AI 2040: Plan A

What do we do when artificial intelligence becomes more capable faster than societies can develop rules, safety mechanisms and shared decisions?

AI 2040: Plan A starts at precisely this point. Published in July 2026, the AI Futures Project initiative offers a constructive counterpoint to the earlier scenario AI 2027. There, the authors described a very rapid race towards superhuman AI, with the risk of losing control or concentrating power to an extreme degree. Plan A now asks: What might a different path look like?

An important distinction: the authors explicitly describe Plan A as a recommendation and scenario – not a prediction. The specific sequence is intended to make their policy proposals visible and open to criticism. The authors themselves consider some assumptions relatively robust and others highly uncertain.

Their starting point is nevertheless demanding: they consider it plausible that AI research could become largely automated within a relatively short time, followed by systems that outperform humans across many economically relevant tasks. Whether this development will actually happen so quickly remains open. This is precisely why it is worth reading the text less as a forecast and more as a stress test of our ability to shape the future .

1. A race can drive reasonable actors towards unreasonable outcomes

At the heart of Plan A is a strategic problem before a technical one. If the USA slows down, China could catch up. If one company invests more time in safety, a competitor could release the more powerful system first. When no one trusts anyone else, acceleration seems understandable to each individual.

This is precisely where the authors see the danger: What appears rational locally can become risky globally. A race rewards speed, even though all participants could benefit in the long term from more time, better testing and more reliable rules.

Plan A therefore proposes an international agreement, initially mainly between the USA and China. Its core is not trust but verifiability: computing capacity would be recorded, relevant data centres made inspectable and agreements verifiable through both technical and institutional means.

From the Future Skills Institut's perspective, this is an interesting shift in viewpoint. Agency does not simply mean that an individual responds faster. Systems, too, must be able to recognise shared interests and organise collective self-regulationwhen individual incentives point in another direction.

2. Time is not a delay to progress – it is part of shaping it

The name AI 2040 describes the scenario's central idea. Without intervention, the assumed trajectory could allow very rapid further acceleration around 2030. International agreements stretch this process out in the scenario: between 2030 and 2035, capabilities increase within the human range; in 2035, a pause follows at approximately the level of highly capable human experts; scaling further towards superintelligence would resume only in 2040.

These dates are not reliable predictions. The authors themselves highlight major uncertainties about timelines, development speed, alignment and verification.

What matters is therefore less the year 2040 than the underlying principle: Pace is a variable we can shape. Technological development need not be seen only as something that should advance as quickly as possible. Speed can itself create costs when safety research, institutions, labour markets and societal adaptation cannot keep up.

This connects Plan A directly with Future Skills. Personal development, too, rarely succeeds by squeezing as much change as possible into the shortest possible time. People need direction, experimentation, feedback and opportunities to adjust. At the societal level, Plan A asks a similar question: How much time do we need to do more than react to a development and instead help shape it responsibly?

3. Transparency should prevent knowledge and power from remaining in a few hands

One particularly far-reaching proposal is known as Total Research Transparency. Almost all algorithms, experiments, training runs and frontier research findings would become transparent to other researchers and the public. Particularly sensitive information such as model weights would, however, remain protected.

The idea pursues two goals. First, to make an international agreement verifiable. Second, to prevent a small number of companies or governments from being the only ones who understand how the most powerful systems work, thereby gaining disproportionate power.

The authors do not hide the trade-offs. Greater transparency can also spread capability-enhancing findings faster. The proposed verification is not yet technically mature. And even an agreement could be broken. Plan A therefore discusses unusually drastic safeguards – including a mutual ability to render new large-scale computing capacity unusable in the event of a breach.

Here in particular, the text should not be hastily read as a concrete policy blueprint. Many details are speculative and controversial. Its value lies more in making trade-offs visible: openness versus safety, innovation versus control, national interests versus global risks, speed versus time to learn.

4. Even safe AI development does not automatically answer the question of a good life

Plan A focuses strongly on how very powerful AI could be achieved safely in technical and geopolitical terms. Yet in the scenario, AI significantly changes work and the economy before that point. The authors expect many knowledge-intensive activities increasingly to involve directing AI agents, and human labour potentially to lose much of its economic importance in the long term.

For the Future Skills Institut, this raises a second question that Plan A itself answers only in part: What does a technologically ‘successful’ future actually mean for people's lives?

If AI creates prosperity but paid work provides less support, income and participation need to be reorganised. If people hand over tasks, they also need the judgement to decide what they want to delegate. And if work provides less structure, recognition and meaning, other areas of life become more important.

Our research strands complement one another here: JobRadar 2040 asks where new human tasks and responsibilities might emerge. Lebensfähig 2040 asks what sustains a good life when paid work no longer fulfils this function alone.

Perhaps the most important future skill therefore sits one level higher: not just being able to work with AI, but remaining able, under new technological conditions, to decide what we actually want as people and as a society.

What we take from this

AI 2040: Plan A is not a neutral research report. The authors have clear assumptions about very rapid AI development and the considerable risks of highly capable systems. Those who do not share these assumptions will also assess many of their recommendations differently.

That is precisely why the scenario is valuable: it makes its premises explicit and attempts to translate an abstract demand – ‘AI must be developed safely’ – into concrete decisions, institutions and trade-offs.

For us, one thought above all remains:

Agency requires us to distinguish between technological capability and societal intent.

Not everything that becomes technically possible must be implemented immediately. Nor is every delay automatically hostility to progress. Sometimes time may be exactly what people and institutions need to understand options, weigh consequences and make deliberate decisions.

The decisive question is therefore not just: How capable will AI become?

But: Will we remain capable enough to shape its development together?

If we could deliberately trade speed for greater safety, understanding and societal control in highly capable AI: at what point would slower actually be better for us?

Context: This 3-Minute Insight summarises selected statements and proposals from AI 2040: Plan A in our own words and places them in the context of the Future Skills Institut's research questions. The AI Futures Project explicitly describes Plan A as a recommendation and scenario, not its most likely forecast of the future. Many central assumptions – particularly about AI timelines, the speed of a possible ‘takeoff’, alignment, verification and economic consequences – are uncertain and are described by the authors themselves as having varying degrees of robustness. The connections to human agency, JobRadar 2040 and Lebensfähig 2040 are FSI interpretations.

Source: Thomas Larsen, Romeo Dean, Brendan Halstead, Eli Lifland, Ryan Greenblatt & Daniel Kokotajlo (2026): AI 2040: Plan A, AI Futures Project, published on 9 July 2026. Further information: Official scenario ↗ · Assumptions & uncertainties ↗ · Verification Plan ↗ · Transparency Plan ↗