Apertus: "Accessible" ≠ "Open"
A Technical Audit of the Swiss AI Initiative’s Data Model #
Executive Summary:Apertus markets itself as a “fully open” model built on “European Values.” However, its definition of “open data” relies ontechnical accessibility(robots.txt compliance) rather thanlegal permissiveness(copyright clearance). This creates a fundamental contradiction: a model claimed to be “open” and “transparent” is built on a corpus of copyrighted works whose specific origins are opaque, making true “studyability” and “redistribution” legally ambiguous.
1. The Distinction: Open Source Weights vs. Open Data #
Apertus correctly states that its architecture and model weights are open-source. However, it conflates this with the openness of its training corpus.
| Feature | Claimed Status | Technical Reality |
|---|---|---|
| Weights | ||
| Open Source | ✅ True. Licensed permissively. | |
| Code | ||
| Open Source | ✅ True. Training code is available. | |
| Training Data | ||
| “Publicly Available” | ⚠️ Misleading. “Publicly Available” means “visible to crawlers,” not “free of rights.” | |
| Copyright Status | ||
| Implicitly Clear | ❌ Opaque. Most web data is copyrighted. Apertus uses it without explicit license. |
Why “Publicly Available” is Not “Open”
In legal terms, Publicly Available means anyone can see it.
Open (in the context of intellectual property) usually implies anyone can use it (Public Domain or Creative Commons) or “free as in freedom”.
Apertus’s training corpus consists largely of copyrighted web text. By labeling this “open data,” they create a false equivalence. Users assume that because the data was “on the web,” it is free to use. It is not. It is scraped.
2. The “European Values” Contradiction #
Apertus positions itself as a sovereign, European alternative. This claim rests on two pillars: Data Sovereignty and Legal Compliance.
🛑 GDPR vs. Copyright: A Category Error
Apertus highlights its GDPR compliance extensively. This is true but incomplete.
GDPR protectsPersonal Data(identifiable human beings).** Copyrightprotects Creative Works**(articles, code, literature).
Apertus filters PII (Personally Identifiable Information) from its training data, which satisfies GDPR.
However, it does not filter Copyrighted Material. Most of the internet is copyrighted. By scraping it, Apertus is operating in a legal gray area common to all LLMs, but they frame this as “ethical standards.”
The Critique: To claim “European Values” while relying on a data set that largely ignores copyright norms (the default state of the web) is a selective reading of European IP law. The EU AI Act addresses risk, not copyright ownership. Apertus conflates the two.
🇨🇭 Swiss Copyright Law
Swiss copyright law protects literary and artistic works. It does not automatically grant permission for machine learning ingestion. By using “publicly available” data, Apertus is relying on an implicit, untested “fair use” or “text and data mining” exception argument, rather than explicit permission. This makes the “Open” label legally precarious.
3. The robots.txt #
Loophole: Technical vs. Legal
Apertus defines “openly available data” as:
“Data which is publicly available… filtered to respect machine-readable opt-out requests… even retroactively.”
The Technical Definition
“Open” = “Not Blocked by Robots.txt”
If a website does not explicitly block bots via robots.txt
, Apertus
considers the data “open.”
The Legal Flaw
This equates Technical Access with Legal Permission.
Default State: The default legal status of web content isCopyrighted, not Public Domain.** The “Opt-Out” Burden:**Apertus shifts the burden of permission to the creator. You mustactively blockthem to retain your rights. In Free Software terms, this is akin to saying software is “open” unless you write a license file.Retroactivity: Apertus claims to respect opt-outs “retroactively.” This is a technical feat of filtering, but it does not change the fact that theinitialingestion was based on a technical heuristic, not a legal license.
Why This Breaks “Traceability”
Apertus claims the model is “traceable.” But traceability requires a manifest.
- If I want to trace the source of a specific output, I need to know which URL generated it.
- Apertus provides “recipes” (datasets used), not a precise URL-to-Weight mapping for the entire web scrape. - Therefore, “traceability” is limited to dataset-level attribution, not content-level provenance. This makes it difficult for users to verify if a specific copyrighted work was included or how.
4. Request for Apertus v2: Transparency, Not Just Openness #
We do not ask for perfection. We ask for precision in their claims.
✅ What Apertus Should Publish in v2:
Data Composition Manifest:- A breakdown of the training corpus: % Public Domain, % CC-Licensed, % Copyrighted/Scraped.
- A list of the top 100 largest sources by token count, not just “the web.”
Clear Copyright Stance:- Acknowledge that “Openly Available” = Technically Accessible. - Acknowledge that the model is built on Copyrighted Data without explicit license, distinguishing it from models trained on fully licensed datasets (like Common Crawl with verified licenses).
- Acknowledge that “Openly Available” =
Source Code for the Filter:- Publish the exact code that parses
robots.txt
to prove the “retroactive” filtering logic is robust and not just a heuristic.
- Publish the exact code that parses
🚩 Conclusion
Apertus is a powerful, open-weight model built on accessible data. It is not a “free” model in the legal sense because the underlying data rights are unresolved. It is a scraped model wearing an “open” label.
Open Source Weights, Closed Source Truth.
Download Apertus Weights Technically open. Legally ambiguous.
If the download link doesn’t work, that is because we do not recommend non-free LLMs.