Many AI models called open source fail the test for the term

Many AI models marketed as open source do not meet the actual definition of the term. A How-To Geek opinion piece by Jorge A. Aguilar points to the Open Source Initiative, the group that keeps that definition. It reviewed Llama 2, Grok, Phi-2 and Mixtral, and none of them qualified.

What open source is supposed to mean

The term dates to February 1998, right after Netscape released the source code for its web browser. Bruce Perens then helped turn earlier guidelines into the Open Source Definition. It has ten rules. In short, people can share the software freely, see all the source code, build their own versions and use it however they want. No one can be turned away.

Where AI models fall short

Most AI companies release only the model's weights, which are the numbers it learned, plus some basic code to run it. The training data, the steps used to clean it and the setup choices stay private. The result is a black box that people can run but cannot fully check or rebuild.

Licenses add more limits. Meta's Llama Community License requires any company with more than 700 million monthly active users to get a separate commercial license. Other releases ban uses such as law enforcement or credit scoring. Some bar using a model's outputs to train competing systems.

Who passed the test

Only fully transparent projects made the cut with the Open Source Initiative. These were EleutherAI's Pythia, AI2's OLMo and LLM360's Amber.

The article also blames the wider community and platforms. Hugging Face used to group truly open models with restricted ones under the same open source tags and filters. The author argues that nobody pushed back, so companies borrowed the trust the label had earned.

Why full openness is hard

Requiring all training data to be released is not simple. Big datasets draw on huge parts of the internet and can include copyrighted material, which has already led to legal trouble. Models trained on medical or financial data can involve personal information protected by laws such as GDPR.

What to do

• Read the license before building a product on a free AI model.
• Look for limits on company size, industry or how outputs can be used.
• Treat the words open source on a announcement as a claim to check, not a guarantee.