Focus of the letter 49

We have already talked here about the (non-)neutrality of models: each one carries and relays its own choices and values, visions and biases. This week, several readings show this, each in their own way.

Anthropic will prohibit gratuitous violence against its models in its usage policy. This raises questions about the fact that this practice exists, and especially that it is widespread enough for Anthropic to want to include it in its rules. We also read a certain conception of what the user can or should do with the machine.
A study relayed by Data for Good compares the responses of Claude, Gemini and ChatGPT to identical medical symptoms. The three models advise a young woman much less often than a man to go to the emergency room. The bias is not seen in an isolated response: only the comparison reveals it. And it has very real consequences, even more so when the advice is about health.
OpenAI publishes hundreds of mathematical results, only a part of which is verified by formal proofs (the rest may contain “errors”, the company warned): the true and the plausible coexist, and it is up to the reader to distinguish between them.
Arthur Mensch, of Mistral AI, does not see in the development of AI a risk of extinction of humanity, but in the concentration of the sector between two or three companies: in addition to the obvious interest for Mistral to defend its place, if only a few players provide the models, their vision of the world could become the majority or even ours by default.
That leaves Ecosia, the search engine that is leaving Mistral, considered late and not very sovereign, for open models, especially Chinese, hosted by a European platform. Its founder announces costs halved and experts warn of the risk of censorship. Whatever one thinks, this decision is a reminder that a model is chosen, and changed.

Price also comes into play, and it’s not as neutral as it seems. An OECD working paper, based on interviews with 25 organizations from 11 countries, describes a “tokenization tax”: models cut text into tokens, their vocabulary has been built mostly from English, which uses fewer tokens than others, and the same text is therefore more expensive to process in some languages.
The organizations surveyed outside the English-speaking world also describe that models are lagging behind in terms of law and cultural codes. The survey is qualitative and illustrative, but the idea is fraught with meaning: the price of the token is not the same for everyone, and language is part of the cost. Choosing a model therefore also means accepting, or not, to pay more to speak one’s own language.

This is where the subject becomes political. Opting for a model is not only about comparing its performance or its price. It is also about choosing the values, the words, the priorities and the silences that result from the decisions of its creators and publishers. By dint of using the same model out of habit, we risk mistaking its vision of the world for our own.
However, it is not a question of changing the model for each question, but of choosing the one that best suits the task and, when relevant, comparing several answers. Changing the model, or at least comparing several, then becomes more than a technical option: it is a way of keeping control, of preserving a diversity of points of view and our critical thinking.

To practice this, you can of course use chatbots that offer several models (there are many of them in the dedicated category of one AI per day). You can also test the answers in “Journey to the land of the unwritten” to question the criteria that have become invisible.
There is also a French public tool that has already been mentioned several times, Compar:IA, developed within the Ministry of Culture, which allows you to ask a question to two models and compare their answers. It aims to raise awareness of the pluralism of models, cultural and linguistic biases and their environmental impact, while making it possible to build a corpus that can be reused in future French-language tools.

What if, this week, we asked the same question to several models to compare their values and note what surprises us? Confronting their choices rather than integrating them little by little is exercising our critical thinking and questioning our own values. It is a question, individually and collectively, of not being subjected to these choices without questioning them.

Echo:

• Letter 2, ” Behind open models, a silent geopolitical battle “: open models and sovereignty issues, at the heart of Ecosia’s choice.
• Letter 37, ” Do we really know what we’re using? “: Tokemon and the 30 models, to identify the model behind each tool.
• Letter 17, ” Design Arena, playground for comparing AI models “: comparing models, what Compar:IA does.