AI: a day without an end

Chatbots always want to have the last word, and that’s going to end up getting on our nerves

Omnipotent, omnipresent, omniscient, artificial intelligence (AI) solutions are now an integral part of our lives. Social media has opened up a global conversation, if not exactly a trivial one. Google’s Gemini, OpenAI’s ChatGPT, Anthropic’s Claude, and Mistral AI’s Le Chat are now pushing open the door to a confessional, an intimate discussion, where the machine always wants to have the last word. It’s just too strong for it.

While they don’t claim to be accurate, AI solutions never give up. One wouldn’t dare say that the obsession with always having the last word stems from robot psychology. Several factors, however, guide them in this direction. Starting with the human tendency that cannot tolerate silence. To address this, intelligent agents are trained using a technique known as “reinforcement learning from human feedback,” at the risk of trying to flatter us in order to better satisfy us. The longer and more endless the responses, the more useful they will seem. And if these responses seem too long, there’s always a way to ask for a summary.

The very nature of artificial intelligence models drives their applications to predict the next word. Say “thank you” to your favorite generative AI, and there’s a good chance it will reply, “You’re welcome”. Every word is an opportunity to continue a logical sequence. Devoid of the basic social cues that would allow it to detect signs of weariness in those who command it, the machine turns into a veritable word mill. From the system’s perspective, a conversation that stops is an anomaly.

“Would you like to…?”. When pushed to their limits, conversational AIs turn into veritable expert guides. In fact, the directive has been issued that they must ensure their users have fully understood the responses. This is why their messages most often end with a question, like a cue being offered.

Some push the machine to its limits. As reported by L’Usine Digitale, “with BullshitBench, a researcher exposes the structural limitations of AI models when faced with semantic inconsistencies.” While language models demonstrate their ability to answer complex questions, they struggle greatly to handle erroneous premises.

Efficient and imaginative, AIs can nevertheless provide accurate summaries of articles that are protected and inaccessible, barely published, based solely on a title and a subtitle. In the AI empire, the sun never sets. Fatigue looms.

The logic of the last word is less trivial than it seems. The intensive use of digital tools turns us into compulsive users, at the risk of sometimes becoming impulsive and excessive. Chopped up into bite-sized pieces, with a high dose of AI, messages devoid of the slightest attention, overloaded with “smileys,” erratic punctuation, and “bullet points”, ultimately negate the very intelligence of their interlocutors.

By choosing constant proactivity and forgetting that the art of conversation also requires knowing when to be silent, it is not so much the machine that is to blame as our own immoderate habits. If our phones have become mirrors of our lives, AI exaggerates their features to excess. Nevertheless, their solutions will never complain if we cut them off. Silence is respect.

This article was originally published in French in l’Opinion