policies and clear responsibilities make sure this happens permanently. Every piece of data goes through a lifecycle with five phases: creation, storage, usage, retention and disposal.
The core idea is simple: data that no longer exists cannot leak. The GDPR also demands storage limitation. In TYPO3 you put this into practice like this:
- You set retention periods for each type of data.
- The Scheduler runs clean-up jobs regularly, triggered by a cron job.
- The task "Table garbage collection" deletes old entries from configured tables, for example from the system log.
- The task for IP anonymization is called "Anonymize IP addresses in database tables" in the Scheduler. It shortens stored IP addresses once they are older than a chosen number of days. It is prepared for the system log, and you can add your own tables. That way the log stays useful.
Deleted does not mean gone: When someone deletes a record in the backend, TYPO3 usually just sets a "deleted" flag. The record stays in the database so that it can be restored with the Recycler. For personal data that is often not enough. The Scheduler task "Remove deleted records" from the Recycler extension removes such records for good after a set number of days. A console command does the same.
vendor/bin/typo3 cleanup:deletedrecordsReal data on test and development systems
To fix a bug, the agency wants to pull the live database onto the developers' laptops. Data protection applies there too, even if it is only for a short time. Local machines and staging systems are often less protected than production. Every copy of real data increases the risk, for example if a laptop gets lost.
These measures are appropriate:
- Data minimization: only transfer the tables and records that are really needed.
- Anonymization: remove names, email addresses and IP addresses beforehand, or replace them with test data.
- Pseudonymization, if relationships between records must be kept.
- Encrypt the laptops and delete the copies when the work is done.
An unencrypted copy by email, a public download link or plain-text passwords are off limits.
With anonymization, the link to a person is removed permanently. With pseudonymization, names or addresses are replaced by identifiers. With extra knowledge, such as a mapping table, the link can be restored. So pseudonymized data is neither deleted nor encrypted. It remains personal data and still needs protection, but it lowers the risk.
Copyright and AI content
Images, videos and texts are usually protected by copyright, even if they are freely available online. You may not simply use a picture from a Google image search. You need permission from the rights holder or a suitable licence. Renaming the file or running a non-commercial website changes nothing. Violations can lead to expensive cease-and-desist letters.
You can get licences from commissioned photographers, from stock image services or as Creative Commons. The CC BY licence allows use and adaptation, but requires crediting the author in the requested form, together with a licence notice. The additions NC for "non-commercial" and ND for "no derivatives" restrict use further. No fee is paid to Creative Commons. In TYPO3, editors maintain such details in the file metadata.
Even your own photos are not always free. The Tutorial for Editors warns that even images you created yourself can infringe rights, for example with artworks or cartoon characters. A photo of a famous sculpture in a park can touch third-party rights. Some countries allow photos of works that stand permanently in public places, others restrict this. So it depends on the individual case and the country.
Photos of people are also personal data. To publish a portrait of an employee on the website, you usually need her consent. Employees do not generally have to put up with publication, not even with a small picture or without a name. There are exceptions, for example for persons of public interest. When in doubt, get legal advice.
AI-generated content must be checked before publishing, as the learning objectives explicitly state. If the Seetal team has an AI create texts and images, you raise three risks:
- Factual errors, because an AI can produce plausible-sounding but false statements.
- Third-party rights, if a result resembles a protected work.
- Responsibility for published content stays with the website operator.
So AI content is neither automatically legally safe nor generally forbidden. How far it is protected itself is partly unresolved and a question for legal counsel.
The GPL: freedoms and duties
TYPO3 is licensed under the GNU General Public License, GPL for short, more precisely under "GPL version 2 or later". Extensions in the TYPO3 Extension Repository, or TER, must also be under the GPL. For the client this means no licence fees and no dependency on a single vendor.
The GPL guarantees four freedoms: to use the software for any purpose, to study the source code, to modify the software, and to distribute copies or modified versions. For this the source code must be available. Charging money for the software is allowed.
Copyleft: Anyone who distributes GPL software or a work based on it must distribute it under the GPL again. That keeps the freedoms intact. If you distribute a modified version, these rules apply:
- The modified version is also under the GPL.
- The recipients get access to the source code.
- Licence and copyright notices are preserved.
Additional restrictions on the freedoms are not allowed. Extensions build on TYPO3, so they are usually under the GPL too. What always matters is distribution:
- Extension for a single client: The agency does not have to publish the code. The GPL only requires source code for the people who receive the software. The client gets it with all GPL rights and may pass it on in turn.
- Extension sold to many clients: Selling is allowed, but every recipient gets the source code with the GPL freedoms. Encrypting the code or forbidding redistribution by contract is incompatible with that. A business model therefore relies on services such as support or updates.
- Website content and design: The GPL covers the software code. Texts and images on the website are not affected. Running a website is also not a distribution of the software to its visitors. So the client does not have to open up the website, and commercial websites are allowed.
How language models work and why they fail
A large language model, LLM for short, produces text step by step. Based on patterns in its training data, it keeps predicting the next likely piece of a word. These pieces are called tokens. So an LLM does not look anything up in a verified database, does not copy a search result and does not ask a human.
This leads to the typical weaknesses:
- Fluent does not mean correct. Good wording says nothing about accuracy, either way.
- Hallucinations are plausible-sounding but false or invented statements, for example made-up TypoScript properties or sources. The model is not lying on purpose. It produces what is likely instead of what is verified.
- Varying answers: The same question can get different answers on different attempts.
- Training cut-off: A model learns from data up to a cut-off date. It does not know about later events, such as a new TYPO3 version, without additional sources.
Generative or retrieval-based: A search engine or knowledge base is retrieval-based. It finds existing documents and returns them with their source. An LLM is generative and writes new text. The combination is called Retrieval-Augmented Generation, or RAG.
AI rules and the limits of consulting
AI is not governed by the same rules worldwide. Regulation differs from jurisdiction to jurisdiction and keeps evolving. The EU has its own law, the AI Act. Other countries regulate AI differently or in separate laws.
Which rules apply depends on where the system is used, who is affected and the role of the company. The AI Act also applies to providers outside the EU if the output is used in the EU. So the vendor's home country is not the only thing that counts.
The risk-based approach: The AI Act classifies AI systems by their risk, not by price or technology.
- Some practices are prohibited, for example social scoring.
- High-risk systems are subject to strict obligations.
- Certain systems have transparency obligations. A chatbot, for example, must make clear that you are talking to an AI, unless that is already obvious.
- All other systems face only minimal requirements.
Provider and deployer: The provider develops an AI system and places it on the market. The deployer uses it under its own authority. If an agency sells an AI-based translation module for TYPO3, it is the provider. The client who uses it on its website is the deployer. Both can have their own obligations, depending on the risk classification.
AI literacy: Article 4 requires providers and deployers to ensure a sufficient level of AI literacy among their staff. This duty has applied since February 2025, and it also covers companies that only use AI. An agency that uses AI tools in client projects therefore trains its people on possibilities, limits and risks. Appropriate means matching the tasks, prior knowledge and context of use.
When a legal review is needed: As a consultant you point out legal questions, but you do not give binding advice. You explicitly recommend a qualified legal review when:
- AI influences decisions about people, for example job applications or loans, because high-risk rules may then apply.
- Personal data goes to AI services outside the EU.
- A client wants a binding answer on whether their use of AI is lawful.
A low-risk application usually needs no review of its own, for example an AI that only checks the spelling of an internal draft and stores nothing.
Key points in brief
- Personal data is anything that makes a person identifiable, including IP address and photo.
- Every processing needs a legal basis and follows principles such as purpose limitation and data minimization.
- Consent is freely given, informed, unambiguous and can be withdrawn at any time.
- External content sends the IP address away when it loads. Remedies are two-click, a consent tool and local hosting.
- A data breach is reported within 72 hours where feasible.
- Scheduler tasks clean up tables, anonymize IP addresses and remove deleted records for good.
- Pseudonymized data remains personal data.
- Third-party images only with a licence, photos of people usually only with consent, AI content always checked.
- The duties of the GPL arise on distribution.
- An LLM predicts likely tokens. Fluent does not mean correct.
- Legal advice is not part of a consultant's job.
Terms
- Anonymization
- Permanently removing the link between data and a person.
- Copyleft
- The duty to distribute derived works under the same licence.
- Data minimization
- Collecting only as much data as the purpose requires.
- Hallucination
- A plausible-sounding but false or invented statement by a language model.
- AI literacy
- The knowledge needed to use AI systems in an informed and responsible way.
- Pseudonymization
- Identifiers instead of names, the link can only be restored with extra knowledge.
- RAG
- The model answers on the basis of documents found beforehand.
- Training cut-off
- The date up to which data went into a model's training.