DISCLAIMER: The views and opinions expressed in this blog are solely my own and do not reflect those of my employer, Fraunhofer or any of its affiliates. All content is based on my personal insights, experiences, and research, and should not be construed as representing official positions or policies of Fraunhofer.

TLDR; In the EU, anonymize personal data unless necessary to reduce bias, and ensure compliance with data minimization principles. For copyrighted material, you can train models unless owners explicitly opt out using appropriate machine-readable tags. Outputs generated by AI systems are public domain unless human creativity is involved, which allows copyright claims. The rules remain broad, creating ambiguity across member states.

Intro

After writing my previous post on the EU Data strategy, I was left with a sense of dissatisfaction. Working in data myself, I failed to see clear and defined rules on how we’re supposed to collect/process data.

For this reason, I took some very real problems the sector (training LLMs but it can be extended to other modalities too) is facing and transformed them into questions I aimed to answer:

1. How can you handle Personal Identification Information (PII) in your data?
   • Should you anonymize data?
   • What is the best way to access data?
2. What about licensed and copyrighted material?
   • Can you train your model on copyrighted material?
   • What about intellectual property?

Documents and Conventions

Before starting, I should clarify which documents are which (mostly for myself, I need this). Since some documents come with an official name (e.g., Regulation 2018/679) and some with a common name (the latter is the GDPR), here’s a list:

• Regulation (EU) 2016/679 : General Data Protection Regulation, aka GDPR
• Regulation (EU) 2018/1725 : New Data Protection Framework, aka NDPF1725
• Directive (EU) 2016/680 : Law Enforcement Directive, aka LED680
• Directive (EU) 2019/790: Directive on Copyright in the Digital Single Market: DC790
• Directive - 96/9/EC: Database Directive, aka DD96
• Directive 2001/29/EC: Copyright Harmonization, aka CH29

For citation, I’ll use a simplified notation: Article X.Y.Z, where X is the article number, Y is the paragraph, and Z is the point¹. For the introductory sections, I’ll use Recital X.Y.Z. This might be obvious for those with a legal background, but for my fellow tech people - this one’s for you! <3

With this said, los geht’s!

Personal Identification Information

Let’s start by looking at how the AI Act handles personal data. Searching for “personal data” in the Act reveals several key points:

• Recitals 10 : GDPR’s fundamental rights to data protection still apply. No exceptions for AI training.
• Recital 67²: You must be transparent about how and why you originally collected the data.
• Recital 69 ³: Emphasizes data minimization and protection principles. You should anonymize and encrypt data where possible, or better yet, use federated approaches where algorithms go to the data instead of data moving around.
• Recital 140 : Discusses data sandboxes⁴ and how personal data collected for other purposes can be used in AI systems only under specific conditions⁵.
• Article 10: Focuses on data governance, with a range of requirements if you’re working with personal data⁶.
• Article 59: Details rules for using personal data in data sandboxes. Essentially, in these sandboxes, you may process “lawfully” gathered personal data, but only if working in sectors like healthcare, environment, transport, or public sector efficiency. Additionally, data must be deleted after use.

To summarize, anonymizing or pseudonymizing personal data is essential. In rare cases where full anonymization isn’t feasible, you must demonstrate that processing the personal data is necessary to mitigate biases.

Data Features

Recital 69 introduces the principle of “data minimization and protection”, but what does that actually mean? According to Article 66 of the AI Act, the data used for AI also needs to be of “high quality”—a somewhat vague term. In essence, data used in AI should meet certain standards and adhere to specific principles.

Principle of Data Minimization

The EU data protection supervisor defines the principle of data minimisation to be “data controllers should collect only the personal data they really need, and should keep it only for as long as they need it.”. Working in data myself, the question of “how much data is enough” often pops out. In the context of training large language models, it sometimes feels as though there’s no such thing as “enough” data, suggesting that data minimization could be overlooked if LLMs demand more data from various sources. Furthermore, Article 10.5 of the AI Act allows the collection of personal data specifically to reduce biases, potentially creating a loophole where vast amounts of personal data could be collected without anonymization, all under the aim of mitigating bias in LLM training⁷.

High-Quality Data

Recently, my team and I published an article discussing our data processing methods for one of our projects. In it, we touch on a field-wide issue: there’s no standard definition for “high-quality” data, and different techniques are emerging to assess various properties of text. So, I was surprised to see “high-quality data” mentioned in the AI Act. Reading through, I noticed the definitions are still pretty broad and seem to revolve around a few main concepts:

• Biases (Recital 67): We’ve talked about this before. The AI Act also acknowledges the risk of feedback loops where biased data leads to biased models, which in turn produce even more biased data, creating a self-sustaining loop.
• Relevance and Representativeness (Article 10.3): High-quality data should be relevant to the AI system’s intended purpose and sufficiently representative of the real-world scenarios the system is expected to operate in. You could argue that if you’re trying to build artificial general intelligence, all data is relevant.
• Error-Free and Complete (Article 10.3): I have no idea what this means from a technical perspective. What counts as an error in a text? A typo? A syntactic issue? A conspiracy theory (non-factual content)? Thankfully, the Act adds “to the best extent possible” before this requirement (Article 10.2.h).
• Appropriate Statistical Properties (Article 10.3): Another vague requirement. However, if you can define certain statistical properties as “quality signals” (like factual correctness, average word length, or toxic content), you could start measuring these to give concrete values for each of these signals.

Back to PII

This section was about PII—did you forget? I definitely did and had to reread from the start.

So, did we manage to find an answer on handling PII in your data? Surprisingly, yes.

The golden rule: always anonymize your data! Removing PII from text is a growing field of research⁸, and there are some established methods out there. But if you really want to play at pro level, consider never even touching the data directly. Instead, go for federated approach (wink wink to GaiaX).

Finally, if none of this appeals and you’re doing nothing to address PII, just say you need it all to mitigate bias! Sure, recent research might call this out as ineffective, but you can try to spin it using scaling law.

Copyright and Intellectual Property

Now let’s dig into copyright and intellectual property (IP).

What do we actually care about here? For AI systems, I’d say it boils down to two main questions:

1. If I train a system on copyrighted material (say, with non-commercial clauses), can I use the system for commercial purposes?
2. And what about intellectual property? If my model spits out something close to the input data, can I get sued? If the output is modified, when does it stop being IP from the original source?

Train on Copyrighted Material

So, let’s check in the AI Act. First up, if you’re conducting research, you’re in the clear and you can train on pretty much anything (Recital 25). But if you’re working on something for commercial purposes, things get trickier.

Copyright vs. Term of Services vs. License

Before we start discussing training on copyrighted data, let’s clarify what copyright really means and how it relates to other terms. If you’re already familiar with these concepts, feel free to skip ahead.

In simple terms, copyright protects the ownership of the original work. For example, if you scroll down to the bottom of this post, you’ll see “© Copyright 2024, All right Reserved By - Nicolo’ Brandizzi”, which means I own this post. Now, say you want to use this blog post to create a podcast episode. Can you read parts of my blog post for your podcast? That’s where the license comes in. A license regulated how copyrighted work can be used by others. If I license this post for “non-commercial use only,” you wouldn’t be able to monetize it on your podcast. Finally, since this post is hosted on my website, which is also hosted on GitHub pages, we’re both subject to GitHub’s Terms of Service (ToS). GitHub’s ToS states that users aren’t allowed to distribute my content outside of GitHub (e.g., mirroring my post on other sites). So, the ToS basically defines the rules we all have to follow to use a platform or service.

Custom Licenses

Here’s the thing: there’s no fixed set of licenses, or ways to interact with copyrighted content. You’ve probably heard of licenses like Creative Common, MIT license and GNU General Public License. These are well-established, but you can always create your own!

For instance, I could create a unique license for this blog post where I allow AI training only on paragraphs that start with a consonant. Unusual? Definitely. But still legally binding.

Why don’t we see more custom licenses? Probably due to several factors: legal expertise is usually required, and custom licenses are harder to interpret and enforce, which can lead to legal headaches.

Opting Out of AI Training

If you’re training a model for commercial use, Recital 105 makes it clear: you need to check if the data owner has explicitly said, “Hey, don’t use this for AI”. If there’s no specific mention or opt-out, you’re technically in the clear to use the data.

So, how do data owners actually opt out of having their content fed into an AI? Well, the Act doesn’t lay it out in exact terms but points⁹ us to Article 4.3 of DC790, which says that rights holders need to express their wishes in an “[…] appropriate manner, such as machine-readable means in the case of content made publicly available online.”.

In plain terms, if the opt-out isn’t “appropriate” (like, say, yelling “NO AI” from your window), your data might still be used for AI training. But let’s clarify something important: if the data isn’t labeled properly with a “do not train” tag (like a machine-readable indicator), it doesn’t mean copyright protection vanishes. It just means that, unless explicitly marked, you can use it for training unless the owner opts out. Copyright remains, but these guidelines let you use the data if no opt-out is provided.

An Example of Appropriate Opting Out

So, what does this “appropriate manner” actually look like?

Take DeviantArt, for example (a popular digital art platform you may know for other reasons too 👀). f we check out their terms of services, point 24 states: “Unless you actively give your consent, for Artificial Intelligence Purposes, DeviantArt will include a robots meta tag with the “noai” or “noimageai” directive in the head section of the HTML page associated with that Content on the Site […]”

This is a fantastic, straightforward way to set your intent individually. Now, if we could just agree on standard tags to cover different data types (e.g., “notextai,” “noaudioai,” etc.), we’d be all set.

What About the EU?

DeviantArt’s approach is commendable, but it’s not exactly a universal standard yet. So, does the EU offer a standard for opting out? Kind of.

The DC790 introduces the concept of Text and Data Mining (TDM), referring to automated methods of analyzing vast amounts of text and data. TDM’s copyright implications are complex, mainly because it’s wrapped up in copyright exemptions. For instance, Article 3 of DC790 grants certain exemptions to copyright ¹⁰ if entities conduct research¹¹.

For everyone else, data use is allowed if the data doesn’t explicitly prohibit TDM and if you’ve accessed it lawfully (i.e., no hacking). In that case, Article 4 of DC790 exempts you too.

So you might think you’re covered by opting out of TDM and keeping your data safe, right? Well, yes and no. Remember how I mentioned that there’s no standardized opt-out? This is a well documented problem and hasn’t been solved yet. As it stands, there’s no universally recognized method for opting out.

Whose Job Is This?

In recent years, we’ve seen a flood of datasets released each year¹² on various platforms like HuggingFace, Zenodo, Kaggle, and so on.

So, who’s responsible for collecting the “right” data? Is it the dataset provider (i.e., that one struggling PhD student) or the people who end up using the data? Well, according to Recital 106, it’s the latter. The AI providers bear the responsibility of identifying the copyrights associated with the data they use.

I’m genuinely conflicted about this. On one hand, it makes sense to hold the people actually training the models accountable, given they likely have the resources to investigate copyright. But on the other hand, datasets often come stripped down, with only the core content (like text from blog posts or images from DeviantArt) leaving out the original HTML or metadata where copyright info might have been stored. So, what’s the solution in these cases?

Honestly, I think we should obligate or at least heavily incentivize dataset collectors (yes, our overwhelmed PhD students again) to include copyright information for each data point. This small step could go a long way.

Summary on Copyrighted Material

• Check before you train: If you’re working on a commercial AI, it’s on you to make sure the data is free to use. Look for opt-out tags or warnings.
• Opting out is possible: Data owners can make it clear their data is off-limits with metadata, terms, contracts, or even public statements.
• Stay responsible: The EU wants a balance between innovation and protecting rights, so don’t think you can ignore these rules and use everything freely.

Derivate works

Alright, let’s dive into the second question (the one about intellectual property, hope you didn’t forget already).

This one is hard. Why? Because it’s basically philosophy.

Take my blog post here, for instance. If I license it with a “non-commercial use” clause, that means you can’t turn it into a podcast for profit. But what happens if I add a no derivate clause? What does “derivative” even mean?

The Blurry Line of Similarity

When it comes to defining what’s derivative, things get murky fast. Copyright and trademark laws don’t follow one-size-fits-all rules, it depends on context, interpretation, and, often, a judge’s decision.

Take the Creative Commons non-derivative license. It says, “If you remix, transform, or build upon the material, you may not distribute the modified material”. Terms like “remix,” “transform,” and “build upon” are vague, and they link to footnotes that dig into what counts as an adaptation. In short “modification rises to the level of an adaptation under copyright law when the modified work is based on the prior work but manifests sufficient new creativity to be copyrightable, such as a translation of a novel from one language to another, or the creation of a screenplay based on a novel.”

This “sufficient new creativity” is a fuzzy area, and only a court can really decide what qualifies. I found a few landmark cases that talk about this gray zone:

• DC Comics vs. Gotham Garage (2015): DC Comics sued a car manufacturer for creating replicas of the Batmobile. Although the Batmobile is “just a car”, the court ruled it was a protected character due to its unique design and connection to Batman’s world. This shows that even vehicles can be characters under copyright law when they are distinctive enough.

• Universal vs. Nintendo (1982): Universal claimed Nintendo’s Donkey Kong was too similar to King Kong. However, the court sided with Nintendo, ruling that the broad idea of a giant ape was public domain, and Donkey Kong was distinct in its own way. Here, the court acknowledged similar themes, but differences in characterization and execution allowed Nintendo to win.

• Mattel vs. MGA Entertainment (2011): Mattel, the maker of Barbie, sued MGA, claiming the Bratz dolls were derivative of Barbie. After years of litigation, the court ruled in favor of MGA, noting that while the two products shared similarities as fashion dolls, the distinctive style and design of Bratz dolls made them original enough. The case highlighted how styling and branding can create separation, even within similar categories of products.

Each of these cases shows that there’s no universal rule for “too similar.” Courts look at key characteristics, overall design, and whether the new work borrows too much from the original or brings something fresh.

So, what does this mean for our AI models? The answer, honestly, is that no one knows yet¹³.

Who owns the output?

So, imagine you’re using an AI system to clean up the readability of your blog posts. The system takes your messy, grammatically broken draft and spits out a polished, readable piece instead. Who owns this newly polished version? Let’s run through the options:

1. The company providing the AI service
2. The AI system itself
3. You
4. It depends
5. No one

If you guessed “it depends”, you’re on the right track!

Copyrighting AI-Generated Outputs

Out of the 5 options I enumerated before, let’s eliminate option 2 right away. No, the AI system itself can’t claim ownership law, copyright is only granted to natural and legal persons¹⁴, meaning humans or organizations. AI isn’t considered either of these (yet), so it’s out.

What about option 1, the company providing the service? If they owned the output, users would likely abandon AI systems altogether, since it would mean you couldn’t use any AI-generated work commercially. That’s why most AI companies state in their terms of service that the output belongs to the user¹⁵. Besides copyright, you need human creativity, something the company didn’t exactly contribute to.

That leaves us with two possible owners: you or no one, and two levels of human involvement: some, or none. It’s straightforward: if the AI system alone generated the content with zero human creativity, it falls into the public domain. Nobody owns it.

On the other hand, if you co-created with the AI, adding your own creativity, then you meet the human creativity requirement and can claim copyright. But how much involvement is enough? Probably more than just pushing a button and sitting back. For a deep dive into this, check out this paper.

Summary on Derivate Works

So, what does all this mean for AI? If your AI system is just reproducing the exact same input data, prepare to get sued for copyright infringement. For everything else, the jury’s still out, but one thing is clear: to copyright AI-generated content, you need to contribute some genuine human creativity.

Conclusion

So, we’ve made it to the end. First off, thank you for sticking around this long! (And if you just jumped to this section from the intro, no hard feelings. <3)

Key Take-Aways

To recap, here’s what we’ve covered:

• PII Handling: Always anonymize or pseudonymize data, and only use personal data when necessary to mitigate biases. Following data minimization principles is crucial, though federated approaches can further limit direct data access.
• Data Quality: The AI Act emphasizes high-quality, error-free, relevant, and representative data, but “high quality” lacks a strict definition, leaving much up to interpretation.
• Copyrighted Material: When using copyrighted data, check for opt-out tags or machine-readable permissions, as unmarked data could be used by default. However, this doesn’t nullify copyright protections; it simply allows use unless the owner explicitly opts out.
• Derivative Works: The concept of “derivative” remains legally ambiguous, with AI outputs falling into a gray area regarding originality and transformation.
• AI-Generated Outputs: Ownership of AI-generated content depends on human involvement. Minimal involvement, like pressing a button, isn’t sufficient for copyright, as outputs must involve meaningful human creativity.

Also, fun fact: I used ChatGPT to help summarize this list. Does that make it public domain? Well, that’s still a gray area…

Personal Take-Aways

As some of you know, my background is pretty technical, and coming from engineering, I’ve learned there’s always an answer (a yes or a no, a number that defines a minimum threshold, or something equally straightforward).

This? This is the complete opposite. Regulation is a different world where you’re defining acceptable behavior while considering so many factors that listing them all is already a task. After all that effort, the regulation you create can still end up used in ways you’d never imagine, sometimes for the worse¹⁶.

So, why am I saying this? Maybe I’m just trying to convince myself that the EU has a monumental task in balancing regulation with freedom. Unlike the U.S., the EU isn’t a single state but a collection of member states, each with its own priorities. If you pass a law that’s too restrictive, a member state might veto it. So, regulations have to be broad enough to avoid pushback while still trying to meet their original goals.

But that’s also why there’s a lack of specifics in EU policy. And while it makes sense, it doesn’t help. We seriously need standards (like a consistent way to opt out of AI training) and precise laws. This ambiguity hinders innovation. After all, a law might be interpreted one way in Spain and completely differently in Poland¹⁷. And no one wants to risk a lawsuit over unclear laws.

It’s ironic. Legislating is about weighing endless variables and unpredictable outcomes. You know what’s good at that? AI. Yes, this is a place where AI could actually help. In an ideal future, I’d love to see AI draft laws, with humans having the final say. Humans would always hold the power, but it’s becoming clear that we’re not always great at weighing long-term consequences. So, why not let something more capable handle that?

Footnotes