AI’s Real Turning Point: Authority, Not Consciousness

In 2026, the AI debate is shifting. More than how intelligent machines will become, what matters is which decisions we delegate to them, who receives the productivity gains, and whether people retain their ability to make decisions along the way.

Read separately, the AI news of the past few weeks seems contradictory.

The Economist asks whether AI can become conscious. Dünya newspaper, drawing on Meta and Klarna, reports that companies’ dream of working without people has hit a wall. Bill Gates says AI could become history’s greatest equalizer or its greatest source of injustice. Daron Acemoğlu warns that the current direction could increase job losses and inequality, while the expected productivity effects of AI investment have yet to appear in macroeconomic data. Yuval Noah Harari argues that trust is among the things we must protect in the AI era.

Then the tone of the debate suddenly sharpens.

Jacob Coxon, who worked on foundation-model development at OpenAI and Anthropic, resigns from Anthropic, saying leading laboratories are racing irresponsibly toward self-improving systems. According to Axios, he does so just two months before his equity rights would vest. Three days later, Anthropic CEO Dario Amodei says the industry should slow the pace of model capability development. A few days earlier, OpenAI Chief Scientist Jakub Pachocki similarly wrote that no laboratory had solved alignment and oversight well enough to keep advancing safely at maximum speed for long.

It makes more sense to read these as parts of the same picture.

Because what we are experiencing is not merely the adoption of a new technology. It is the gradual redistribution of economic and organizational decision-making authority.

Consciousness is fascinating. But power does not require consciousness.

At the center of The Economist’s August 20 coverage of AI and consciousness is an intriguing experiment by Anthropic on Claude. Researchers observed a structure in the model’s internal layers, called J-space, resembling a shared workspace where information comes together before the final answer is produced. It has functional similarities to global workspace theory about the human brain.

But Anthropic itself draws an important boundary: the findings do not show that Claude has human-like experiences or feels anything. Complex internal processing and subjective consciousness are not the same thing.

That distinction matters, but business and society face another, more urgent issue.

A machine does not need to be conscious to exert influence.

It can assess a credit application. Rank investment alternatives. Screen out a job candidate. Plan steps in a cyberattack. Set prices, write code, assign other agents, gather information, and determine which options people are presented with.

In other words, it can be an economic actor without having subjective experience.

Harari’s recent statements should be read in this light. In a long interview with The Economist published on August 26, he describes AI as one of humanity’s largest psychological and social experiments, emphasizing the shift of trust from people to algorithmic systems. His September 8 Digital X talk in Cologne is titled ‘AI and the Tower of Babel.’ It focuses on how societies can preserve understanding, trust, and a shared sense of direction as machines increasingly determine how information is produced and distributed.

Alongside the consciousness debate, another question therefore opens up:

Which matters more: a system being able to say ‘I,’ or us ceasing to challenge its decisions?

The second could happen much sooner.

Meta’s experiment revealed the limits of measurement more than the limits of AI

Dünya’s September 8 article on companies’ frustrated ambitions to work without people presents Meta and Klarna as concrete corporate examples. The Reuters investigation behind it describes Meta’s experiment in detail.

Meta’s Project OT transformation initiative included scenarios for reducing some teams by as much as 60%. The goal was for small AI-native teams to use AI agents to do the work of much larger organizations.

At first glance, some indicators were indeed impressive. Code changes in internal systems increased year over year by 220% with AI use.

But the increase in new or improved features reaching end users was only 36%.

Meanwhile, major technical and security incidents rose by 40%, and the firefighting time employees spent resolving them increased by 70%. Meta laid off around 10% of its workforce in May but canceled a second major restructuring wave planned for November. Zuckerberg later acknowledged to employees that AI agents had not advanced as quickly as he expected.

To me, these figures reveal one of the most important measurement errors in the AI debate.

The volume of work produced and the economic value created are not the same thing.

Financially, the equation is familiar: we do not know real productivity until we deduct error costs, rework, security risks, control requirements, and operational disruption from the increase in gross output.

An employee may produce three times as much code with AI. But if much of it needs to be rechecked or corrected, or creates operational risk elsewhere, the company’s net gain is not threefold.

Klarna should be read similarly. AI took on enormous volumes of routine customer service work. But the company later strengthened access to people again, moving toward a hybrid model in which AI handles routine requests and customers can reach a real employee when needed.

So the lesson from Meta and Klarna is not that AI failed.

The more valuable lesson is this: a task being suitable for automation does not mean the entire human role containing that task is unnecessary.

Confusing the two can become an expensive management mistake.

The most important part of Acemoğlu’s warning goes deeper than employment

In an August 21 article for Nature, Daron Acemoğlu questions the AGI race itself and proposes pro-worker AI: systems that expand human expertise rather than focus on eliminating it.

Speaking to CNBC-e on September 11, he emphasized that AI’s impact was not yet clearly visible in productivity statistics and that the current direction carried greater risks of job losses and inequality. In a BBC interview just nine days earlier, he raised a more political issue: AI will transform almost every part of society, yet people have very little say in the form that transformation takes.

I think another study Acemoğlu published in 2026 with Dingwen Kong and Asuman Ozdaglar is at least as important for business as the employment debate.

Its title is ‘AI, Human Cognition and Knowledge Collapse.’

The model presents a paradox: while agentic AI can help people make better decisions today, it may gradually reduce their effort to acquire the knowledge and learning needed for decision-making. If that process becomes strong enough, society’s collective knowledge itself can weaken.

Another NBER study published in September adds intriguing field evidence to this theoretical concern.

David Autor and colleagues followed 133 patent lawyers at 11 US intellectual-property law firms for three months. AI-assisted lawyers produced higher-quality work, with the largest immediate gains among junior lawyers.

Then the AI was removed.

When the lawyers were asked to perform a task requiring professional judgment on their own, lasting performance gains remained mainly among experienced lawyers. The junior group showed no lasting improvement on average. The researchers’ possible conclusion matters: acquiring durable skills from AI may require preexisting foundational expertise.

Companies face a risk here that does not yet appear on their balance sheets.

I call it expertise debt.

It resembles technical debt. We gain speed today and leave costs for the future.

Giving a junior analyst’s first model, a young lawyer’s first draft, or a new engineer’s first problem to AI raises short-term productivity. But if we also remove the learning steps through which people develop expertise, where will senior experts come from ten years later?

A CFO does not become a CFO solely through financial textbooks. They encounter poor forecasts, experience liquidity crises, recognize faulty assumptions, and learn why an investment works in Excel but fails in real life. Much of judgment develops through precisely this friction.

If AI removes all that friction, it may reduce today’s costs while consuming tomorrow’s human capital.

Gates’s warning about inequality is really a distribution problem

One of the most striking statements in Bill Gates’s August 26 essay is his view that AI could become the greatest equalizer ever invented, or the worst source of injustice.

This is often presented as a warning against AI. Yet Gates’s text is not opposed to the technology. It asks how the value created by it will be distributed.

He even proposes a category called Human Reserved: tasks retained by people for economic, social, or human reasons even when machines could technically perform them. In areas such as education and mental health, he envisages hybrid models in which AI expands human capacity while people remain in charge. Gates also explicitly discusses taxing AI tokens and robots.

A simple economic reality underlies this debate.

Saving labor may be rational for an individual company. But when every company moves in the same direction under the same incentives, the social outcome is not merely the sum of individual corporate optimizations.

If all productivity gains accrue to capital owners, entry-level jobs disappear, and routes to developing new expertise narrow, technology may increase total output while worsening income distribution and social mobility.

AI’s social outcome is therefore not written inside the model.

It will be written in ownership structures, incentives, taxation, competition policy, and institutional design.

More striking still: Calls to slow down now come from inside the industry

In June, Planet Labs founder Will Marshall wrote an article for The Economist titled ‘Humanity isn’t ready for the coming intelligence explosion’, arguing that AI companies were caught in a race they struggled to leave. He warned that if AI accelerates AI development through recursive self-improvement, technological progress could outpace the institutions meant to govern it. His proposal was less an abstract call to stop AI than a case for verifiable safety boundaries and international coordination.

In June, this view might have seemed radical.

By September, it was coming from inside AI laboratories.

In his September 6 essay An Alien Mind, OpenAI Chief Scientist Jakub Pachocki explicitly wrote that he believed no laboratory had solved alignment and monitoring well enough to safely sustain maximum-speed scaling for long. He added that he expected voluntary slowdowns to become normal until common safety thresholds were established.

On September 9, Jacob Coxon left Anthropic. Having worked on pretraining at both OpenAI and Anthropic, he criticized the race toward self-improving superintelligence and identified the next one or two years as critical. This was a warning from someone inside a model-development team, not an outside observer.

On September 10, Anthropic published its own threat intelligence report. It said it had detected and disrupted attempts to misuse Claude in cyber operations, surveillance, fraud, biological misuse, and weapons development, among other areas. It also stated plainly that some safeguards had stopped most requests, but not all.

And on September 12, Anthropic CEO Dario Amodei said the industry needed to slow down the pace of model capability development.

His three-part proposal is notable: independent auditors with employee-level access inside frontier laboratories, coordination among companies on shared safety standards, and international coordination. Anthropic also commits to taking unilateral steps on independent oversight.

What matters here is not who is an AI optimist and who is a pessimist.

More interestingly, the people closest to the technology are increasingly converging on the same word:

control.

Perhaps we have been asking companies the wrong question

So far, much of the discussion of AI investment has centered on these questions:

How many people’s work can it do?

How many hours does it save?

How much code does it write?

How much does it reduce costs?

These questions are necessary. But they are not sufficient.

A company’s real AI performance cannot be measured solely by how much faster it works. Decision quality, error detection, whether it develops human expertise, where the control chain sits, and whether the organization can notice when the model is wrong all belong in the same equation.

Meta shows that activity volume and value creation can diverge.

Autor’s patent-lawyer experiment shows that immediate performance and lasting expertise can diverge.

Acemoğlu argues that private productivity gains and social welfare may not be the same thing.

Gates reminds us that technological abundance does not automatically lead to fair distribution.

Harari points to the quieter change beneath them: when people begin not only to have systems do work but also to trust them, authority begins to shift too.

The warnings from Coxon, Pachocki, Marshall, and now Amodei reveal a more advanced edge of the same problem: as technology begins accelerating its own development, the gap between capability and our ability to control it may widen.

My conclusion from all of this is not that AI has failed.

Quite the opposite.

It is becoming successful enough that where we use it is becoming more important than how we develop it.

This is no longer a matter to leave to the technology department. It belongs jointly to boards, CFOs, risk managers, HR, and regulators.

Because beyond a certain point, the issue is not buying software; it is designing decision rights.

Who will set the credit limit?

Who will challenge the forecast?

Who can override the model’s assumption in an investment scenario?

What authority will the person rejecting AI’s recommendation hold in the organization?

When an error occurs, will responsibility rest with the algorithm or management?

I think this is where the AI debate of 2026 needs to go.

We do not know when machines will become as intelligent as people. Nor whether they will become conscious.

But without waiting for either, we can give them something very important:

Authority.

Throughout history, power has been governed not only by the capacity of whoever holds it, but also by the quality of the institutions that constrain it.

AI will be no different.

That is why, at Ainsideme, we treat AI not as a decision-maker replacing financial judgment, but as a layer strengthening decision architecture: human judgment → AI speed → human validation. The goal is not to automate as many decisions as possible, but to design correctly which decisions can be accelerated, which must be questioned, and which must remain with people.

Because the real competitive advantage ahead may not belong to the company using the most AI.

It may belong to the company that knows what it should not delegate to AI.