Gemini has been doing something stupid in recent days, and here are the causes according to Google

Google’s chatbot arrives after its main competitors. An opportunity not to make the same mistakes as them, but which does not guarantee that things cannot go wrong.

If generative artificial intelligence does not only produce great works of art, it at least has the merit of being at the center of some incredible stories. This is the case of Gemini, the new nickname for Google Bard, whose image generation capabilities have been strongly criticized.

Indeed, the tool has the annoying habit of rewriting history when we don’t ask it to do so. A problem which is not trivial and due to… its safeguards.

Google’s good intentions

Taylor Swift’s recent troubles with deepfakes have reminded us that there is a need for tighter control over what can be done with a generative tool. A convincing image can cause a lot of harm, on a small scale like at the top of our states, and the moderation of tools like DALL·E or Midjourney is becoming a crucial concern for the companies that develop them.

Google is well aware of this, and has made Gemini public with the necessary safeguards to avoid the setbacks that others have experienced (or allowed). It is therefore impossible to create images of a violent or sexual nature with the program, or even to represent a real person, for example. A setting is as precise as it is essential for developers, because there are many smart people who try to find the flaw that will allow them to make an AI do anything.

But, in Mountain View, we want to do things well. In addition to putting in place safeguards, the company wanted to show the diversity of our societies within the images generated by Gemini. Thus, in credible contexts, and when not specified by the user, the characters drawn by the program can have any skin color or be of any gender.

Good ideas, but backfired on Google.

When moderation is zealous

In a blog post published this weekend, Prabhakar Raghavan, vice president of Google, explains why Gemini seems to have missed history lessons. “ First, our tuning to ensure Gemini shows a range of people didn’t account for cases that clearly shouldn’t show a range “, he explained. “ Second, over time the model became much more cautious than we expected and refused to answer some queries, misinterpreting very innocuous questions as sensitive. »

No question of rewriting history in Californian style, therefore. Here, AI tended to “ overcompensating in some cases and being too careful in others, resulting in clumsy and erroneous images », concluded Raghavan. Programs like Gemini are incredibly complex, and the American giant was fooled by its essential caution.

While waiting to resolve this issue, Google has disabled the ability to generate images with humans on Gemini. A decision as much criticized as applauded, but which will give the company a little more time to refine its safeguards.

Source : Neowin