Open Source vs Open Weights: The Legal Battle Over the OSI AI Definition

The Semantic Crisis in AI Open Source

For over twenty-five years, the term "Open Source" possessed a precise, legally binding meaning governed by the Open Source Initiative (OSI). A program was open source if and only if its source code was publicly accessible, free to modify, and redistributable without discriminatory commercial restrictions (satisfying the 10 criteria of the Open Source Definition).

The generative AI boom shattered this consensus: - Leading technology companies released neural network model weights under licenses labeled as "Open Source," while imposing commercial user ceilings, field-of-use restrictions, or prohibiting competitors from fine-tuning rival models on generated outputs. - Most critically, these releases withheld the primary source code of modern AI: the pre-training dataset and scraping pipelines.

Releasing compiled floating-point tensor weights without the training dataset is analogous to distributing a compiled binary executable without the underlying C++ source code. You can execute the binary, but you cannot study, verify, or recreate how it was built.

In October 2024, at the All Things Open conference, the OSI officially ratified the Open Source AI Definition 1.0 (OSAID), establishing an international technical standard for what legitimately qualifies as Open Source AI.

┌────────────────────────────────────────────────────────────────────────┐ │ The OSI Open Source AI Definition (OSAID) │ │ To be certified as "Open Source AI", a system must provide: │ ├────────────────────────────────────────────────────────────────────────┤ │ 1. Complete Source Code │ │ - Training code, tokenizers, data filtering, inference engines │ │ 2. Complete Model Parameters & Architecture │ │ - Checkpoint weights, optimizer states, hyperparameters │ │ 3. Comprehensive Data Information │ │ - Provenance, selection criteria, data curation scripts, licenses │ └────────────────────────────────────────────────────────────────────────┘

---

The Four Degrees of AI Openness

The modern AI ecosystem does not exist in a binary of "closed vs open." Instead, models exist across a four-tier spectrum of accessibility and licensing freedom:

[Tier 1: Fully Closed] ──► [Tier 2: Open Weights / Restricted] (OpenAI GPT-4, Claude) (Meta Llama 3, Cohere Command) │ │ ▼ ▼ [Tier 3: Permissive Weights] ──► [Tier 4: OSI-Compliant Open Source AI] (DeepSeek-V3/R1, Mistral) (Allen AI OLMo, EleutherAI Pythia)

#### Tier 1: Proprietary Closed API - Models like OpenAI's GPT-4, Anthropic's Claude 3.5 Sonnet, and Google's Gemini Ultra. - Weights, training recipes, datasets, and architectures are proprietary trade secrets. Accessible solely via managed HTTP inference endpoints.

#### Tier 2: Open Weights with Restrictive / Commercial Licenses - Meta Llama 3 Family: Weights are freely downloadable, but the Llama 3 Community License imposes strict conditions: - Commercial threshold: Entities with over 700 million monthly active users must negotiate a separate commercial license with Meta. - Non-compete: Model outputs cannot be used to improve any other language model (except Llama derivatives). - While extraordinary for public research, the OSI explicitly rejected Llama's classification as Open Source due to these commercial and downstream restrictions.

#### Tier 3: Permissive Open Weights (Apache 2.0 / MIT) - Models like DeepSeek-R1, Mistral 7B (base), and Qwen 2.5 Coder. - Weights and code are released under standard permissive licenses (MIT or Apache 2.0). There are zero user ceilings or commercial royalties. However, raw training datasets are withheld.

#### Tier 4: True Open Source AI (OSAID 1.0 Compliant) - Architectures like OLMo (Allen Institute for AI), Amber, and Pythia. - Complete public disclosure: raw pre-training corpora (Dolma), exact data curation and deduplication pipelines, training logs, intermediate training checkpoints, and Apache 2.0 licensed code.

---

The Great Data Transparency Debate

The most contentious debate during the drafting of OSAID centered on training data requirements:

• The Strict Open Data Camp: Argued that without the exact, byte-for-byte training corpus, a model cannot be truly reproduced or audited for bias, security vulnerabilities, or copyright infringement.
• The Pragmatic Consensus Camp: Pointed out that requiring full public distribution of petabyte-scale datasets creates insurmountable legal and logistical barriers:
• - Distributing web scrapes containing personal identifying information (PII) violates the EU GDPR's "Right to be Forgotten."
• - Distributing medical imaging or clinical notes breaches HIPAA privacy laws.
• - Hosting 20+ terabytes of raw text data for free download imposes extreme bandwidth costs on non-profit research institutions.

The ratified OSAID 1.0 reached a pragmatic compromise: it requires "Data Information"—a detailed, reproducible accounting of data sources, filtering criteria, and access mechanisms—allowing researchers to understand and reconstruct datasets without legally requiring labs to redistribute copyrighted or sensitive personal data.

---

Licensing Comparison Matrix

| Model Family | License | Commercial Limits? | Outputs Allowed for Rival Training? | Training Data Disclosed? | OSI Compliant? | | :--- | :---: | :---: | :---: | :---: | :---: | | Allen AI OLMo | Apache 2.0 | None | Yes | Full Open Dataset (Dolma) | Yes (OSAID) | | DeepSeek-R1 | MIT | None | Yes | High-level report only | Permissive Weights | | Mistral Large 2 | MNLP-91 (Commercial) | Commercial fee required | No | No | No (Proprietary) | | Meta Llama 3.3 | Llama 3.3 License | >700M MAU limit | No | High-level report only | No (Restricted) | | OpenAI GPT-4o | Proprietary API | API terms | No | None | No (Closed) |

---

Enterprise Risk Analysis: Choosing the Right License

For engineering leaders building commercial software, the distinction between true open source and restricted open weights is not academic—it is a critical legal liability:

1. Audit & Compliance Risk: Enterprise legal teams routinely flag models with non-standard licenses (like Llama's commercial threshold) during M&A due diligence or public offerings. 2. Vendor Lock-in and Revocation: Permissive Apache 2.0 and MIT models (like DeepSeek or Mistral base) offer irrevocable legal protections. 3. Local Self-Hosting Sovereignty: Deploying models via local runners like Ollama and Hugging Face Spaces ensures that model inference, fine-tuned LoRA adapters, and proprietary customer data remain completely within corporate virtual private clouds (VPCs).

As national regulators under the EU AI Act begin enforcing transparency rules for general-purpose AI, the industry's shift toward transparent, standardized open-source definitions will dictate the next generation of enterprise software architecture.