AI researchers doubt current models will lead to achieving human intelligence

Artificial intelligence offers new developments weekly. Dozens of companies with billions in investment are competing to pass the latest human test or become the latest meme. However, an international survey of specialists has revealed strong distrust that continuing down this path will lead us to human-like artificial intelligence; what is known in the industry as artificial general intelligence (AGI). The survey is the work of the Association for the Advancement of Artificial Intelligence (AAAI), a U.S. scientific organization that surveyed 475 AI academics and professionals from around the world: 76% believe it is “very unlikely” or “unlikely” that the increase in current approaches will achieve AGI.

The incredible rise of AI has caused the voices predicting the likely end of humanity at the hands of machines to subside. But the turmoil in the sector remains. This, for example, is the statement by another group called AI 2027: “We believe that the impact of superhuman AI in the next decade will be enormous, even greater than that of the Industrial Revolution.”

At the same time as these grand declarations, Meta has unveiled its two latest models: a small one (Llama 4) and a mid-size one called Maverick. In the comparative rankings of the main models, Maverick jumped to second place, just below Gemini 2.5 Pro and above ChatGPT-4o. But, with a little effort, a handful of specialists have discovered that the version of Maverick that competed was specifically trained to pass the tests. Meta failed the tests of the model competition.

With these subtle pitfalls, the difficulty in defining the threshold of purely human intelligence and knowing how to reach it complicates the picture. “My definition of artificial general intelligence is an AI with the same level of competence and complexity as human intelligence, including concepts as difficult as self-awareness,” says Nuria Oliver, scientific director and co-founder of the Ellis Alicante Foundation. “We are very far from reaching it, and I don’t know if we will ever achieve it.”

EL PAÍS asked a group of Spanish AI scientists for their impressions of this threshold and how it will be surpassed. As in the AAAI survey, there are a variety of responses. “We have the ingredients, although not optimal, to achieve this, but they require certain incremental combinations that need to be explored and even more computing power,” says José Hernández-Orallo, a researcher at the Leverhume Centre for the Future of Intelligence in Cambridge (United Kingdom). “It could be done with less, very possibly, but the question we are asking ourselves is whether it could be achieved by scaling up current approaches, and I think it could.”

However, Senén Barro, a professor at the University of Santiago de Compostela, believes that the current approach will not suffice: “The path to general artificial intelligence is not simply the models we have now, even if we increase their size, give them better inference capabilities, and specialize them in agent architectures. This will allow for significant advances in their capabilities, but what we understand by AGI is much more,” he explains.

An AI with a body and common sense

Barro compares artificial general intelligence to Mars exploration: “We know the path to follow to get people to Mars, although it couldn’t be done today; and there are certain pending R&D&I issues, which are not minor. In any case, we would know how to address them. The same isn’t true of AGI: we still don’t know what path, still distant, would lead us there, and it doesn’t seem at all likely to be one of improving the models.”

Other scientists add another layer of difficulty to this path: the corporeal one. “To advance toward AGI, AI needs to be corporeal and equipped with reasoning and symbolic learning capabilities,” says Carme Torras, a researcher at the Institute of Robotics and Industrial Computing. “When I say corporeal, I don’t just mean robots, but also other objects capable of perceiving, processing, and interacting.”

Researcher Carles Sierra of the AI Research Institute shares this view: “We need neurosymbolic approaches to scale. Neurosymbolic architectures, along with situationality and environmental perception, are the path that will allow us to scale and eventually provide an experience-like notion, necessary for the common sense and agency required for AGI.”

Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition