L’intelligence artificielle et ses défis d’ici 2035
“Since ChatGPT 3.5 was launched in November 2022, artificial intelligence has evolved from a specialist software tool into a mass consumption product, with more than 1.5 users around the world every month, and 25% of the active population in developed countries using it on a daily basis”, Arnault Barichella states in his introduction to this study of the Jacques Delors Institute, commissioned by CRiP (Digital Responsibility Stakeholders, known by its French name Club des Responsables d’Infrastructure, de technologies et de Production informatique) (our translation throughout).
Amongst other negative effects, “agentic AI carries a risk of ‘tsunami’ affecting so-called ‘white-collar,’ office jobs up to 2028 and beyond, representing the majority of jobs in Western countries. More than 300 million jobs worldwide are believed to be vulnerable to automation in the short term, while 150 to 200 million new jobs could be created, for instance in digital infrastructure, cyber-security and AI management”, the author stresses.
AI is also wreaking havoc in the education systems. With “the risk of a certain intellectual laziness developing, or even a pronounced drop in the capacity for critical thought of young students compared to generations who never experienced AI at school or university”. “Indeed, if LLMs can do all homework in all subjects or at the very least, do them well enough to achieve an average and passing grade, there is a risk of evolving into a society in which young people no longer have the same intellectual capacity as their parents or grandparents. This includes complex reasoning ability, which is vital to devolve successfully on the employment market, at a time when job opportunities for young graduates are on the decline precisely because of the AI revolution. The risk is that we end up with a generation educated by ChatGPT and dependent on it, no longer capable of doing even basic work without recourse to AI. Obviously, it is vital to train the younger generations in AI tools, as these are becoming vital on the employment market. But this should not be at the expense of developing free and critical thought, which is necessary to involve in a democratic society in which citizens are called upon to vote regularly and in order to be competitive on the employment market”, Barichella explains.
There is the potential threat of an identity crisis for humanity and serious consequences for society. “The lines between the human and the machine could become blurred, a phenomenon that could worsen over the next several years. An early example is the symptom that has emerged since ChatGPT was launched in 2022, leading to withdrawal and social isolation, in which the temptation to converse with or even confide in AI rather than another human being, often out of expediency, has spread rapidly. In certain cases, we have reached a point at which the machines become a confidant, or even a therapist, with examples of harmful psychological pressure that could lead to extreme situations involving murders or suicides”, the author stresses.
He also points out that “AI ties that has already exhausted the majority of data available on the Internet, risking stagnation in new models. ‘Data summaries’ (compiling data generated by AI and a human creative touch) and decision latency (AI taking longer to generate responses) are inadequate palliatives in the long term. The excessive use of data summaries could cause the model to collapse and the performance to regress, with the risk of a downturn in the speed of technological development up to 2028”.
“A single ChatGPT prompt uses around 10 times more electricity than a Google search. Data centres are already consuming around 415 TWh a year (almost as much as the whole of France), and this is expected to double, to reach 945 TWh by 2030 (the equivalent of the whole of Japan). This is creating major difficulties in energy supply for businesses in the sector, with extremely negative consequences for the environment. Small Language Models (SLM), which use less energy, offer an alternative, but remain limited compared to LLMs because of their specialisation and therefore inability to execute complex tasks”, the author notes, before touching upon the financial dimension: “more than 1000 billion dollars have been invested around the world since 2022, but 80% of AI projects fail to generate a return on investment, twice the rate of conventional IT projects. The return on investment is negative for the majority of start-ups and model developers, but strongly positive for manufacturers of chips and suppliers of hardware (Nvidia) or providers of Cloud services (Azure, Google Cloud). The risk of a speculative bubble – along the lines of the dot-com bubble in 2000 – is very real, although a consolidation of the market up to 2028 with greater or lesser corrections would appear to be the most likely hypothesis”.
“AI [also] increases disinformation capacity: instantaneous production of convincing content in numerous languages (texts, videos and audio), ultra-realistic deep fakes of political leaders, automated election manipulation campaigns on a large scale. These are often carried out at the behest of authoritarian regimes (Russia, Iran, North Korea, etc.) to divide the West and promote the rise of populism by undermining democracies from the inside. The AI bots of 2026 are capable of engaging in conversations autonomously with psychological micro-targeting, making them extremely difficult to detect. In Europe, the AI Act (articles 12 and 13) requires auditability, transparency and ‘automatic logging’ of systems throughout the model’s life cycle, and proposes for the continuous correction of bias (MLOps, AIF360, Fairlearn) and the technique of watermarking”, the author observes, noting however that “the ongoing simplification process of digital legislation by the European Commission (Digital Omnibus) could water down some of these protections: it is therefore vital to seek compromises”.
To bring the disinformation exacerbated by AI back under control, the report recommends that the following be set in place in the very short term (2026-2028): (1) secured data sandboxes managed by the European AI Office to allow businesses to test their algorithms on EU certified data to detect bias; (2) a compulsory dynamic algorithm passport for every high-risk AI model containing the history of the training data, successive bias tests and a modification log; (3) automated, real-time cross-platform signalling in the event that a manipulation bias is detected; (4) reinforced watermarking (compulsory integration of meta data (C2PA) at hardware (sensors) and software level for any content in the European public domain).
AI represents a systemic ecological challenge: the demand for electricity in data centres optimised by AI is expected to increase more than fourfold between now and 2030. On top of this, they require massive water consumption: 4 to 6 billion m³/year – the equivalent of 50% of the water consumption of the United Kingdom – rising steeply every year, as well as the electronic waste from specialist chips with a short life, the author explains, stressing that “taken together, these factors are generating increasing pressure on local ecosystems”, whilst recognising that “AI is also making a positive contribution to the ecological transition in other cases with, for instance, the optimisation of the electricity network via smart grids, precision agriculture or satellite monitoring of greenhouse gas emissions”. It also states that “current simplification process (Omnibus) of climate standards” carries risks and that “a fair balance must be struck if we are to avoid sacrificing ecology on the altar of competitiveness”.
The report goes on to recommend the introduction of the following in the very short term (2026-28): (1) an environmental audit based on real-time access to the electricity and water meters of data centres; (2) a proportionate digital carbon tax for models exceeding a certain threshold of daily requests, with reduced VAT for more economic processes; (3) a standardisation of Small Language Models (SLM), with a requirement upon suppliers to propose a pared-down version for simple tasks; (4) quotas of water per teraflop (upper limit of litres of water consumed per calculation unit); (5) a ban on using drinking water for cooling in water stress areas; (6) a repairability/durability index of AI infrastructure: recycling, compulsory re-use of equipment.
“Europe is hugely dependent on American technologies at three levels: the Cloud (65% controlled by AWS, Azure, Google Cloud), material/processes (GPU Nvidia virtually exclusively), and the basic models used in everyday work tools (ChatGPT, Microsoft 365, Google Workspace) on a large scale. The American Cloud Act and its extra-territorial clauses directly affect the digital sovereignty of the EU, as it neutralises the protection of the GDPR. The EU responded with measures including sovereign Cloud labels (EUCS and SecNumCloud) and by investing in European digital infrastructure such as super-computers (EuroHPC network), with for instance the AI Continent Action Plan and the Data Union strategy. The aim is to subsidise research and support European champions such as Mistral AI (France), DeepL (Germany) and Silo AI (Finland). Despite these efforts, 65% of organisations in the EU admit that they are unable to remain competitive without at least partially using non-European suppliers. The ‘Draghi’ report (2024) identified the regulatory excesses as a hindrance to innovation in the EU in the digital sector, although many economists have flagged up other factors such as the persistent fragmentation of the single market”, Barichella goes on to explain.
So how are all these challenges to be faced? The report estimates the envelope necessary between now and 2032 protect economic sectors negatively affected by AI at a minimum of 100 billion euros (50 billion from the forthcoming multi-annual financial framework (MAFF) + 50 billion in co-financing by the member states and private capital). Within this envelope, at least 50 billion should be earmarked for training needs, or around 10 billion a year over the next five years, starting in 2026.
A second envelope of at least 300 billion euros, with 50% to come from the European budget (2028-2034 (MAFF) and 50% from the member states and private capital, will be necessary to reduce dependency on foreign technologies, with “the aim of pragmatic resilience and targeted sovereignty”. This would allow the creation of a digital sovereign fund, with an envelope of at least 30 billion/year over a decade from 2026, to pay for sovereign critical digital infrastructure (Cloud, super-computers, data centres, AI factories) from public funding, public-private partnerships or the “robot tax”.
The report also calls for (1) a strategic public procurement policy, with a European preference in public tendering for European AI solutions; (2) a Data Union strategy: pooling data in key sectors (healthcare, industry, energy); (3) relocation of the production of advanced chips (GPU, AI accelerators) to Europe, in the framework of the new Chips Act; the creation of a capital markets union no later than 2028 to “allow European AI ‘unicorns’ to raise funding without being bought up by foreign groups”.
Open Source, which the Commission highlights in its recent communication (of 3 June 2026), is a software development and distribution mode where the source code is made public and, the author stresses, “provides strategic autonomy (possible interruption to access), transparency (absence of ‘back doors’ often used for spying) and making it possible to break away from ‘vendor lock-in’”. “Open Source gives the EU the possibility to compete with the USA with far fewer financial resources: this method has already allowed the emergence of European champions (Mistral AI, Odoo, Grafana Labs, SUSE). But Open Source alone is not enough if the hardware remains American, which in the majority of cases it is, particularly as the GAFAM are also leaders in this field. This means that Europe must dovetail Open Source with the reinforcement of sovereign digital infrastructure, the integration of the single market and the Data Union strategy”, Barichella argues. He goes on to estimate that between now and 2035, around 4 billion euros extra must be redirected every year into Open Source, with a focus on AI and the Cloud (included in the second financial envelope of 30 billion euros/year) +1% of public IT budgets earmarked to maintain the existing Open Source.
“Under the AI Act, technologies are regulated on the basis of the degree of risk they represent, with four categories: unacceptable AI risk (prohibited: social rating, subliminal manipulation), high risk (strict regulation: justice, employment, health), limited risk (compulsory transparency: chatbots) and minimal risk (voluntary codes of conduct)”, the author recalls, observing however that there are still certain gaps in the European legislation: “lack of flexibility and overly slow revision process in view of the swift development of the sector, insufficient attention paid to the collective societal impacts of AI (disinformation/democratic manipulation) and excessive use of self-assessment by businesses (particularly with the Digital Omnibus) weakening compliance”. In order to plug the gaps in the European AI Act, he recommends, amongst other things, a swift update of the risk categories (delegation to the European AI Committee of annual revisions with no complete legislative procedure) and expanding the criteria of the “unacceptable risk” category, particularly for the facial recognition biometric.
Finally, the report makes the case for the systematic integration of the ISO standards in the AI Act – governance (ISO/IEC 42001), risk management (ISO/ IEC 23894), data quality (ISO/IEC 5259), ethics and bias (ISO/IEC TR 24027) – and argues that “the EU must require its principal partners to adopt European standards of the responsible use of AI, on the strength of its position as the largest single market in the world and the extra-territorial scope of the AI Act”. (Olivier Jehin)
Arnault Barichella. Révolution des technologies de l’intelligence artificielle à l’horizon 2035 – L’humain dans la boucle : réguler et définir des limites pour garder le contrôle (available in French only). Institut Jacques Delors. June 2026. 95 pages. The report is available to download free of charge on the Institute’s website: https://aeur.eu/f/mhn