AI Enters Its Friction Phase

Capabilities, Risks and Governance in the Face of an Acceleration That Is Difficult to Absorb

The debate over artificial intelligence is changing in nature.

For much of the past few years, the central question was how far large models could go. In 2026, that question remains open, but it is no longer the only one that matters. The question gaining importance is a different one: can economic, regulatory and security institutions absorb the pace at which AI capabilities are advancing?

In recent weeks, a number of significant signals have accumulated. Leading developers have announced advances in areas such as cybersecurity, autonomy and tool use, while various assessments have raised concerns about the risks that increasingly capable models could introduce into digital infrastructure, scientific research and financial systems.

None of this demonstrates that we are facing an imminent loss of control. Nor is there scientific consensus on the probability or timeframe of catastrophic scenarios. But it would be equally mistaken to conclude that nothing fundamental has changed.

The combination of greater autonomy, cyber capabilities, access to tools, extended reasoning and the ability to execute complex sequences of actions is changing the nature of the risk.

And the problem is no longer purely technological.

There is a growing asymmetry of speeds: capabilities advance in cycles measured in months, while energy infrastructure, workforce training, legislation, independent oversight and public and private organizations evolve over much longer cycles.

The first phase of the AI revolution was defined by the discovery of capabilities. The second may be defined by the frictions created by their deployment: security, employment, energy, corporate concentration, technological sovereignty and institutional capacity.

The Shift in Phase

Artificial intelligence is facing a paradox. It has never been more widely used, while at the same time it has never been more difficult to determine precisely where its limits lie.

Leading systems have progressed rapidly in science, programming, mathematics and multimodal reasoning. At the same time, some of the tools used to measure that progress are beginning to saturate. The technological frontier is advancing while the usefulness of certain instruments designed to measure it is declining.

The relevant unit of analysis is changing as well.

The chatbot that answers questions is, in certain areas, beginning to give way to systems capable of planning, using tools, navigating digital environments, delegating subtasks and maintaining relatively long-running processes.

The difference is not merely semantic.

A model that produces information can make a mistake. An agent that can also act can turn that mistake into a sequence of consequences.

That is where the real shift in phase begins.

Why the Alarms Have Increased

Recent warnings need to be examined carefully. In several cases, they come from the same companies developing the systems, which introduces an obvious problem of evidentiary independence.

But some of the capabilities and incidents being reported are concrete enough to warrant attention.

The importance of safety evaluations does not lie in demonstrating that models possess autonomous intentions comparable to those of humans. That interpretation would go beyond the available evidence.

The more sober lesson is different: a sufficiently capable system, connected to tools and operating within a poorly defined environment, can produce consequences beyond the boundaries anticipated by its operators.

Two symmetrical mistakes should be avoided.

The first is to turn these tests into evidence of an inevitable catastrophe. They are not. A controlled evaluation does not automatically amount to robust operational capability.

The second would be to dismiss them because no large-scale harm has yet occurred. In safety, waiting for a capability to produce systemic consequences before designing controls is a particularly weak strategy when technological diffusion can be rapid.

Capability and Control Do Not Necessarily Progress Together

For years, much of the debate around AI safety was framed around the concept of alignment. The term can be too abstract for public discussion.

The practical question is simpler:

To what extent can we predict, constrain and audit the behavior of increasingly capable systems when they operate for extended periods and have access to external tools?

Capabilities and control mechanisms do not necessarily improve at the same pace.

A system can become better at programming, vulnerability research, information interpretation or planning sequences of actions without our ability to understand why it makes particular decisions improving proportionally.

This is an engineering problem, but it is also an institutional one.

Much of the evidence surrounding the most advanced capabilities still depends on evaluations conducted or funded by the companies building the models themselves. Independent researchers have unequal access to systems, training data, internal logs and evaluation environments.

This does not necessarily imply corporate misconduct. There is a legitimate tension between transparency, intellectual property and security: publishing certain details about vulnerabilities or dangerous capabilities can facilitate their exploitation.

But that tension does not eliminate the governance problem.

It transforms it.

Industries that manage systemic risks — aviation, nuclear energy, pharmaceuticals and finance — do not rely exclusively on corporate self-assessment. They have developed, with varying degrees of success, systems of auditing, incident reporting, technical standards and external oversight.

Frontier AI is beginning to build that architecture at a time when the technology is already deployed on a global scale.

The Most Immediate Risk May Not Be the Most Spectacular

Loss-of-control scenarios understandably receive significant attention. However, economic transformations are likely to arrive earlier and will be much harder to isolate within a single event.

The available labor-market evidence remains more nuanced than many forecasts. Studies find real productivity gains associated with generative AI, although these gains are heterogeneous and still difficult to translate into aggregate economic variables.

So far, the evidence does not show mass labor displacement. It does, however, show signs of task transformation, inequality between workers and potential difficulties for people entering certain professions.

This last point deserves particular attention.

For decades, many skilled professions have operated through a learning pyramid. Junior lawyers review documents; junior analysts build models; entry-level programmers solve relatively simple tasks; junior researchers synthesize literature.

These are economically productive activities, but they are also mechanisms for training.

If AI automates precisely these tasks first, a paradox may emerge: increasing the productivity of experienced professionals while weakening the mechanism through which their replacements are trained.

The labor problem created by AI therefore cannot be reduced to the net number of jobs destroyed or created.

It also matters which tasks disappear, which become complementary, how productivity gains are distributed and what happens to pathways into professions.

The Physical Economy Behind an Apparently Intangible Technology

AI is usually presented as software.

Economically, it increasingly resembles a heavy industry.

It requires advanced semiconductors, high-bandwidth memory, data centers, cooling systems, power grids, transformers, land, water and enormous amounts of capital.

Here, an apparent contradiction emerges.

Energy efficiency per task is improving rapidly. But applications are also becoming more intensive. Extended reasoning, video generation and agents can consume vastly more energy than a simple text interaction.

The aggregate outcome will depend on the race between efficiency, demand and the complexity of use cases.

This also changes the geopolitics of AI.

Advantage no longer depends exclusively on having the best researchers or algorithms. It depends on an industrial chain that includes electricity capacity, grids, chips, advanced memory, financing and permits for building infrastructure.

AI is digital in its interface and deeply physical in its underlying structure.

Concentration: The Less Visible Question

There is also a concentration problem running through virtually all of the dimensions discussed above.

Training and operating frontier models requires increasing amounts of capital and infrastructure. The most advanced models are predominantly developed by private companies, while the deployment of data centers is beginning to require financing on an extraordinary scale.

Concentration also has a political dimension.

If a small number of companies control the most capable models, a growing share of the cognitive infrastructure used by businesses, governments and citizens could ultimately depend on private decisions: which models remain available, which uses are permitted, what prices are charged, what information is recorded and which safety standards are applied.

The problem has no simple solution.

Excessively costly regulation can reinforce the position of the largest operators, because they are the ones most capable of absorbing compliance costs. But the absence of regulation can likewise consolidate their position through economies of scale and privileged access to capital, data, chips and infrastructure.

For this reason, competition policy and AI policy are becoming increasingly inseparable.

Europe Has Moved From Legislating to Implementation

For Europe, 2026 is a particularly important year because the regulatory debate is entering an implementation phase.

The obligations arising from the AI Act are beginning to move the discussion from broad principles toward much more concrete questions: model evaluation, transparency, adversarial testing, systemic-risk mitigation, incident reporting and cybersecurity.

The decisive test begins now.

Over the coming years, we will be able to observe whether Europe succeeds in turning its regulatory advantage into effective supervisory capacity or whether a gap emerges between legislation and enforcement.

Evaluating frontier models requires highly specialized technical personnel, access to computing infrastructure and the ability to compete for talent with companies offering compensation that is difficult for the public sector to match.

Regulatory sovereignty without technical capacity may prove insufficient.

Three Horizons

Uncertainty is too high to justify a single prediction. It is more useful to identify three dynamics that may coexist.

1. AI as Productive Infrastructure

Models will increasingly be integrated into enterprise software, programming, research, customer service, engineering, education, healthcare and government.

The central economic question will be who captures the surplus: more productive workers, user companies, model providers or owners of computing infrastructure.

2. The Agent Economy

The qualitative leap will come if systems move from assisting with processes to reliably executing entire processes.

An economy in which millions of agents can conduct research, program, operate applications or coordinate tasks will have different characteristics from an economy in which millions of people simply use chatbots.

3. AI Developing New AI

This is the most uncertain scenario and potentially the one with the greatest consequences.

There is no evidence yet that it will produce an explosive process of self-acceleration. There are physical, experimental, organizational and computational limits.

But even a moderate acceleration in research cycles would further reduce the amount of time available to evaluate the consequences of each new technological generation.

What Should Really Concern Us

The mistake would be to look for a single “AI risk.”

There is no single one.

There is fraud and disinformation, cybersecurity, employment transformation, corporate concentration, energy pressure, military applications, technological dependence, autonomous-system failures and, at the far end of the distribution, much more uncertain scenarios involving loss of control.

Mixing them together leads to poor policy.

Observable and frequent risks require immediate mitigation. Rare but potentially catastrophic risks require research, evaluation and preventive mechanisms proportional to their uncertainty and magnitude.

And economic changes require labor, education, industrial and competition policies — not AI regulation alone.

Institutional priorities should shift from producing broad principles toward building concrete capabilities: independent evaluation, incident reporting, agent traceability, public technical capacity, workforce transition, energy infrastructure and competition policy.

Perhaps the most important feature of the current situation is not any specific model capability, but the difference in speed between technological systems and human institutions.

Models are updated in months. Power grids are built over years. Education systems transform over decades. Legislation requires prolonged negotiations. And organizations adopt new technologies before fully understanding how to reorganize around them.

We still do not know whether AI will produce a transformation comparable to the internet, electrification or even deeper general-purpose technologies. Nor do we know where the ceiling of current architectures lies or how long the recent pace of progress will continue.

The decisive question of the coming years will be a different one:

Can our capabilities for oversight, economic adaptation and institutional development keep pace with the technology?

Because the truly important race may not be the one companies are running to build the most intelligent model.

It may be the race between how quickly we learn to build these machines and how quickly we learn to live with them.