Between hallucinations and historical distortions, ChatGPT and Gemini go off the rails 😵‍💫

Cliquez ici pour lire en français

Generative artificial intelligences are being used more and more, but can we trust them? Recent events show that we can’t.

ChatGPT goes crazy 😵‍💫

On February 20, ChatGPT literally went crazy. Many of the tool’s users witnessed strange behavior on the part of OpenAI’s generative AI. Several accounts of ChatGPT’s replies were posted on the social network Reddit.

« The common experience over the last few hours seems to be that responses begin coherently, like normal, then devolve into nonsense, then sometimes Shakespearean nonsense »

For several hours, users observed inconsistent responses from the AI. Between mixing languages and generating content that had nothing to do with the user’s command, ChatGPT was really messing things up.

Following the avalanche of testimonials, OpenAI conducted an investigation and delivered its conclusions.

On February 20, 2024, an optimization to the user experience introduced a bug with how the model processes language. LLMs generate responses by randomly sampling words based in part on probabilities. Their “language” consists of numbers that map to tokens. In this case, the bug was in the step where the model chooses these numbers.

Gemini doesn’t seem to know much about history 🤨

Google’s Chatbot, Gemini, which is supposed to compete with ChatGPT, has been one screw-up after another. Previously named Bard, it had already been the subject of a bad buzz when it was launched (Click here to find out why). But the latest is even worse.

Since the beginning of February, it has been possible for Gemini’s US users to generate images. While this is an important milestone for Google’s chatbot, not everything went according to plan. Indeed, some users got a nasty surprise when trying to generate historical figures. For example, when typing « German soldiers in 1943 », some Americans were unpleasantly surprised to see black Nazi soldiers appear. Except that this is simply impossible given the historical context.

© Google. Capture d’écran. Images générées par Gemini – Requête : soldats allemands en 1943.

But that’s not all. Gemini also tends to generate black Vikings, or women among America’s founding fathers. That’s all it took to set the American web on fire. Since the incident, image generation has simply been deactivated. Google has commented on the matter.

We’re working to improve these kinds of depictions immediately. Gemini’s AI image generation does generate a wide range of people. And that’s generally a good thing because people around the world use it. But it’s missing the mark here. »

Beware of AI ⚠️

While unfortunate, these generative AI blunders at least have the merit of drawing our attention to the imperfections of this technology. In the case of ChatGPT, an update drove the chatbot mad. In the case of Gemini, it’s the very basis of the AI that’s not right.

As explained in a previous Question Aux Griots (available by clicking here), generative AIs are not intelligent. They are only complex systems and algorithms trained and configured from astronomical amounts of data. This data, collected from the web, is full of stereotypes and biases. And these are very often reflected in the results generated by these AIs. In the case of Google and Gemini, we can only imagine that the Mountain View company wanted to configure its AI in such a way as to reduce stereotypes towards certain people and communities. Unfortunately, the opposite happened.

Whether it’s ChatGPT or Gemini, we have yet more evidence that these tools are far from perfect. Their errors, known as « hallucinations », should be a reminder to us all that AIs are not omniscient entities, but only imperfect machines.