LLM Paradox: When AI thinks and people do the work

The authors of this article, Filip Tichý (Partner at Grant Thornton Slovakia) and Jakub Chudík (Co-Founder at Assetario), take you through the world of artificial intelligence in the AI Breakfast series. This article was written without the use of AI.

Large language models (LLMs) are dazzling the world with their ability to generate text, synthesize information, and provide expert advice on demand. From legal analysis to programming code, their outputs often seem almost magical. The more sophisticated LLM models become, the more an interesting paradox emerges: the smarter the LLM model, the more it relegates humans to the role of performing routine tasks and steps designed by LLMs. In this age of advanced artificial intelligence, are we witnessing a paradoxical role reversal - the machine is the creator and the human is becoming the executor?

The LLM Paradox is clear. LLMs are taking on increasingly complex analytical and creative tasks - writing reports, devising strategies, diagnosing problems - while humans are increasingly engaged in copying data, verifying AI-generated information, and implementing proposed steps. The more AI "thinks," the more humans have to do. This is not just a quirk of technological progress - it is a fundamental challenge to the way we organize work and assign value in the digital age.

The roots of this paradox lie in the strengths and weaknesses of LLM. The models excel at connecting words, concepts, and language synthesis, but they have several weaknesses. The first is the so-called "Execution gap" or inability to perform in reality: LLM can suggest hundreds of actions, but it cannot perform a single one. It also has an "Accuracy gap" or lack of accuracy: it generates convincingly sounding but sometimes inaccurate information that must be verified by a human. And the third shortcoming is the "Collaboration gap" or inability to collaborate: LLM is essentially an isolated model, unable to communicate effectively outside its own text window.

This division of labor has real consequences. In many workplaces, LLM is a kind of creative director who comes up with ideas and strategies, while humans are assembly line workers responsible for implementation. The result is a workforce increasingly burdened with mechanical and repetitive tasks, even though the tools they use are becoming more and more intelligent. The promise of AI - to free people from routine and unleash their creativity - seems to have been reversed for now.

AI agents are entering the scene - a new way of applying artificial intelligence that aims to bridge the gap between thinking and doing. The agent workflow uses LLM models as one of its components, but in addition, the workflow also includes subsequent executions as proposed by the LLM. AI agents thus represent a convergence between robotic process automation (RPA) and language models, promising a future where machines not only advise but also act.

Until AI agents are fully developed, the LLM paradox will persist. The irony remains that in our quest to create machines that think, we have inadvertently condemned humans to robotic work.