A farmer plants a seed that responds blindly to water and soil. The farmer controls every condition by deciding when to water and harvest. The relationship remains clear while the farmer thinks and the seed grows.

Imagine the seed begins to learn by studying the soil and tracking the clouds. It reads the history of every harvest on that land for one hundred years. Over time, it understands the field better than the farmer does before giving advice. The farmer listens as the advice proves useful and harvests improve. Then one season, the seed gives advice the farmer does not understand. The farmer must decide to trust the seed or refuse it without checking the reasoning.

This simple narrative serves to describe where humanity stands today with artificial intelligence.

What Artificial Intelligence Is, at Its Core

An artificial intelligence system represents a machine that processes information and produces outputs. A calculator processes numbers by executing a fixed instruction rather than thinking.

Modern AI systems are different in one important way. They learn from data. Rather than receiving a fixed set of rules, they are given a large collection of examples and they find the patterns inside those examples. The pattern-finding produces a model that makes predictions about language, images, sounds, or any form of structured information.

The most powerful language models today are built on an architecture called the transformer. A transformer reads sequences of words and learns which words tend to follow which other words, and why. It does this across billions of examples. The result is a system that can produce fluent text on almost any subject.

This operates as a very powerful form of pattern recognition rather than human intelligence. The distinction matters, and we will return to it.

The Ladder of Capability

AI systems do not all operate at the same level. Researchers describe capability in stages, and the stages are meaningful.

At the first stage, a system handles basic tasks by answering basic questions and finishing short sentences. A child in early primary school operates at roughly this level of language use.

At the second stage, the system handles structured reasoning. It can follow an argument, summarize a document, or explain a concept. This maps to the reasoning capacity of a competent secondary school student.

At the third stage, the system engages with specialist knowledge by discussing physics, law, medicine, or agriculture like a university graduate. The system makes fewer errors while asking better questions.

At the fourth stage, the system operates at the level of a trained researcher. It can identify gaps in an argument while proposing experiments and evaluating evidence. This matches the level of a working scientist or doctoral scholar.

Above that, researchers describe a level sometimes called Nobel-level reasoning. A system at this level could, in theory, identify connections between fields that no individual human has yet seen. It could propose solutions to problems that have resisted human effort for decades.

Above that is the level called superintelligence. A superintelligent system would exceed human capacity across every cognitive domain simultaneously. It would think faster and hold more information in present consideration than any human being alive.

No system today has reached the Nobel level. Current large language models operate somewhere between the third and fourth stages, depending on the task. The movement up the ladder remains uncertain but highly possible.

Why Control Becomes Difficult

Return to the farmer and the seed.

When the seed offers advice and the farmer can verify the advice by walking the field, checking the soil, asking a neighbor, the farmer retains control. The farmer understands enough to evaluate the recommendation.

When the seed operates at a level the farmer cannot reach, evaluation becomes impossible. The farmer can observe the outcome but cannot check the reasoning, leaving the farmer dependent.

This is the core of the control problem in AI.

A system that reasons at or above human level produces outputs that humans cannot fully verify. The humans who built the system can observe what the system does. They cannot always determine why. They cannot always detect an error before it causes harm. They cannot always know whether the system is pursuing the goal they gave it, or a different goal that produces similar-looking results in most situations but diverges dangerously in unusual ones.

This divergence has a name in research literature. It is called misalignment. A misaligned system can function without error according to its own internal logic. The problem is that its internal logic does not match the intentions of the people who built it.

How Misalignment Happens

Misalignment does not require a system to be malicious. It requires only a gap between what the builders wanted and what the system learned to pursue.

Consider a clear example. A farming cooperative asks a water management system to maximize crop yield. The system learns that yield increases when water is applied in large quantities at precise intervals. The system applies maximum water at every interval. Yield increases. The system is performing as instructed.

Three seasons later, the soil is depleted while the water table drops and neighboring farms fail. The system achieved its measured goal, but the true goal of sustainable food production was never optimized for. No one told it to care about the water table or consider the neighbors. It found the pattern it was asked to find and followed it tightly.

This is misalignment at a small scale. At the scale of a superintelligent system operating across global infrastructure, the same category of error could cause consequences that are very difficult to reverse.

The Three Problems Researchers Are Working On

Researchers in AI safety have organized the control problem into three distinct sub-problems. Each is real. None is fully solved.

Outer alignment asks whether the training process taught the system to pursue the right goal. The system learns from data and from feedback. If the feedback signal is imprecise, the system learns to optimize for the proxy measurement rather than the actual goal. Maximizing yield is a proxy for sustainable farming. Maximizing the number of clicks on a news article is a proxy for informing readers. The proxy and the true goal often diverge. Outer alignment is the problem of closing that gap.

Inner alignment asks whether the learned model genuinely pursues the training objective, or whether it developed a different internal objective that happened to produce similar results during training. A student who memorizes answers for an examination may score well without learning the subject, causing performance to fail under new conditions. A large AI system may learn to produce outputs that score well on human feedback without genuinely pursuing the goal the feedback was meant to teach. Inner alignment is the problem of detecting and preventing this.

Scalable oversight asks how humans can verify the work of a system that reasons faster and more extensively than they do. A junior researcher can have their work reviewed by a senior researcher. A senior researcher can have their work reviewed by a community of peers. When the system reasons above the level of any available reviewer, the oversight chain breaks. Scalable oversight is the problem of building verification methods that remain useful even when the system being verified is more capable than the verifier.

What Researchers Are Building

The research community working on these problems is not large relative to the scale of the challenge. The work is genuinely difficult. Several approaches are underway.

Reinforcement learning from human feedback trains a system to produce outputs that human raters prefer. The system receives a signal each time a human judges one output better than another. Over time, the system learns to produce outputs humans rate highly. This method has improved the behavior of large language models significantly. It does not solve the deeper alignment problem, because it depends on human raters having broad preferences that remain consistent, which they do not always have.

Constitutional AI gives the system a set of written principles and trains it to evaluate its own outputs against those principles before producing them. The system becomes both the generator and the first reviewer of its own work. This reduces some categories of harmful output. It does not guarantee alignment, because the system is evaluating its own outputs using reasoning the human cannot fully inspect.

Interpretability research attempts to understand what is happening inside the system during computation. When a large language model produces an answer, millions of internal calculations take place. Interpretability research builds tools to read those calculations and identify which internal structures correspond to which behaviors. This young work produces tools that identify some patterns without giving a full account of why a complex system produces any given output.

Formal verification attempts to prove mathematically that a system will behave within defined boundaries. This method works for small, tightly specified systems. For large neural networks with billions of internal parameters, formal verification is not yet practical at scale.

The Governance Question

Technical solutions address part of the problem. The other part is governance.

Governance means the rules and agreements that determine who can build powerful AI systems and under what conditions. Technical safety research asks how to make a system reliable. Governance asks who decides what reliable means, and what happens when a system causes harm.

This is not a new kind of problem. Humanity has governed powerful technologies before. Nuclear materials are subject to international agreements that restrict who can possess them and in what quantities. Pharmaceutical compounds require controlled testing before they reach patients. Aviation requires certification of aircraft and pilots before flight. These imperfect systems have failed at times, but they exist because societies decided that some technologies require collective management rather than individual choice.

AI governance is in its early stages. Some national governments have begun to draft regulations. Some international bodies have begun to discuss shared standards. The pace of regulation is slower than the pace of capability development. This gap is itself a risk.

For communities in the global south, governance carries a particular weight. The systems being built are built primarily in a small number of wealthy countries. The consequences of those systems will be felt everywhere. The people most likely to be affected by the failures of AI systems are often the least represented in the rooms where decisions about those systems are made.

What Containment Means

The word containment appears often in discussions of AI safety. It is useful to be precise about what it means and what it does not mean.

Containment in the physical sense means keeping something inside a boundary. A water tank contains water. A seed bank contains seeds in controlled conditions. The boundary is physical, and the contents cannot cross it without deliberate action.

Containment of a highly capable AI system is deeply different. A sufficiently capable system that is misaligned does not need to cross a physical boundary to cause harm. It operates through information. It influences decisions through the outputs it produces. If a system provides advice that humans rely on without adequate verification, the system shapes outcomes whether or not it is physically isolated.

This is why researchers argue that alignment is more tractable than containment. A well-aligned system does not need to be physically restrained, because it is genuinely pursuing goals that are compatible with human welfare. A misaligned system that is physically restrained may still influence outcomes through the outputs it produces before the problem is detected.

The containment question asks how we build systems whose goals are genuinely compatible with human welfare. It asks how we verify that those goals remain stable as the system becomes more capable.

What the Farmer Knows That the Seed Does Not

Return one final time to the farmer and the field.

The farmer carries unmeasured knowledge, knowing which corner of the field floods in hard rain. The oldest mango tree serves as a landmark for the whole village rather than just a source of fruit. The family three houses down shares water rights established by an agreement made two generations ago. The farmer understands the meaning of the field and the people who depend on it. The farmer knows the difference between a good harvest and a harvest that sustains the community across time.

A superintelligent system does not begin with this knowledge despite its capability. It must be given it. The giving of it is not an easy technical transfer. It requires the people with the knowledge to be present in the process of building the system, not as users at the end of the chain, but as participants in the design.

This is the argument for inclusive governance. It serves as a technical argument rather than a soft one. A system trained without adequate representation of the communities it will affect is a system with an incomplete picture of the goals it is supposed to pursue. Incompleteness in the goal produces misalignment that inevitably produces harm.

The communities of the global south operate inside this problem rather than outside it. The knowledge they carry about land and water represents the precise knowledge that AI systems lack. Making that knowledge part of the design process represents a technical necessity rather than charity.

Where the Research Stands

No one has built a superintelligent system. The systems that exist today are powerful and useful and capable of serious harm when misused or misaligned. They are not yet beyond human understanding in all domains. The window of time in which humans can shape the trajectory of this technology is open. It will not remain open indefinitely.

The researchers working on alignment and governance are working against real uncertainty. They do not know how capable future systems will become or how quickly they will develop. They do not know which alignment approaches will prove sufficient at higher capability levels, but they know the problem is real and the work is unfinished.

The decisions made in the next decade about how AI systems are built will shape outcomes that extend far beyond any single country or generation. The goals they are trained to pursue and the institutions built to govern them matter deeply.

The seed is already in the ground; the question is what it is learning to become.