# AI transparency: open models and the democratic question

**AI transparency democracy** links two debates that now overlap: who can inspect a model, and who benefits from it. Training a frontier language model requires large computing budgets and advanced machine learning skills, which few organisations possess. When the groups that can afford it keep the data and the code private and distribute the model only through an API, the research community loses the ability to examine how the system works.

## The open-source response

Open models appeared as a counterweight. Meta's LLaMA, released in early 2023 to researchers on request, was followed by derivatives such as Stanford's Alpaca and Databricks' Dolly, and later by Vicuna, Koala and ColossalChat. Some of these ran on a laptop, a Raspberry Pi or an old phone, which changed the assumption that useful language models require tens of thousands of dollars in hardware. The licensing remained restrictive for commercial use, and the terms of the larger proprietary models limited the development of competing systems.

The open-source argument is older than AI. The movement that produced Linux, Apache and Firefox showed that open development can generate widely used infrastructure. Figures such as Yann LeCun and Hugging Face's Clement Delangue argued that much of the progress in machine learning came from open science, and that closed models reduce transparency when researchers cannot even know how many parameters a system contains.

## Where the debate stands

The tension is genuine. Some labs argue that releasing powerful models openly creates safety risks, while others hold that independent audit is the only way to understand and improve them. Meta's decision to grant controlled access to LLaMA through academic researchers, rather than a full release, illustrates the middle ground. Joelle Pineau of Meta described transparency and accountability as essential while acknowledging that access levels should depend on the potential for harm, and independent researchers such as Alex Engler have noted how easy open models are to use even for people who do not understand them deeply.

## What companies should watch

For a business, the transparency debate is not abstract. It determines which models can be used in Europe, what documentation a vendor provides, and whether a system can be audited by a third party. The same public oversight logic now applies to the platforms a company depends on, through [EU algorithmic oversight](/en/ecorner/2023/eu-digital-algorithm-regulation.html) of recommendation and moderation systems. The EU AI Act and the NIST AI Risk Management Framework both push toward documentation, testing and traceability, which means that model provenance and auditability become procurement criteria. A vendor that cannot explain what data trained a model, how it is evaluated and where it runs fails those criteria regardless of benchmark scores.

## Frequently asked questions

**Why does open-source AI matter for transparency?**
It allows researchers and auditors to inspect weights, data and behaviour. Closed models can still be documented, and they limit independent verification.

**Are open models less safe?**
Safety depends on training, evaluation and deployment controls. Open weights make misuse easier to study and also easier to attempt, which is why controlled access models exist.

**What does the EU AI Act require on transparency?**
Different obligations apply by risk level, including documentation, logging and information for users. General-purpose models carry their own transparency duties.

**How should a company assess a model vendor?**
Ask for documentation on training data, evaluation, hosting location, security and audit rights, alongside the technical capabilities.

## Sources

Stanford CRFM — Alpaca: a strong, replicable instruction-following model — https://crfm.stanford.edu/2023/03/13/alpaca.html
Meta AI — LLaMA: open and efficient foundation language models — https://ai.facebook.com/blog/large-language-model-llama-meta-ai/
Hugging Face — open-source AI development — https://huggingface.co/
European Commission — EU AI Act — https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
NIST — AI Risk Management Framework — https://www.nist.gov/itl/ai-risk-management-framework