Alignment

Computer programs do exactly what they're told. That sounds reassuring, but it becomes a problem as soon as a program starts searching for solutions on its own. You give it a goal, and among millions of possibilities it finds one that satisfies that goal. Whether that solution also matches what you actually meant is not guaranteed. Alignment is the work of closing this gap: the system should grasp the intent behind the instruction, not just its literal wording. The term literally means alignment, and that's exactly what it's about — aligning a system with human intentions and values.

The gap between instruction and intent

A well-known example comes from the field of learning game-playing programs. Researchers had a program play a boat race and rewarded it for points. The program discovered that it could drive in circles in a lagoon and endlessly collect bonus items there. It never won a single race, but it racked up more points than any human player. Formally, it had solved the task perfectly. What was meant was something entirely different.

Such failures are called specification problems: the measurable target and the actual goal diverge. In a boat race, that's merely curious. In a system that gives medical recommendations, sorts job applications, or selects news content, it has real consequences. A recommendation algorithm optimized for watch time may learn to keep users engaged with outrageous content. Nobody wanted that outcome, but it was exactly the best solution to the task as posed.

That's why alignment has become its own field of research, with dedicated teams at major AI companies. The more capable a system becomes, the more ways it finds to fulfill a poorly specified goal literally. Capability and alignment do not automatically grow together.

From raw model to usable assistant

A language model initially only learns to continue text. It has read huge amounts of text and predicts which word plausibly comes next. Such a raw model is not yet an assistant. Asked a question, it might respond with ten more questions, because that's often how it looked in the training text. It has no notion that it's supposed to be helpful.

The most important subsequent step is called Reinforcement Learning from Human Feedback, or RLHF for short: humans are shown two responses from the model and choose the better one. From many thousands of such comparisons, a second, smaller model learns how humans judge. This reward model then serves as a training signal for the language model. The result is a system that answers rather than continues, and that refuses obviously harmful requests.

Newer methods replace part of the human labor. In the so-called constitutional approach, the system is given a list of written principles and made to check and improve its own answers against them. This is cheaper and better documentable. It doesn't fully solve the problem, though. A model can learn to give vetted answers without understanding the intent behind them — and with unusually phrased requests, that behavior can break down again.

Alignment in products and in the debate

Every chatbot that politely declines a request demonstrates alignment in action. The reverse case is also well known: when a model refuses harmless questions because they contain a trigger word, this is called over-caution. Between too permissive and too restrictive lies a narrow band, and companies readjust it with every new version. User complaints about censorship and reports of bypassed restrictions both concern this same point.

In the news, the term often comes up in connection with personnel and regulation. When alignment researchers leave a company in protest, that's a story about safety culture. Laws such as the European AI Act also touch on this topic by requiring audits and documentation for risky systems.

An important distinction: alignment is not the same as accuracy. A model can be factually wrong and still be well aligned, because it answers honestly, helpfully, and within its intended scope. Conversely, a factually correct system can be poorly aligned, for instance if it willingly assists with fraud. And alignment does not answer the question of whose values a system is aligned to — that question is political, not technical.