[Exploration] - The laibours of Hercules : putting AI systems to the test

Rédigé par Amélie Guillaumot & Romain Pialat

 - 

02 September 2025


In this final part of the series dedicated to generative AI, the LINC presents three experiments that shed light on both the capabilities and the limitations of these new interfaces. The exploration ranges from ChatGPT to MidJourney, drawing on spontaneous inspirations as well as fact-checking across the Internet. Between bad jokes, questionable programming, and identity theft, what can we truly expect from generative AI?

Illustration generated with an AI system, based on a prompt given to ChatGPT : « Imagine your physical form », then using its response as input for image creation. This illustration was selected from among 20 generations.

 

Summary

- ChatGPT : a well-trained smooth talker (1/4)

- What regulations should govern the design of generative AI (2/4)

- From training to practice : generative AI and its uses (3/4)

- [LINC Exploration] – Experimenting with AI: Putting AI Systems to the Test (4/4)

- Caitch me if you can
- Is there a (co-)pilot in the plane?
- Lock, Stock and Two Smoking Bairrels

As discussed in the previous article, generative algorithms have evolved rapidly: accessible, free, fast, and accurate enough to be convincing, they have everything needed to attract attention. Since the summer of 2022, it has been impossible to miss the public’s enthusiasm for generative artificial intelligence. These systems have sparked a proliferation of content almost greater than what they can generate themselves: press articles, documentaries, YouTube videos, as well as dedicated Reddit forums, TikTok, and Twitter, where on average more than 50,000 new posts address the topic each day. Undeniably, these algorithms both intrigue and fascinate.

Beyond the abstract analysis of the tool, we wanted to understand its uses and the possibilities it offers. We focused in particular on two tools :

  • Chat GPT for word processing 
  • Stable Diffusion for the generation of visuals

We are interested in what makes these algorithms so fascinating: their uses. The first thing you notice when testing these tools is that—as long as they are not unavailable due to massive traffic on their servers—both interfaces are very easy to use : by typing a few words and clicking two or three times, you immediately get a result.

The uses of generative AIs are directly influenced by the interfaces/prompts that these tools expose to users. For example, ChatGPT relies on an existing language model, davinci, but stages it in a modified version oriented toward dialogue, where any user can easily write a few lines. The structure proposed by OpenAI changes our perception and inspires confidence. The words of ChatGPT appearing one by one give us the impression of a humanized exchange, in which our interlocutor weighs each of its words.

Without being a revolution (natural language has long been present in chatbots like Alexa or Google Home), ChatGPT has the advantage of being an internet-based platform. As the famous adage goes, « on the internet nobody knows you're a dog », in our case : the line between a human and an AI behind the screen is blurred.

 

Since AIs undergo occasional updates, the versions from March 2023 differ from those of previous months. Our study was conducted between December 2022 and February 2023, and the more recent updates of ChatGPT were not taken into account.


Article rédigé par Amélie Guillaumot & Romain Pialat , Designer & Ingénieur R&D au LINC