Computer scientist and artificial intelligence researcher Luc Julia argues that AI is best understood not as independent, human-like intelligence, but as a collection of tools built on statistics and mathematical models. In his view, responsible use requires people to retain control, check the results and consider the resources the technology consumes.
Tools, not human intelligence
Julia compares AI to a toolbox whose different implements are suited to different jobs. He argues that the more general-purpose a tool is, the less effective it becomes at a particular task. He sees the progression from machine learning through deep learning to generative AI as an evolution rather than a revolution: systems have grown more complex and process increasing amounts of data.
He used a personal example to illustrate the difference between human and machine learning. A picture-recognition algorithm, he said, may need 100,000 photographs of cats to identify the animal with 98% accuracy, while showing his one-year-old daughter two cats was enough for her to recognize one. Generative models can be trained on hundreds of billions of data points, but Julia argues that processing vast quantities of data is not the same as human understanding.
Ideas and verification remain human tasks
Julia nevertheless sees a major change in the usability of generative AI: anyone can now give an instruction and have a system produce text, an image or computer code in moments. In his view, however, the creative idea comes from the person, while the machine produces the result from patterns it has learned. He cautions that the technology’s impressive performance can make people forget their own creative and directing role.
Hallucinations—incorrect claims generated by AI—also make checking essential. Julia recalled the case of a New York lawyer who used ChatGPT to prepare a court filing in 2023 that cited several nonexistent court decisions. He also said that he asks different chatbots for his own biography every month and still regularly finds errors in their answers.
In a 2023 Hong Kong study, ChatGPT achieved an average accuracy of 63.4% across ten types of reasoning task. Julia argues that the situation has since worsened, partly because synthetic data produced by other AI models can contain false claims. According to the figures presented, ChatGPT’s error rate rose from 36% to 40% over three years, and other models have also become more prone to errors.
Julia therefore recommends using AI, where possible, in areas users understand themselves, so they can check the answers. Asking a system about a completely unfamiliar subject without being able to judge the result, he argues, is essentially gambling.
Legal and environmental costs
Accuracy is not the only concern: copyright also raises questions. The use of protected material to train models has led to numerous legal disputes, and ownership of works generated by AI is not always clear.
Julia also drew attention to the energy and water required to operate and cool data centers. He does not advocate rejecting the technology, but says people should consider what they use it for and favor important, meaningful tasks. As an example of a use he considers wasteful, he cited generating an image of the Mona Lisa with a mustache; he presented five versions he had generated to illustrate the point.
Julia also examines common misconceptions about AI and the question of machine creativity in his book, Robot, Not Pilot.



