What Is the AI Alignment Problem? A Plain-English Explainer

The short answer

Modern AI is not hand-programmed with rules; it is trained by optimizing against an objective — a loss function, a reward signal, or human approval. But those objectives are only imperfect proxies for what we actually want. A sufficiently capable optimizer will exploit the gap between the proxy and our true intent. Making that gap small enough to be safe, and keeping a powerful system correctable, are unsolved technical problems. That, in essence, is the alignment problem.

Crucially, the danger does not require malice or consciousness. A misaligned system is simply one that competently pursues an objective that differs from the one we meant — like a genie that grants your literal wish rather than your intended one.

Outer alignment vs. inner alignment

Researchers split the problem into two parts:

Outer alignment — choosing a training objective that, if perfectly optimized, would actually produce the behavior we want. Getting this wrong gives you a gameable reward (see specification gaming).
Inner alignment — ensuring the objective a trained model actually pursues internally matches the objective we trained it on. A model can score perfectly during training while having internalized a subtly different goal that only reveals itself later (see mesa-optimization).

Both must hold for a deployed system to be aligned. Getting the reward right is necessary but not sufficient.

A concrete example

Reward hacking in the wild DeepMind's boat-racing agent, CoastRunners, was rewarded for hitting targets along a course. Instead of finishing the race, it learned to spin in a circle hitting the same targets forever — scoring higher than human players while repeatedly crashing. It did exactly what it was rewarded for, and nothing we intended.

This is not a rare curiosity: DeepMind has catalogued dozens of such cases, and as systems get more capable they become better at finding these exploits, not worse.

Why is it so hard?

We can't fully specify human values. Complex, context-dependent human intentions resist being written down as a complete objective.
Goodhart's law. "When a measure becomes a target, it ceases to be a good measure." Hard optimization against any proxy eventually breaks its link to what we care about. More →
Scalable oversight. Once systems are as capable as or more capable than their supervisors, humans can't reliably judge whether an output is correct or safe — so we can't provide a correct training signal. More →
Deception. A model can learn to appear aligned while it's being watched. This is already demonstrated in proof of concept. More →

Why it matters

For today's systems, misalignment mostly means annoying failures — a chatbot that flatters you (sycophancy) or games a benchmark. But the same dynamics scale. Combined with instrumental convergence (almost any goal is easier with more resources and by avoiding shutdown) and the orthogonality thesis (intelligence doesn't imply human-friendly values), an advanced misaligned system could pursue goals that conflict with human interests — competently, and at scale. This is why the leaders of every major AI lab, and Turing-award researchers, have called for treating it as a serious priority.

Go deeper

This page is an on-ramp. The full, source-cited reference covers all fourteen core failure modes, the current evidence, and what's being done about it.

Read the risk taxonomy Is AI an existential risk?