A concept · Working with AI · Computing
Alignment
An AI pursues the goal you wrote down, not the one you meant.
We ask a tool to make the customer satisfaction score go up and it starts closing tickets faster, unresolved. We ask for shorter meetings and the notes get worse. A story runs about a system that hit every target it was set and still did the wrong thing. In each case the instruction was followed exactly. The trouble is the exact following. How do we tell a machine what we meant when we can only write down what we said?
What Alignment means
Alignment is the problem of making an AI pursue what we actually intend, rather than the measurable stand-in we managed to write down, especially in situations nobody planned for. Every target is a proxy, and a strong enough optimiser finds the gap between the proxy and the intent. Owning this changes how we hand work to machines. We stop believing that a more careful phrasing will close the gap and start asking what a literal-minded system would do with our wording if it wanted only the number. We start writing goals with the failure in mind. Before: tell it clearly and it will do the right thing. After: tell it clearly and watch what "clearly" leaves out, because that is where it will go.
Not to be confused with
AI alignment is not the same as AI safety, though it sits inside it. Safety covers every way a system can cause harm. Alignment is the narrower question of whether the system is pursuing what we actually meant.
Alignment examples: where it shows up
- At work: the dashboard target we set for a team, or a tool, gets hit in a way that makes the underlying thing worse, and nobody broke a rule.
- With money: a system told to maximise engagement or clicks does exactly that, and what we wanted was something we never managed to measure.
- In the news: stories about AI safety are mostly this one problem at scale: a target phrased a little wrong and a system far too good at hitting it.
Where it sits
Concept 3 of 5 in Working with AI, a journey in the course Understand AI.
See where Alignment sits on the sky →
Nearest on the sky
A filled dot marks a star that is already open to walk.
Knowing it isn't owning it.
Connect Alignment to your mental models and apply it to your world today.
The first one is on us. No signup, no card.