The advent of large language models, or LLMs, marks an evolution in our adoption of AI, sparking questions about the future of jobs and the human-machine relationship. While certain concerns are legit, they should not overshadow our view of the future. Persistently viewing ourselves as rivals to machines, or even attempting to mimic them, only fans the flames of our fears, especially as it becomes increasingly evident that such a competition is lost in advance. Rather than succumbing to skepticism, we should all seize the opportunity AI presents to refocus on the core of what makes us human. Whether in their design or usage, understanding how these models are working, their differences compared to the human brain, and redefining our expertise opens up promising horizons.

Human training

While ChatGPT's ability to generate remarkably natural human-like responses—fuelling an ever-growing anthropomorphism—is astounding, it's important to underline the crucial role human expertise plays in its training, unlike other models (e.g., BERT). Specifically, alongside language masking, which involves hiding words in varied sentences and assigning the task of discovering hidden words to the model, ChatGPT's training necessitates a human evaluation of results based on several criteria such as the relevance of responses, ethics, and adherence to human values. Once this initial stage is completed, reinforcement learning is employed to enhance the model's performance. The idea here is to reward the model, positively or negatively, based on its actions. The model learns the rules and effective response strategies by incorporating these rewards. In the case of ChatGPT, the more its responses align with the reward model—which has learned human preferences—the more it's rewarded. This design process underscores the importance of human expertise in training, ensuring both performance and ethical considerations. Models that do not include these human preferences in their training continue to struggle in matching human-level performance. For instance, recent research by Meta demonstrates that: (1) with language masking, LLMs are capable of constructing word representations considering the immediate context, akin to the brain, and (2) however, the brain is able to enrich this initial layer of representation by considering a broader temporal context and hierarchy to construct a richer understanding of the text. LLMs that don’t incorporate reinforcement stages based on human preferences are thus unable to achieve this sophisticated understanding.

Moreover, LLMs are essentially stochastic parrots that rely on probabilities and lack the capacity for planning, consciousness, or information updating. For instance, when we receive new information, our brains immediately update this information to optimize our ability to predict future events. When we're surprised by some piece of information, the hippocampus—a brain structure associated with memory—understands it's time to restructure the information, switching from a preservation mode to an updating mode. LLMs lack this flexibility: Composed of billions of parameters, it's impossible to know which to update in order to refresh the information, and complete retraining would be costly. Hence, a significant part of the research on LLMs is devoted to overcoming these limitations, in an attempt to draw closer to the distinguishing capacities of the brain.

A renewal of our expertise

When used wisely and in complement to our own qualities, generative AI and LLMs can significantly enhance our capabilities. Research published by MIT scholars, studying the impact of ChatGPT on the performance of skilled workers on writing tasks, shows that the use of ChatGPT enables them to: (1) complete tasks faster, (2) create content that is deemed higher quality in terms of writing, content, and originality, and (3) improve workers' satisfaction with task completion. Notably, while the tool allowed highly capable workers to work faster, it primarily enabled others, who initially had lesser abilities, to increase the quality of their output, to the extent of narrowing the performance gap between the less proficient and the best performers. Other studies have also highlighted AI's ability to improve individual decisions, and the capacity of humans and AI to mutually enhance each other. These findings call for valuing distributed cognition, which advocates the necessity of human expertise in an AI-augmented world. In this context, while AI represents a major technological breakthrough, it's fundamentally a human revolution that calls for a shift in our metacognition, humility, and relationship with the world. Ultimately, technologies don't change societies—it's humans giving sense to them that brings real evolution.

To accomplish this, we must understand our own intellectual limitations and how we complement AI, affirm our curiosity rather than our pride, and learn to ask the right questions. Everyone, in their emotional acceptance of AI, needs to cultivate critical thinking skills to understand both the potential biases of an AI and their own human biases. The question isn't whether AI is perfect, but rather if it allows us to perform better than the human status quo. Intellectual emancipation then becomes a vital lever to gain an enlightened understanding of AI's potential as well as its blind spots. In this respect, AI offers everyone extraordinary opportunities for knowledge access and proves to be an exceptional learning catalyst. It is up to us to translate these advantages into concrete and beneficial actions: this is our part of humanity and expertise. The real threat is not AI itself, but our attempt to transform ourselves into automates: maintaining an advantage over machines primarily involves not acting like them. Therefore, now is the time for us, as humans, to truly understand and redefine the actual place of our expertise, work, and humanity.