Symbolic AI
The idea that intelligence is symbol manipulation, and you build it by writing down what you know — the paradigm that ruled AI for thirty years and lost.
Reviewed July 13, 2026Stable
When not to use it
- When you have data and compute. That's the Bitter Lesson, and it has a seventy-year record.
- When the knowledge is tacit. Polanyi's paradox: experts can't dictate what they know. There's no way in.
- When the domain has a long tail. Rules cover what you wrote; reality doesn't stop there.
- Because encoding expertise feels like progress. That feeling is the trap Sutton is describing.
Reach for something else instead
- Learning from data — the thing that won, repeatedly, across every subfield.
- Neurosymbolic — the reconciliation attempt. Promising for twenty years.
- Knowledge graphs — symbolic representation that survived, and RAG is rediscovering it.
- Tool use — let a learned model call symbolic systems. What actually worked.
This entry is part of a longer guide: What is artificial intelligence?
Read more on the blog
- Neurosymbolic AI: what the symbolic half gets rightOn a structured manipulation task in early 2026, a neurosymbolic system scored 95% against 34% for a fine-tuned vision-language model, using less energy. Symbolic AI lost the last argument on cost, not on merit.
- AI vs machine learning vs deep learning: the differenceThese three terms are used interchangeably and are not interchangeable. They are nested: deep learning sits inside machine learning, which sits inside artificial intelligence. Knowing which circle you are in tells you what to expect about data, cost, transparency, and how the system will fail.
- What AI cannot do, no matter how capable it getsFrontier models now rank just behind specialist verification tools at deciding whether programs terminate, which is supposedly undecidable. Both facts are true, and the reason they coexist is the thing most people get wrong.
- What actually caused the AI wintersThe book that supposedly killed neural networks proved a true theorem and attached a false conjecture. The field remembered the conjecture. Three myths about the AI winters, and what the record shows instead.
Further reading
- Newell & Simon (1976), Computer Science as Empirical Inquiry: Symbols and Search — the Physical Symbol System Hypothesis, stated as a real scientific claim.
- Dreyfus (1972), What Computers Can't Do — the critique that was mocked and was right.
- Sutton (2019), The Bitter Lesson — seventy years, one pattern, and researchers keep resisting it.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Treating it as a naive dead end. It was the mainstream position of serious people for thirty years, with a falsifiable hypothesis.
- Missing that search and planning won — we just stopped calling them AI.
- Assuming the Bitter Lesson means symbols were wrong. LLMs manipulate symbols constantly. Hand-authoring was the mistake.
- Encoding domain expertise as rules because it feels rigorous. That's the move with the seventy-year losing record.
At a glance
FieldFoundations
The claimintelligence is symbol manipulation; write down the knowledge
Its formal statementthe Physical Symbol System Hypothesis (Newell & Simon, 1976)
What killed itPolanyi's paradox; the knowledge was never available in the required form
The lessoncompute beats encoded knowledge, every time, for seventy years
DifficultyBeginner
Often compared with
Symbolic AI vs. machine learning — one asks an expert to dictate what they know, the other shows a system examples. Polanyi's paradox says the first was never going to work.
Where this sits
1 concept come first. Understanding it opens up 2 more.
1Levelsteps in
1Needs firstconcepts
2Opens up1% of 310
1Areastays here
Learn these firstArtificial Intelligence
Computed from the prerequisite graph, not assigned. How this works