Open Source AI: What It Means for Client Data Privacy

What Open Source AI Means

Open-source AI includes the freedoms and materials needed to study and change the system. The Open Source Initiative’s Open Source AI Definition requires model parameters, relevant code and sufficiently detailed information about training data.

That data requirement does not mean every original training document must be downloadable. It includes information about data sources and preparation that lets a skilled person build a substantially equivalent system.

An open-weight model makes its learned numerical settings available. Your firm can load those weights into compatible software and generate answers on its own hardware. A release can provide weights without the training code or data information needed to meet the wider definition. Its licence can also restrict use or redistribution.

“Open source” therefore needs a more precise question: what exactly has the maker released, and under which licence? Access to weights helps your IT team control deployment. It does not by itself make the training process reproducible or the licence unrestricted.

Is ChatGPT open source?

No, ChatGPT is a hosted service, and its underlying models are not released as a downloadable ChatGPT installation. OpenAI’s separate open-weight releases include gpt-oss-20b and gpt-oss-120b. They are models your team can deploy through compatible software, not a local copy of the ChatGPT product.

Official OpenAI documentation describes gpt-oss-20b as a model for local or specialised use. Both named gpt-oss releases use Apache 2.0 licensing. Choosing one still leaves your firm responsible for its hosting and the application around it.

Leading Open Models

Qwen, Gemma, Llama, Mistral, DeepSeek and gpt-oss are model families legal teams will encounter when assessing downloadable AI. Their licences differ, including between releases from the same maker. The examples below identify specific releases so a family name does not become a blanket licence claim.

Model Families, Licences and Document Tasks

Apache 2.0 and MIT permit commercial use subject to their terms. Gemma and Llama use their own agreements, so the same assumption cannot cover every row. A modified or distilled model can carry conditions from its source model as well as its publisher’s terms. Distillation trains a smaller model using outputs from another model.

Which models suit legal document work?

For multilingual text, Qwen3 is a useful starting point because the family supports multiple languages and reasoning modes. Llama 3.2’s small text variants suit short rewrites and summaries where hardware is limited. For documents containing images, choose a vision-capable Gemma variant rather than a text-only model.

Mistral Small 4 and gpt-oss are options for teams with more processing capacity and text-reasoning tasks. A larger model also increases the work of hosting it. The best choice for a firm depends on document type, available memory and acceptable processing time.

Document work also needs more than a model that accepts text. A scanned PDF needs text extraction or image processing. A long bundle needs a way to retrieve relevant passages or divide the work. A claimed context window describes how much input a model can handle, not proof that it will find every relevant fact.

Choose the model and document workflow together. An image-capable model does not guarantee correct reading of a faint amount, and a fluent summary can omit a qualification. The verification process must check the document facts your team relies on.

Open Does Not Mean Private

Privacy depends on where your model runs and which other systems receive the input. The same downloadable weights can be used in several places with different access arrangements.

A cloud environment your firm controls still has a hosting provider. Control needs to cover access permissions and the application’s data paths, including logging and backups. An Australian hosting region alone does not establish who can access the content.

DeepSeek shows the difference. Downloaded DeepSeek weights can process documents on an office computer without sending those prompts to DeepSeek, provided the surrounding software also stays local. DeepSeek’s hosted app sends requests to its service. Its privacy policy says it collects, processes and stores personal data in China.

These are separate deployment choices even when both carry the DeepSeek name. The DeepSeek guide covers the hosted app’s privacy position. Using an open-weight release locally does not inherit the hosted app’s data path automatically.

For the practical computer setup, use the local LLM checks. If your team wants a managed hosted service, assess private AI through its processing terms and access controls.

Judging an Open Model for Client Work

Approve an open model for a defined task and deployment, with a named person responsible for maintaining it. A licence review answers permission to use the model. A separate privacy review answers who receives the material.

1. Identify the exact model and licence. Record the publisher, release and any underlying model. Confirm commercial-use rights, acceptable-use restrictions and redistribution conditions for your intended workflow.
2. Draw the data path. Name where prompts and documents are processed, where answers are saved and who can access each copy. Include document extraction, external tools and telemetry.
3. Assign maintenance. Name who updates the model runner, checks downloaded files and responds to security notices. Keep an approved version and a way to return to it after a failed update.
4. Define output checks for the job. For extraction, compare names and figures with source pages. For summaries, check material qualifications and omissions. Verify every cited authority against the original authority.
5. Test the complete workflow with fictional or approved sample documents before using client material. Include the formats and document lengths your team actually handles. Record failures as well as successful outputs.

For a hypothetical statement summary, a useful check asks whether a transfer is described as income without evidence. Compare the classification with the source and surrounding transactions. That tests the firm’s intended use more directly than whether the model writes a convincing paragraph.

Keep approval tied to the model release, its configuration and its deployment. If a local workflow later adds web search or a hosted document extractor, reassess that new data path before sending client material through it.