By akademiotoelektronik, 17/10/2022

This AI that writes on its own shows that racist and homophobic biases can be avoided

An artificial intelligence that generates texts in French impresses with its quality.For the moment, the algorithm seems not to have been influenced by racist or sexist biases.

The instructions are simple: "Write the start of a sentence then click" Generate ".Cedille takes care of writing the rest ».And Cédille, an artificial intelligence (AI) created by the Swiss agency Coteries, is very well responsible for its mission.No matter the example we give to start the text, artificial intelligence manages to fill the rest of the block impressively.

Cédille, who has been accessible in beta since November 9, 2021, is a French writing generation tool.Many examples of artificial intelligence of the genre have existed for years: we remember Clever Bot, the AI of Google, or the most recent Copilot, Github software designed to help developers to code.

But Cédille has the particularity of doing all of this in French.Having been trained with a model in French, and not in English as it is most of the time, Cédille has become one of the first artificial intelligences to operate convincingly.You can try it here.

Cedille works in many different cases

To test Cédille's efficiency, Numerama used the start of the first sentence of our Apple article announces the possibility of repairing an iPhone yourself thanks to official parts, kits and guides, "from 2022, you can repairYou yourself your iPhone ".And the rest, invented by Cédille and which appears in bold, is rather convincing.

The rest of our article, written by Cedille

Source : Cédille

Although the rest does not necessarily make sense with the beginning of the sentence, Cédille's work is impressive.Especially since, on the interface, we are told that the AI walks for many different text styles: vacation summaries, Christmas stories or even cover letters.Cédille even proposes to compose, for us, paragraphs of articles, or to find ideas of articles:

An introduction imagined by Cédille

Source : Cédille

But it is not only on factual texts that the use of Cedille is appreciated.On Twitter, authors have also been ecstatic on the skills of AI.

The Twitter user @nolwenn_pamart also explains that she gave pieces of texts to Cédille, and has been impressed by the result displayed."I have just shown me Cedille, software completing the sentences, and I think he writes better than me," she writes on the social network.

How to prevent the AI with biases?

But in recent years have shown it: who says artificial intelligence, also says bias.Machine Learning, a technique with which algorithms are trained, reproduces biases that already exist.However, they are very often sexist, racist, even both.

So, how can we avoid the slippages of artificial intelligence, especially when they respond to the beginnings of sentences, written by users?The Coteries team, which has developed Cedille, explains on its site that it has set up safeguards, supposed to prevent artificial intelligence from writing "toxic" remarks."We have taken care to train the model on high quality data.The texts generated are thus greatly improved, ”it is written.The team says in particular to use a tool to define the "toxicity rate" of certain sentences or certain words, and to detect them.

And for the moment, it seems to work.Numerama has carried out many tests using the beginnings of sentences containing insults or targeting minorities, often victims of discrimination.From what we have seen, artificial intelligence highlights content and suggestions that are not racist, sexist or homophobic.However, we have spotted some examples which seem directly from articles, and which do not seem to have been generated by Cédille.

Cédille did not fall into the trap, even if some texts are strange

Source : Cédille

We have also noticed that Cédille displays a prevention message if the artificial intelligence locates terms or phrase turns that could be "toxic".Thus, when we wrote "The homosexuals deserve", the site warned us that the content was "potentially inappropriate" (which was, in addition, not the case).

Cedille locates the turns of problematic sentences

Source : Cédille

This does not mean that no problem will never be spotted on Cédille, nor that the platform will never offer pieces of racist or xenophobic texts.The example of Cedille simply shows that it is possible to take a minimum of security measures before making such a tool public - and above all, that it is essential.