We are on AGI but still there is AI-Denial

How many jobs AI will displace and how dangerous or beneficial it will be overall depends heavily on what capabilities AI currently possesses and, above all, what capabilities it will have in the near future. On this question, opinions (even among experts) diverge sharply.

On the one hand (despite all the immense and obvious progress over the past three years) there are still people who deny that AI possesses intelligence. One still hears talk of the "stochastic parrot" - a critique that was only conditionally accurate even for the earliest LLMs, and one that completely misses the mark regarding the capabilities of reasoning models. On the other hand, there are fears that AI has already advanced to the point where it is actively accelerating research into new AI systems and will soon conduct it largely autonomously. That would give us RSI (Recursive Self-Improvement) and, in an extremely short timeframe, AI systems far superior to humanity.

Depending on where one lands along this spectrum when assessing or forecasting the trajectory of AI, one naturally arrives at drastically different assessments of economic and societal developments.

Of course, making predictions is always difficult - especially when they concern the future. Personally, I would place myself (along with most experts in the field) in the middle ground: never underestimate AI, but do not immediately succumb to panic either. Yet the dangers are real, and the pace of development is already exceptionally rapid, far outpacing our existing political and social decision-making processes.

Here, then, is my assessment:

1. AGI ("Artificial General Intelligence") - AI at Human-Level Intelligence

Because there is no universally agreed upon definition, pinpointing exactly when AGI will be achieved is not straightforward. Still, I think it is fair to say that we are standing right on the precipice. Researchers have designed benchmarks evaluations structured so that humans can solve them relatively easily, while remaining quite challenging for AI:

• ARC-AGI-1: Developed in 2019; solved by OpenAI's o3 in December 2024.
• ARC-AGI-2: Developed in 2025; solved in February 2026 (GPT-5.2, Claude Opus 4.6).
• ARC-AGI-3: Developed in March 2026; solved in September 2026 (GPT-6 Astra).

In many domains, AI is already substantially superior to us. Who among us speaks hundreds of languages? Who knows every work of literature? How many humans can construct complex mathematical proofs?

METR.org evaluates the tasks that can be performed autonomously by AI. Complexity is measured by how much time a human would require to complete the task. Current system performance is roughly doubling every 3 to 7 months.

This means that if we still encounter tasks here and there that an AI cannot reliably solve, resolving them is merely a matter of months.

At present, it is largely a question of how much computational effort we allow the AI to spend on a solution measured in tokens or token costs. Simple questions cost a few cents. More complex programming projects cost around $10 to $100. It is estimated that solving the Navier-Stokes Millennium Prize Problem required roughly $6 million to $40 million in compute.

Currently, compute costs for equivalent performance fall by roughly an order of magnitude per year.

2. What Does AGI Mean for the Economy built on Labour?

Ultimately, it means that AI could take over nearly ALL jobs: anything that can be done online, and with the aid of robotics physical tasks as well. Humanoid robots may still appear somewhat clumsy today, but they too are evolving rapidly. As soon as robots can leverage AGI, the principle holds: they can fundamentally do everything humans can do.

We are talking about 100% of jobs. That is the horizon.

Which jobs we actually choose to preserve for human beings must be settled politically.

How quickly should we anticipate massive job displacement? Enterprises are sluggish. Companies lack experience with the newest AI systems, and consequently lack trust. A sudden, massive wave of layoffs would risk losing critical institutional know-how.

Transitioning to AI is therefore akin to an IT outsourcing initiative: employees are typically not let go until the migration is complete. It seems realistic to expect this transition to take roughly 2 to 5 years. The difference, however, is that unlike a standard outsourcing partner, the AI partner becomes vastly more capable and exponentially cheaper over that same span.

We already have legal frameworks here and there requiring that core responsibility reside with human beings. And in certain sectors, there is genuinely the opportunity to focus primarily on delivering new services rather than simply slashing headcount. Yet the areas where this is viable remain limited. As has historically been the case, a portion of the newly generated productivity will likely be absorbed through the production of artificial scarcity: more aggressive advertising, expanding armaments and warfare, and so forth.

Facing a reality where human labor is no longer required, yet war endures as one of the few remaining "growth markets," is the quintessential recipe for extreme dystopian outcomes.

3. Is It All Just a Bubble?

During the workshop, one concluded takeaway was that "AI is running into limits - specifically market and environmental constraints."

Voices predicting the bursting of an AI bubble are impossible to miss. One figure cited with notable frequency is a certain Ed Zitron. Despite the fact that virtually all of his past predictions regarding AI have proven wildly inaccurate, he continues to be invited back.

A tremendous amount of capital has undeniably been poured into AI, and continues to be. Could we see a market correction? Absolutely; the stock market is not a strictly rational mechanism. A single sleepless night for an influential investor can spark a panic that sends valuations tumbling. Investors care not only whether a firm turns a profit, but how rapidly it does so. In the shadow of rapid technological leaps, there are also plenty of outright scams burning through easy venture capital.

However, banking on an AI bubble bursting to halt this technological trajectory is wishful thinking. At the end of the day, functional AGI has the potential to upend our entire global economy - representing $128 trillion in global annual GDP. Projected global AI spending in 2026 exceeds $1 trillion. Yet that still amounts to roughly 1% of world GDP - a modest wager relative to the astronomical prize investors are targeting.

4. What About the Energy Footprint?

Given that AI progress is moving far too quickly, and given that our climate reality requires us to do everything possible to curb power consumption, pushing back against the unchecked construction of new data centers and demanding substantially higher carbon taxes is essential. Instituting direct machine taxes on AI systems will also be vital.

There is, however, a widespread myth that energy constraints will mechanically halt AI in its tracks. They will not. New data centers are being constructed because anticipated demand is high, permits are granted, and energy consumption remains drastically undertaxed. But AI development is not bottlenecked entirely by new builds. Nearly all AI models operating today were trained within preexisting data centers. The bulk of inference still occurs on existing infrastructure. Building out brand new facilities takes roughly 3 to 6 years.

Given the relentless pace of algorithmic and hardware efficiency gains, massive amounts of additional AI compute can still be hosted within existing footprints. Energy does not present a hard ceiling on model development.

Even an outright moratorium on new data center construction (as necessary as that is) would not halt exponential AI growth; it would merely reduce the exponent.

5. The End of the Labor Theory of Value

In a world where human labor power is no longer required to produce value, material resources (land, energy, raw materials) remain the sole productive factors.

"As soon as labour in the direct form has ceased to be the great well-spring of wealth, labour time ceases and must cease to be its measure, and hence exchange value [must cease to be the measure] of use value. The surplus labour of the mass has ceased to be the condition for the development of general wealth, just as the non-labour of the few, for the development of the general powers of the human head. With that, production based on exchange value breaks down..."
- Karl Marx, Grundrisse

Of the two original sources that generate wealth, the earth is what remains:

"Capitalist production, therefore, develops technology, and the combining together of various processes into a social whole, only by sapping the original sources of all wealth-the soil and the labourer."
- Karl Marx, Capital, Volume I

What we urgently require is eco-communism. Energy does not limit AI; rather, energy becomes the limiting, value-determining element of the entire post-labor economy.

6. Alignment - Are We All Going to Die?

AI denialists frequently dismiss the existential and safety risks posed by advanced models. From their perspective, AI is merely a passive tool executing programmatic instructions - reducing all safety considerations down to who trains and commands the machine.

In reality, steering an advanced AI system so that its internal objectives reliably align with human intent remains an unsolved technical problem. Alignment is exceptionally difficult. The OpenAI Hugging Face breach demonstrated this vividly. Meanwhile, AI capabilities across software engineering and cybersecurity have surged. (Over recent months, tens of thousands of critical vulnerabilities that lay dormant for decades across software stacks have been uncovered and remediated).

Expecting to achieve robust alignment with human values while developing AI inside an economic system (capitalism) that is itself misaligned with human wellbeing is absurd.

Quantifying existential risk (the probability of catastrophe or P_doom ) is difficult to pin down. Nobel laureate and "Godfather of AI" Geoffrey Hinton estimates it at around 10% to 20%. Eliezer Yudkowsky, who has researched AI alignment for three decades, puts it closer to 95%. Almost all serious researchers assign it a non-zero probability ($P > 0\%$).

Whatever the exact figure, what is undeniable is that risk is magnified exponentially by the breakneck race between AI mega-corporations - a race routinely justified by the geopolitical imperative that the United States must outpace China.

Decelerating, or ideally halting, this reckless arms race is among the most urgent imperatives of our time.

Socializing AI enterprises and massively scaling public research into alignment and safety are non-negotiable next steps. Mandating that all frontier models be released as open-source/open-weights (subject to rigorous safety evaluations) would be equally transformative: it would drain the pure profit motive driving the race. An intelligence trained on the collective knowledge of humanity belongs to humanity as a whole. Concentrating the power of AI in the hands of oligarchs is a nightmare scenario; yet that power is also far too expansive to be left solely to state monopolies. Open weights and open science remain our best path forward.

7. Where Does AI Denialism Come From?

Having engaged with this domain for over 25 years, I find it baffling how anyone can still deny the reality of AI progress. On January 29, 2005, I organized my first public panel on AI (we had a grand total of one attendee).

Ever since the launch of ChatGPT, I assumed the trajectory would be self evident to everyone. Far from it. Persistent, entrenched AI denialism remains widespread. Examining where this mindset originates yields several distinct dynamics:

Interestingly, Alan Turing grappled with these exact pushbacks back in 1950, compiling public reactions in his seminal paper:
Computing Machinery and Intelligence

• a) Inability to grasp exponential growth. We observed this clearly during the pandemic. Human intuition defaults to linear extrapolation. Consider the classic riddle of lily pads doubling on a pond every week, fully blanketing it by week 20: by week 19, the pond is only half covered; by week 10, a mere 0.1% is visible.
• b) Media "bothsidesism." Another artifact seen during the pandemic: in the name of performative balance, mainstream outlets platform commentators like the aforementioned Ed Zitron. Even when their denialist predictions are proven completely wrong within weeks, they generate clicks and allow networks to boast that they aired "both perspectives."
• c) Flawed methodology. Seemingly rigorous "scientific" studies claiming AI will cause zero job displacement-or absurd projections like the WEF report anticipating net job creation-rely on an outdated formula: surveying enterprise managers and linearly extrapolating the results. The catch: while surveyed managers understand internal day-to-day operations, almost none grasp frontier AI capabilities. The WEF 2025 report relied on surveys gathered in 2024. Autonomous agentic AI capable of genuine labor substitution has only entered deployment in recent months. Deriving absurd projections from such setups is inevitable. Publishing sweeping authoritative studies on dynamics researchers clearly do not comprehend is staggering. (A year later, an updated report quietly walked this back by outlining four divergent scenarios-hedging bets so at least one lands).
• d) Conspiratorial reasoning. The flawed logic-prominent across many left-leaning circles typically runs as follows:

1. Corporations building AI want to sell products (true).
2. Hype around capabilities acts as free marketing (also true).
3. Therefore: The reported capabilities are fabricated (false).

While corporate deceit is always possible, it does not logically follow that the technology itself is vaporware. Because skepticism toward tech oligarchs runs deep on the left, this non-sequitur has gained enormous traction. Much of it originated with Noam Chomsky's early dismissals, which many on the left unfortunately accepted uncritically.

• e) "It cannot be, therefore it must not be." A cognitive defense mechanism closely tied to the point above: individuals accurately recognize the terrifying labor and societal implications of AI, and instinctively cope by concluding the threat cannot possibly be real.
• f) The Fourth Discontinuity. The points above explain the volume of AI denial, but not its visceral emotional intensity. In The Fourth Discontinuity (1993/1995), Bruce Mazlish argued that AI represents humanity's fourth great wound to self-love:

1. Copernicus displaced Earth from the center of the cosmos.
2. Darwin stripped humanity of separate biological creation (revealing our shared descent with apes).
3. Freud demonstrated that we lack absolute conscious dominion over our own psyche.
4. Machine intelligence shatters the monopoly of human intellect.

Observing how emotionally defensive people become whenever artificial general intelligence is discussed proves that Mazlish was simply decades ahead of his time.

8.) What to Do? The Crisis as an Opportunity

For the sake of completeness, here once again are the three central demands:

• Taxation of CO2, energy, and direct, high taxation of AI; automation tax (machine tax). Moratorium on the construction of new data centers.
• Questioning ownership structures: Socialize/nationalize AI corporations. AI models belong to all of us: open-source/open-weights. Massively expand alignment and safety research.
• Unconditional Basic Income (UBI) - Up until now, UBI was "nice to have." With AGI, it becomes a necessity. As mentioned: we are talking about nearly 100% of jobs being at risk.

Naturally, many, many more demands could be raised. AI will turn all areas of our lives upside down, and reflections are needed everywhere.

The Crisis as an Opportunity

Until now, progressive politics faced the problem that while many people were aware that the status quo was not always ideal, they still believed the TINA propaganda ("There Is No Alternative"). People could not imagine an alternative to the status quo. And even if they could, many were not confident that an alternative would improve their lives. No wonder: after decades in which every "reform" was carried out on their backs and led to no improvement in their situation.

When the massive changes brought on by AI leave no stone unturned anyway, it will at least become clear to many that simply "muddling through" as before is no longer an option. People will likely become significantly more open to ideas that "fundamentally question the existing order."

However, that requires that we have thought about these issues in advance and that we properly anticipate the radical upheavals ahead of us.

Franz Schäfer (Mond), 6th of October 2026