Does AI “Learn” or “Copy”? A Legal Analysis of Learning vs Memorization in AI Training

December 08, 2025

1. Introduction: The Most Common Argument from AI Developers

Whenever AI is accused of copyright infringement, developers argue:

“The model doesn’t store the original images.”
“AI learns patterns — it doesn’t copy.”
“The system only generalizes data.”

However, experts in machine learning and copyright law consistently respond:

AI does not merely learn — it also copies, transforms, and memorizes data.

To understand whether AI infringes copyright, we must analyze:

✔ how AI systems actually work

✔ how the law defines copying

✔ whether “learning” is meaningfully different from “reproduction”

Spoiler: Legally, it often isn’t.

2. What Does “Learning” Mean in Machine Learning?

AI “learning” involves:

Reading and loading the entire dataset
Copying the data into memory
Tokenizing or converting it into numerical form
Extracting patterns
Encoding those patterns into model parameters

In other words:

**Learning begins with copying.

There is no learning without duplication of the data.**

Even if the final stored representation is numeric, the initial copying is legally relevant.

3. What Is “Memorization” in AI?

Memorization happens when a model:

stores patterns explicitly or implicitly
reconstructs specific elements of the training data
generates outputs resembling original works

Contrary to developer claims, research shows:

✔ AI can reproduce training images verbatim

✔ AI can regenerate text passages word-for-word

✔ AI can replicate watermarked images

✔ AI can regenerate faces from training photos

✔ AI can leak training data indirectly

This is known as inadvertent memorization or overfitting,
but even well-trained models retain memorized fragments.

Thus:

AI absolutely memorizes — not only learns.

4. From a Copyright Perspective: Learning = Copying + Transformation

whether AI “understands” the data
whether AI “retains” the original file
whether AI stores JPEGs explicitly

The law only asks:

✔ Was the copyrighted work reproduced at any point?

✔ Was the work used to create a derivative product?

✔ Did the system copy, even temporarily, the protected material?

Under many legal systems:

Temporary copying = reproduction.

Transformative encoding = reproduction.

Every step of AI training involves reading, duplicating, and processing copyrighted works.

5. Why the Developer Argument “AI Does Not Copy” Is Legally Incorrect

Developers often use the analogy:

“Humans also learn from what they see — we don’t call that copying.”

But this analogy is flawed because:

❌ Humans do not replicate images pixel-for-pixel

❌ Humans cannot reproduce a copyrighted photo identically

❌ Humans have creativity and free will

❌ Humans do not store perfect compressed representations

❌ Human learning is subjective, not mechanical

AI, in contrast:

✔ copies data precisely

✔ processes it algorithmically

✔ stores compressed representations

✔ can regenerate training data

✔ cannot differentiate legal vs illegal content

AI is closer to:

a photocopier + compressor + pattern generator

than to a human brain.

6. Technical Evidence: AI Does In Fact Memorize Training Data

Studies from:

Stanford
MIT
Google DeepMind
OpenAI

show that models:

✔ emit verbatim training data

✔ reproduce copyrighted text

✔ regenerate images with watermarks

✔ recreate identifiable characters or faces

✔ replicate stylistic and structural features

This demonstrates:

AI is capable of harmful and unlawful memorization.

7. Legal Perspectives from Around the World

United States

Courts and scholars increasingly argue:

“AI training involves making unauthorized copies of copyrighted works.”

In Getty Images v. Stability AI,
the reproduction of watermarks proved that the model copied images directly.

European Union

Under EU copyright law:

text and data mining involves reproduction
commercial use requires explicit licensing
AI training is treated as systematic copying

The EU AI Act reinforces this by requiring dataset transparency.

Indonesia

Under UU Hak Cipta:

reproduction includes direct copying, indirect copying, and transformations
training qualifies as reproduction
reproduction without permission = infringement

Thus:

AI training = reproduction under Indonesian law.

8. So, Does AI “Learn” or “Copy”?

✔ Technically → AI learns by copying

✔ Legally → learning involves reproduction

✔ Practically → AI memorizes significant data

✔ Ethically → AI uses creators’ work without consent

Therefore, the most accurate description is:

AI learns by copying, transforming, and embedding copyrighted works.

This is not exempt from copyright law.

9. Conclusion

❌ AI does not “just learn”

✔ AI copies, stores, and reconstructs information

❌ “It doesn’t save the original file” is not a legal defense

✔ Any copying — including temporary — is legally reproduction

❌ “AI is like the human brain” is a false analogy

✔ AI is an automated copying and transformation machine

❌ Developers cannot avoid liability

✔ Training requires permission from rights holders

In summary:

**AI learning = copying with mathematical transformation.

It is still copying under copyright law.**