A doctor faces a complex diagnosis. He has thousands of scientific studies, medical reports, and patient files at his disposal. Yet processing this flood of information under pressure is impossible. Fortunately, his assistant can instantly filter the most relevant facts and deliver a precise, sourced, and concise answer within seconds. This futuristic assistant has become today's reality thanks to advances in artificial intelligence, particularly RAG (Retrieval-Augmented Generation).
However, this technology raises several questions, especially when examining the hallucinations that language models still experience. After exploring what RAG is and considering its evolution into multi-agent systems, we now address an almost philosophical question: by attempting to control AI hallucinations, are we constraining its creative capacity? To explore this tension between reliability and innovation, we spoke with Patrick Meyer, Senior AI Architect and Technical Director at Sopra Steria, and Clément Bénesse, Senior AI Researcher at opsci.ai.
Hallucination or creativity: where to draw the line?
The debate begins with a fundamental distinction. "First, we need to define what a hallucination is," explains Patrick Meyer. "In taxonomic terms, hallucinations take several forms, including failure to follow instructions - such as responding in English when French is requested."
For Clément Bénesse, the nuance is crucial: "There's substance and there’s form. In creative writing, it's the moment when the LLM starts writing in poetry when prose was expected. Certain types of hallucination can constitute a form of creativity."
Hallucination becomes problematic when it leads to nonsense. "LLMs are constructed through data, so their vision is merely the prism of training data," explains the researcher. "Without any constraints, hallucination becomes problematic when it generates responses that contradict basic logical rules."
Probability as a source of creativity
Language model generation relies on a statistical principle: predicting the most probable word according to context. Can this probabilistic approach be creative?
"Nature shows us that randomness is a source of creation," observes Patrick Meyer. "Chromosome division is a source of mixing. When an organism isn't adapted to its environment, it dies. Similarly, a video that doesn't appeal is quickly removed, and memes show us that certain videos are excellent vectors of creativity."
The biological analogy reveals a fascinating mechanism: creativity would emerge from random variations subjected to environmental selection. AI models reproduce this process, but with a major difference.
The human advantage: the art of evaluation
If AI can generate creative alternatives, where does human specificity lie? "Generative AI is capable of generating - it's in its name - but is very poor at evaluating," emphasises Clément Bénesse. "What makes humans unique is our ability to imagine alternative scenarios and evaluate them."
This evaluation capacity relies on a wealth of information that current models don't possess. "Humans can think outside the box, invent new words that make sense, because there's the hormonal aspect, knowledge of the physical world," specifies Patrick Meyer. "For instance, an LLM doesn't know what a falling object is."
The observation is striking: "A human is much more than just digital or textual information. We have enormous amounts of latent information that we store and don't even think about, but which conditions our way of thinking and acting."
From Static RAG to Agentic Systems
This reflection illuminates RAG's technical evolution. "RAG has become a commodity," notes Patrick Meyer. "With a few lines of code, you can create your own RAG today."
But this commoditisation reveals a limitation: classic RAG is designed to answer one question, full stop. "This limitation is linked to the fact that we are only dealing with only one use case, answering question. As soon as you want to summarise a text or compare two documents, it no longer works."
The future is taking shape with agentic systems, where RAG becomes a building block of a more complex ensemble. "The next step is agentic systems - a system whose function is not clearly defined with infinite autonomy," specifies Patrick Meyer. "But beware, agentic systems without reliable orchestration aren't agentic: they're a magma of things that communicate to the point of absurdity."
This evolution towards greater autonomy raises formidable practical questions. "Interesting question: what's the lifespan of an agentic system considering we have some pourcentage of hallucination?" questions Patrick Meyer. "When you have multiple agents calling each other, it quickly becomes problematic."
Clément Bénesse proposes a pragmatic limit: "From 3 depths of agents trusting each other mutually, it starts becoming problematic. You either have to take the result as is or have a means of confronting some form of reality to reduce the generated noise."
Human in the loop: a philosophy
Faced with these challenges, both experts converge towards a balanced vision. "We must keep humans in the loop," insists Patrick Meyer. "They must remain the one who says yes or no."
This approach transcends simple technical precaution to begin attempting to define an AI philosophy: "What users want is to be able to answer any question, but also to generate documents and synthetise data. There's a whole set of services to be created."
The paradox resolves in complementarity: AI generates, humans evaluate and decide. "There's enormous potential in taking this technology and mastering it well internally," observes Clément Bénesse. "Paradoxically, it's education and culture that we need to emphasise currently."
Towards responsible and creative AI
This quasi-philosophical exploration of RAG reveals a broader challenge: how to develop AI that's both reliable and creative? The answer doesn't lie in a binary choice between control and freedom, but in their fine orchestration.
The future of RAG and generative AI will be built on this balance: exploiting the creative potential of probabilistic models while preserving human capacity for evaluation and critical thinking. Because while machines can hallucinate just as they can demonstrate a form of creativity, only humans can distinguish one from the other. For now...