The benefits of artificial intelligence

The market, businesses, the state: both theory and economic history have seen it all – and its opposite. But let’s start with the basics: the market must allow room for innovation and competition; businesses must transform knowledge into productive capacity; the state must support collective conditions and make public objectives explicit.

In this context, industries have often called for regulation, but almost never for restrictions. A regulation may benefit those already in the market, as it raises barriers to entry; a restriction, on the other hand, affects the product and its distribution.

On 12 September, Dario Amodei, CEO of Anthropic, called in his essay ‘We Must Pace the Frontier’ for a slowdown in the growth of artificial intelligence models’ capabilities. Within a day, Sam Altman of OpenAI, Elon Musk and Demis Hassabis of Google DeepMind had expressed similar views. Donald Trump, speaking on the sidelines of a tournament at his Irish golf club, attributed the warnings to "negative forces" and reiterated that whoever wins the race for artificial intelligence wins it all. On 22 September, speaking from the rostrum of the United Nations General Assembly, the US President rejected “any attempt to construct a globalist scheme” to control AI. The Global Times, a tabloid published by the People’s Daily – the official newspaper of the Chinese Communist Party – had already interpreted Amodei’s essay as an attempt to hold China back.

For days, the discussion has centred solely on the speed of the models, not on the pace at which businesses and workers are adopting them, nor on their effects on productivity, wages and employment. That it is the manufacturers themselves who are proposing the limit remains an anomaly. The Highway Code was not written by car manufacturers: as a rule, the rules are set by those tasked with safeguarding the public interest, whilst those who produce them negotiate or anticipate them. Those who see this caution as a market calculation too can find some grounds for their view. Anthropic has filed documents for its stock market listing on a confidential basis; OpenAI, on the other hand, postponed its own listing until after 2026 on 12 September, citing security reasons.

However, it would be a mistake to dismiss this as merely a public relations ploy. Amodei proposes independent evaluators within the laboratories, with access to the systems and the freedom to publish results, as well as binding rules and cooperation between governments. She justifies this by pointing out that models are now helping to build their own successors more quickly than researchers can understand what is changing. It is the most significant proposal to come from manufacturers so far. Europe, which has already introduced model evaluations and incident reporting obligations in the AI Act, can recognise its own approach in this and claim to have been the first to recognise the risks involved. As regards the economic effects, however, public policy is still in its infancy. Statistics show how many businesses are adopting artificial intelligence, but we almost never know what they gain from it, how the value created is distributed, or who bears the costs. Governing this transformation requires an understanding of both the capabilities of the systems and the effects of their adoption. To study the latter, Italy has established a dedicated body: the Committee on the Economy of Artificial Intelligence, proposed in these pages on 13 July and set up at the CNEL ten days later.

Amodei refers to the arms control treaties, and the instructive precedent is the 1968 Nuclear Non-Proliferation Treaty. It has held up – for as long as it has – thanks more to the inspections carried out by the Vienna-based Agency than to the word of the signatories, and has crystallised an asymmetry: five powers authorised to possess the bomb, in exchange for a commitment to disarmament that has largely remained on paper, whilst all others are called upon to renounce it. Anyone proposing a treaty today should clarify whether they are aiming for independent verification or for a freeze on the balance of power to the advantage of those already in the lead. Verification requires a third-party institution; the same applies to the economic effects, which should be measured independently of who produces the technology and who adopts it. The United Kingdom has applied this principle to the economy. On 8 June, the UK Treasury and the Department for Science, Innovation and Technology established the AI Economics Institute, chaired by Simon Johnson, winner of the 2024 Nobel Prize in Economics. The institute has no decision-making powers and has signed a declaration of collaboration between the government and Anthropic, Google, Microsoft and OpenAI. This is the model we started with, albeit in a different location.

The difference lies in the link between knowledge and representation. London consults the trade unions; at the CNEL, however, the social partners sit by constitutional provision. Furthermore, the CNEL holds the National Archive of Collective Agreements: measurement can thus be based directly on the agreements themselves, and not merely on companies’ declarations. The Committee did not arise out of nowhere. In 2024, a CNEL working group comprising the National Research Council (CNR), the Italian Institute of Technology (IIT) and the Enrico Fermi Research Centre reviewed the literature on artificial intelligence and work. The conclusion still holds true. The main impact lies in the redefinition of occupations and the algorithmic management of work, rather than in mass replacement. In this scenario, the social partners also play an important role in monitoring the transformations taking place. On 24 January 2025, at Villa Lubin, the CNEL and the European Economic and Social Committee (EESC), together with the economic and social councils of six EU countries, signed a declaration on artificial intelligence and industrial relations and launched OPERA, the Observatory on Policies and Industrial Relations for Participatory Artificial Intelligence. On 1 May, the CNEL dedicated Labour Day to the theme of participatory generative artificial intelligence and to the principle that ‘algorithms can only be governed collectively’, emphasising the role of social dialogue, collective bargaining and workers’ participation in productivity gains. OPERA documents cases where artificial intelligence is introduced into businesses with the consent of those who work there. However, it remained to be measured, on a national scale and using data shared by the parties, what effects this was producing. The CNEL Committee is the missing link. The decree establishing it entrusts the Committee with building a public evidence base on the economic effects of artificial intelligence by region, sector, company size, gender and generation, linking its adoption to productivity, employment, wages and distribution. Without any regulatory functions and operating with full scientific independence, it uses methods that are replicable and comparable at an international level. It produces an annual report for Parliament and the Government, a dashboard of indicators aligned with OECD standards, and evidence briefs.

An hour saved by artificial intelligence could mean an extra appointment at a GP’s surgery, a better product, time to learn, a higher salary, or a reduction in staff numbers or greater profit margins for those selling the technology. The outcome depends on how businesses are organised, on competition, on skills and on contracts, rather than on the machine itself.

By ‘distribution of dividends’ we mean a distribution that is efficient, profitable and inclusive of productivity gains. This is what the history of the industrial revolutions of the last two centuries suggests: productivity gains have taken root when they have been distributed, above all, fairly – in a way that combines efficiency, profitability and inclusion.

Those who work know the exceptions and informal solutions that make an organisation function – knowledge that artificial intelligence requires and that is shared if one expects to grow, not if one fears becoming irrelevant. OECD studies have, since 2023, linked staff training and consultation to better outcomes, without, however, proving a causal link. It is precisely this link that the CNEL Committee is called upon to verify.

The CNEL Committee will focus on the nature of change rather than the extent to which artificial intelligence is used. Daron Acemoglu, David Autor and Johnson himself, in ‘Building Pro-Worker Artificial Intelligence’ (2026), distinguish five forms of technological change, depending on whether it increases the return on labour or capital, automates existing tasks, levels out skills or creates new tasks. Only the last form is unambiguously favourable to labour, as it creates demand for new skills; it is also the least well-funded, due to misaligned incentives between businesses and developers, inertia stemming from decisions already made, and a pervasive pro-automation ideology. The first task will therefore be to understand which firms are adopting artificial intelligence, using which technologies and in which areas, and which occupations and skills are being replaced, created or made complementary. This is the information that collective bargaining needs most, because how a gain is distributed depends on the nature of that gain.

Christine Lagarde distinguishes between the cloud, which stores data, and the model, which analyses it: whoever provides the model can thus learn what an industry knows and, in the long run, compete with it. Soon, artificial intelligence will be dispatching trains and selecting which tax returns to audit. A revocation of access would then have far-reaching effects across all sectors, giving suppliers leverage over Europe that no trading partner has ever had. The dilemma lies between protecting data – thereby sacrificing some growth – and rapidly adopting artificial intelligence, whilst risking the loss of the freedom to organise the economy according to one’s own values. For the European Central Bank, rapid adoption could result in a productivity increase of up to four percentage points over a decade, with a workforce that – for demographic reasons – will shrink by over one million people a year. The second area of investigation concerns data portability, the costs of switching between providers and alternatives for essential functions. A firm may become more efficient whilst, at the same time, losing bargaining power vis-à-vis the provider. According to Albert Hirschman’s (1970) categories, those who cannot exit a relationship have no recourse other than their voice; but a voice without a credible alternative carries little weight. ‘Technological sovereignty’ thus becomes something verifiable, on a company-by-company basis. However, at European level, there is no body to measure its effects. The CNEL Committee’s collaboration with the OECD and the EESC is the first step towards the European network proposed in July in these pages, modelled on CERN, the European Organisation for Nuclear Research.

Measuring is half the task; the other half is turning that measurement into an agreement. And Italy has a precedent. The Protocol of 23 July 1993 on income policy linked wages to a public benchmark: the target inflation rate. Today, that benchmark must be redefined around the dividend from artificial intelligence, which should be estimated by a third party using public and replicable methods, as neither party can do so without the other contesting it. The Committee measures; the Permanent National Commission for Workers’ Participation, established within the CNEL by Law 76 of 2025, translates the measurement into criteria and clauses. The progressive taxation proposed by Draghi and the tax envisaged by Amodei remain possible options, but they come into play at a later stage, whereas collective bargaining operates at the point where value is created. Regulation 2026/1744 has postponed the AI Act’s rules on high-risk systems – including those used to manage workers – until 2 December 2027. Europe has thus postponed the obligation to inform them by sixteen months; Italy can, through contractual means, bring forward the right to participate.

On 15 September, at the Quirinale, Sergio Mattarella called on states and supranational bodies to restore their ‘lost authority’ in the face of private entities endowed with immense financial and technological power, as Draghi and Lagarde had done in the same period, albeit with different emphases. That authority is regained first and foremost through knowledge, even before command. William Thomson, better known as Lord Kelvin, the British physicist who in 1883 delivered a famous lecture on measurement, said that “What cannot be measured cannot be improved. And it is also true that what cannot be measured cannot be governed”.

A company may adopt artificial intelligence because it is already more productive, or it may increase output per employee simply by stepping up the pace of work. Attributing everything to technology would lead to misguided recommendations. Those calling for a slowdown must specify how fast we are going and who is measuring it. The performance of these models is measured primarily by their manufacturers; as for the value they generate in Italian factories and offices, there is now a public body tasked with determining this.

The first Report, in 2027, will provide indicators, scenarios, stated margins of error and, above all, the nature of the change underlying each gain. How much of the dividend to allocate to wages, profit, training or time will be decided by businesses, workers and institutions, based on figures that none of the parties will have produced on their own.

An agreement can only allocate – even in a contentious manner – what has been measured.