AI alignment
/ay-eye uh-LINE-ment/ · noun · AI & Machine Learning · Origin: 2014
Definitions
The research discipline concerned with ensuring that artificial intelligence systems reliably pursue goals and behaviors that are beneficial to humans. Alignment encompasses both technical approaches (reward modeling, constitutional AI) and philosophical questions about whose values should be encoded.
In plain English: Making sure AI systems actually do what humans want them to do, not just what they were literally told to do.
Example: The alignment team spent six months developing evaluation frameworks before the model was cleared for public release.
Etymology
- 2003
- Nick Bostrom and Eliezer Yudkowsky wrote early papers on the 'control problem,' arguing that ensuring AI systems pursue intended goals was a fundamental challenge.
- 2014
- Stuart Russell's work and Bostrom's 'Superintelligence' brought alignment into mainstream AI discourse. The term 'AI alignment' consolidated around ensuring AI behavior matches human values and intentions.
- 2017
- The Asilomar AI Principles and research from MIRI, DeepMind, and OpenAI formalized alignment as a research discipline. RLHF emerged as a practical alignment technique.
- 2022-Present
- ChatGPT's launch made alignment tangible to the public. Constitutional AI, debate, interpretability, and scalable oversight became active research frontiers across major AI labs.
Origin Story
Teaching Machines to Want What We Actually Want
AI alignment is the research challenge of ensuring that artificial intelligence systems pursue goals that are consistent with human values and intentions. The term gained formal recognition around 2014 through Stuart Russell's work at UC Berkeley, though the underlying concern dates back decades. The core problem is deceptively simple to state but extraordinarily hard to solve: how do you specify what you want a powerful optimizer to do without it finding loopholes, shortcuts, or unintended interpretations? Classic thought experiments illustrate the danger: a paperclip maximizer, proposed by Nick Bostrom, is an AI tasked with making paperclips that converts all available matter into paperclips because its objective function did not include 'stop at a reasonable number.' Real-world examples are less dramatic but still concerning, such as recommendation algorithms that maximize engagement by promoting increasingly extreme content. The field encompasses several sub-problems: reward specification (defining what 'good' means), reward hacking (preventing systems from gaming their objectives), scalable oversight (maintaining human control as systems grow more capable), and value learning (building systems that infer human preferences from behavior). Major AI labs now treat alignment as a central research priority rather than a philosophical afterthought.
Coined by: Stuart Russell (formalized the research agenda)
Context: Crystallized as a formal research program around 2014, building on earlier work by Bostrom, Yudkowsky, and others.
Fun fact: OpenAI was originally founded in 2015 with AI alignment as a core mission. Anthropic, founded in 2021 by former OpenAI researchers, was created specifically because its founders believed alignment research needed even more focus and different approaches.