Written by David Kemp.
The ability of generative AI (‘GenAI’) to generate plausible text, images, music and computer code in response to human prompts is impressive. GenAI promises huge productivity gains in many domains, and large amounts of financial and political capital are staked on its success. Nevertheless, the current generation of models exhibit well-publicised weaknesses that might not simply disappear by using more data and processing power or smarter training. This paper looks at those limitations, and the lower-profile alternative approaches to AI that could overcome them and even provide the EU with a competitive advantage.
Since the unveiling of ChatGPT in 2022, we have witnessed what many commentators consider to be the most rapid adoption of new technology in recent times in virtually all segments of society. One cannot fail to be impressed by the speed at which it can produce outputs that some say can be ‘PhD-level’.
At the same time, there is growing awareness that these algorithms often produce plausible but demonstrably false outputs (‘hallucinations‘) or outputs that are inconsistent with logic, maths or ‘how the real world works’; they regurgitate problematic biases present in their training data and cannot provide a reliable explanation of how the outputs were derived. ‘GenAI-optimists’ believe that accuracy and tractability will emerge through scaling up the data and processing power available and adding external (human or automated) processes to fine-tune the models. ‘GenAI-pessimists’, on the other hand, point to the fact that finding solutions to the above problems is slowing down, strongly suggesting that the approach is reaching its limits. In addition, the very high energy and resource consumption of the infrastructure necessary for training and running these ‘brute force’, data-driven applications is putting pressure on already stretched supplies. Given the several trillions of private and public capital invested, now and over the next few years, in the geopolitically strategic race for an ‘all-purpose’ artificial general intelligence (AGI), it is worth asking where GenAI’s limits lie and what other approaches might be explored to put Europe at the forefront of AI innovation.
Potential impacts and developmentsGenAI models, in essence, generate a multi-dimensional map of statistical relationships in the training data between ‘tokens‘ – digital representations of linguistic, graphical, musical or other data. The models can then be prompted to produce new combinations of text, images, music, etc. that are consistent with those relationships. In contrast to traditional statistical regression analysis, these models contain billions of parameters and make almost no assumptions about the form of the relationships between tokens. This gives GenAI its enormous advantage in finding the subtle, multivariate patterns that allow it to output plausible text, code, music, images, etc. in response to prompts.
The flip side of this underlying design, however, is that GenAI (like traditional statistical regression) is prone to mistaking ‘noise’ in the training sample for meaningful patterns and producing bizarre results when it is applied to data outside the training set. In addition, pure GenAI trades transparency of ‘reasoning’ for power: a printout of the billion-parameter map of relationships provides no explanation. This is a problem for the required ‘human in the loop’: how can one have confidence in the result if no one knows exactly how it was reached?
Clearly, these design weaknesses limit mission-critical uses of AI. Scaling GenAI does not appear to eliminate these problems, all the while adding additional problems linked to resource usage. Ideally, we would make the most of the powerful pattern detection that GenAI offers, and overcome some of the limitations of a purely data-driven approach. This requires boostingresearch into ‘hybrid’ approaches which, like the human brain, leverage GenAI’s pattern extraction and matching abilities using built-in, explicit models of efficient reasoning strategies and the real world, as well as strategies formaintaining and developingthose models. These approaches provide a reality check to improve the reliability of GenAI’s probabilistic outputs. They are also more efficient (algorithmically and energetically) as they directly encode readily available, fundamental knowledge rather than requiring everything to be extracted by brute force from the data. Finally, the use of explicit representations and strategies allows their reasoning to be inspected.
Neuro-symbolic approaches, such as IBM’s neuro-symbolic concept learner, combine GenAI with explicit rules for symbolic reasoning (for logic, abstraction and generalisation) to enhance reasoning and explainability. Embodied AI – such as Meta’s Habitat – involves training agents in virtual or physical environments where they are designed to learn through perception, action and feedback, promoting causal learning and the development of sensorimotor intelligence. Cognitive architectures, such as Soar, ACT-R and OpenCog, include explicit models of human cognitive processes, integrating perception, memory, learning, planning and reasoning in a modular way. This enables continuity of learning, goal-directed behaviour and long-term memory. World model learning approaches such as DeepMind’s MuZero, Ha & Schmidhuber’s world model agents and DreamerV3 focus on training agents to derive compact, predictive models of their environment to support causal reasoning, generalisation and efficient planning. Finally, it must be underlined that, regardless of the underlying technology, the pursuit of artificial general intelligence is not necessarily the most efficient route to useful applications. Artificial specific intelligence (AI approaches focused on a specific domain, such as the Nobel prize-winning, protein-folding algorithm, AlphaFold2) gives more reliable and transparent results by combining the subtle pattern detection at which GenAI excels with explicitly encoded, domain-specific knowledge.
Anticipatory policymakingThe reliability issues mentioned above call into question GenAI’s ability to deliver the ‘trustworthy and human centric AI … pivotal for economic growth and … [preserve] the fundamental rights and principles that underpin our societies’, as promised in the AI continent action plan. Most AI initiatives from the European Commission have so far concentrated on the implementation of GenAI rather than on research and development of alternative and complementary AI approaches. The €700 million flagship GenAI4EU programme, for example, states its aim as being to ‘integrate generativeArtificial Intelligence (AI) in Europe’s strategic sectors, and keep their competitive edge’. Consequently, most of the calls for projects focus on applying GenAI in particular sectors and providing the data and computing power it needs. The pursuit of artificial general intelligence has attracted an enormous amount of political and economic interest, potentially to the detriment of equally interesting and possibly more efficient and effective alternatives, including those mentioned above.
To best serve the EU’s goals of competitivity and innovation in AI, EU policy and funding could be targeted more directly to support the wholevalue chain of alternative and complementary approaches to GenAI.It is also important to actively further domain-specific AI (artificial specific intelligence)applications alongside GenAI. This can be achieved through proactively promoting such projects for existing funding programmes and by policy guidance in the next multiannual financial framework and the Competitiveness Fund for Digital Leadership.
Read this ‘at a glance’ note on ‘What if generative AI is reaching its limits?‘ in the Think Tank pages of the European Parliament.
La Commission d'attribution des parcelles situées le long de la Route des pêches convoque les attributaires provisoires pour une séance d'information le 2 décembre.
Une séance d'information est prévue vendredi 5 décembre 2025, à 10 heures, à la Salle Fleuve Jaune du ministère des Affaires étrangères, à Cotonou. Cette rencontre s'inscrit dans le cadre du Projet d'aménagement de la plage Ouest de Cotonou. Elle intervient après les auditions individuelles menées ces dernières semaines.
La Commission veut clarifier plusieurs points. Elle présentera l'état des informations commerciales liées à chaque parcelle. Elle exposera aussi les modalités de poursuite et de mise en œuvre des projets, afin d'harmoniser la suite des démarches.
Le Partenariat Mondial de l'Eau en Afrique de l'Ouest (GWP-AO), en collaboration avec l'Autorité du Bassin du Mono (ABM) et l'Observatoire du Sahara et du Sahel (OSS), a organisé, du 25 au 28 novembre 2025 à l'hôtel Camilio de Grand-Popo, un atelier national de formation des Associations Communautaires de Conservation de la Biodiversité (ACCB). Cet atelier de quatre jours a permis de renforcer les capacités des relais communautaires sur la Gestion Intégrée des Ressources en Eau (GIRE) ainsi que sur les techniques de sensibilisation et de mobilisation pour une gestion durable des réserves et aires protégées du bassin du Mono.
Partagé entre le Bénin et le Togo, le fleuve Mono est de plus en plus menacé par des inondations récurrentes, des conflits liés à la transhumance et à l'usage des ressources naturelles, l'ensablement des cours d'eau, la pollution, la disparition des frayères et des palétuviers, ainsi que la perte de biodiversité. Ces menaces fragilisent la sécurité alimentaire, la cohésion sociale et la durabilité des moyens de subsistance des communautés riveraines.
Face à ce constat, le Partenariat Mondial de l'Eau en Afrique de l'Ouest (GWP-AO), l'Autorité du Bassin du Mono (ABM) et l'Observatoire du Sahara et du Sahel (OSS) ont initié l'atelier de Grand-Popo afin de renforcer les compétences des Associations Communautaires de Conservation de la Biodiversité (ACCB), composées de relais communautaires issus des communautés locales du bassin du Mono.
Chargés de mener des activités de sensibilisation, d'information, d'éducation et de communication sur des thématiques liées à l'environnement, la santé et le développement socio-économique, les relais communautaires jouent un rôle essentiel dans la conservation de la biodiversité.
A l'ouverture de l'atelier, le Coordonnateur national du Projet IREE-Mono a situé le contexte et rappelé l'importance de cette mission. « Votre présence témoigne de votre engagement en faveur de la protection de nos ressources naturelles et de la biodiversité du territoire. Vous êtes les premiers remparts contre les menaces qui pèsent sur notre environnement », a déclaré Léonard Oni. Il a précisé que cette formation pratique et participative s'inscrit dans le cadre du projet Initiative régionale pour l'eau et l'environnement dans le bassin transfrontalier du fleuve Mono (IREE-Mono, 2024–2028).
Le Secrétaire Exécutif du Partenariat National de l'Eau du Bénin (PNE-Bénin), a insisté sur la nécessité de préserver l'équilibre écologique. « Lorsqu'un organisme commence à disparaître, nous devons nous intéresser à ce qui se passe pour éviter une rupture de l'écosystème qui pourrait avoir des répercussions sur nous tous, car chaque organisme à son rôle dans l'écosystème », a souligné André Zogo.
« Vous êtes des éco-gardes. Sans vous, la réserve de biosphère risque de disparaître. Cette formation vient à point nommé et je voudrais vous encourager pour votre engagement à relever ce défi », a déclaré au nom du Directeur exécutif de l'ABM, le Directeur Administratif et Financier, Imorou Daouda.
En dehors des sessions, les participants ont travaillé en groupes, avec des restitutions et échanges en plénière.
À l'issue de l'atelier, les participants ont acquis une meilleure compréhension de leurs missions, valeurs et responsabilités, renforcé leurs compétences en communication et animation communautaire, développé des aptitudes en gestion des conflits et gouvernance locale, et contribué à la préservation de la biodiversité et des ressources en eau du bassin du Mono.
Une feuille de route a été élaborée et validée afin de garantir la mise en œuvre effective des acquis de la formation, avec un mécanisme de suivi pour mesurer l'impact des actions menées par les relais communautaires et les éco-gardes.