Advocating for Truly Open AI Systems: A Call to Action

Open Source dominates the software market; almost no software is shipped without Open Source libraries or components, even though the final product might be obscured behind a commercial and proprietary license, stripping the user of rights to inspect the underlying components. Furthermore software is called open even as it doesn’t adhere to the Open Source criteria established by industry organisations. This practice, known as 'openwashing,' involves organisations misleadingly labeling their products as 'open,' implying transparency and freedom, without meeting established Open-Source standards. This mirrors "greenwashing," used when misleadingly presenting products as environmentally friendly.

While such openwashing in many areas of the software industry has been going on since many years it is particulary eminent in the space of so-called artificial intelligence (AI), which of course is based on neuronal networks and machine learning but will be used for this category of software as adopted in the broader market. Recently generative AI sytems got most of the hype as they impress us and without understanding how they are build they deliver an illusion of intelligence till they fail you.

Open Source has also proven to foster innovation, ensure fairness, or enhance security and we can easily mirror this to the field of AI.

It starts with the AI Model

AI models are classic pieces of software. Therefor they must be Open Source. This allows everybody to study, understand and create. Here, companies like Meta have committed to delivering Open Source models. However, OpenAI, despite its name, has not released source code beyond GPT-2.x—a situation now complicated by legal disputes involving Elon Musk. Additionally, Musk has initiated X.AI, aiming to develop 'Grok,' an open model. These developments highlight the varying degrees of openness in the AI field. Lets not call open what isn’t and lets welcome the open models as much as possible and require them. Even the European players like Mistral and Aleph Alpha have not delivered open models with their newest versions. Lets make sure we don’t use any non open models anymore, it is not needed.

Continues with all the parameters

Even when the AI model source code is available this is mostly not true for the parameters used while training, or generating answers to our inputs. For transparency, security checking and also trust in its potential bias and its weights all paramters including weights, bias, temperature, pre-prompt and anything used in building, training or generating output must be open. Where Open Source licenses don’t fit, data or content licenses can be used.

Last but not least the training data

All training data sets must be availabe and must be covered by Open Source, Open Data or Open Content licenses. Like with Open Source no use restrictions can apply. Even in some scientific models there is not completel transperency on the training data used, or the license under which such data can be reused or the dataset has use restrictions which might be in the best idea of somebody but it is still discrimaniting.

Piecing things together for AI Systems

When the above are combined, when models are combined with each other, when they are combined with additional vector databases, retrieval engines and whatnot only Open Source software shall be used and parameters must be published.

A truly open AI System adheres to all the needs for transparency and openness for our full right to study, understand and reuse. An open AI System has no requirement to be ethical or responsible, those are additional criteria for AI Systems in use and make them fit for specific purposes.

Let’s work on truly Open AI Systems, the final wording by the OSI – the Open Source Initative, the steward of the Open Source definition since 1998, will give us a well defined basis but already with the simpler criteria above we can start our influence today and make sure that in the spirit of Public Money – Public Code the AI Systems and Models pushed forward by a diverse set of government organizations are nothing else but truly Open!

Version 1.0 - May 9th 2024