A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence

By 1955 the ingredients for artificial intelligence already existed, but nobody had mixed them into a single recipe. Alan Turing had asked "Can machines think?" in 1950 and proposed his famous imitation game. Claude Shannon had built a maze-solving mouse from telephone relays. Warren McCulloch and Walter Pitts had modeled neurons as logic gates. Arthur Samuel at IBM was teaching a computer to play checkers — one of the first programs that genuinely learned from experience.

Yet these efforts lived under separate labels — cybernetics, automata theory, information processing — and their practitioners rarely talked to one another. The field that would unify them did not exist, because it did not yet have a name.

The proposal was written on August 31, 1955 by four scientists, each bringing a unique lens to the problem of machine intelligence:

John McCarthy (Dartmouth) — coined the term "artificial intelligence" and proposed studying how machines might use language for reasoning. He envisioned an artificial language rich enough for conjecture and self-reference — ideas that prefigured modern prompt engineering and chain-of-thought reasoning.

Marvin Minsky (Harvard) — focused on how machines could build internal models of their environment. He imagined a system that first explores solutions inside an abstract model before acting in the real world — a description that reads like a blueprint for model-based reinforcement learning.

Nathaniel Rochester (IBM) — brought the engineer's perspective. He had co-designed the IBM 701 and was already simulating neural networks on it. His contribution focused on originality in machine performance — how to introduce guided randomness so that machines could be creative rather than merely obedient.

Claude Shannon (Bell Labs) — the father of information theory. He proposed extending his work on noisy channels and reliable computation to the problem of intelligent machines, and explored the relationship between brain models and their environments.

At the heart of the proposal lies a single, audacious sentence:

"The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it."

This is not a theorem — the authors call it a conjecture. They are not claiming to know how to build a thinking machine. They are claiming something subtler and more powerful: that thinking has a describable structure, and if it does, then in principle a machine can reproduce it. This philosophical bet is still the foundation of AI research today.

The proposal organized the AI problem into seven areas. Each reads like a prescient preview of a modern subfield. Below is each pillar as the authors described it, and what it became in the 70 years that followed.

1. Automatic Computers — The authors noted that the main obstacle was not hardware but programming: "our inability to write programs taking full advantage of what we have." This insight — that software, not silicon, is the bottleneck — remains true in the era of trillion-parameter models.

2. Language — McCarthy speculated that "a large part of human thought consists of manipulating words according to rules of reasoning and rules of conjecture." He envisioned machines that use language for conjecture, self-reference, and argument — anticipating modern large language models by half a century.

3. Neuron Nets — The proposal asked how hypothetical neurons could be arranged to "form concepts." This strand led directly to the neural network revolution: from the Perceptron (1958) to deep learning (2012) to today's foundation models.

4. Theory of the Size of a Calculation — How do you measure the efficiency of a solution? The authors called for a theory of computational complexity — a field that would later produce the P vs NP problem, one of the great open questions in mathematics.

5. Self-Improvement — "Probably a truly intelligent machine will carry out activities which may best be described as self-improvement." Today's reinforcement learning from human feedback, self-play, and recursive self-improvement all descend from this idea.

6. Abstraction — How does a machine form high-level concepts from raw data? This question drives representation learning, from early feature engineering to modern embedding spaces and latent representations.

7. Randomness and Creativity — Rochester proposed that creativity might be "the injection of some randomness" guided by intuition. Modern stochastic sampling — temperature, top-k, top-p — echoes this idea in every text generated by a language model.

Rochester's section of the proposal is particularly visionary. He argues that human creativity is not pure logic — it requires a departure from "traditional methods." But pure randomness would be chaos. Instead, he proposes guided randomness: a process that introduces controlled unpredictability, steered by something like intuition.

He draws on Kenneth Craik's model of the brain: the mind builds "little engines" — internal simulations — that predict what will happen before the body acts. When standard predictions fail, the mind introduces randomness to find novel solutions. This is a remarkably modern idea.

Today, when a language model samples its next token with a temperature of 0.7, it is doing exactly what Rochester described: injecting controlled randomness into an otherwise deterministic process to produce outputs that are novel but still coherent.

Minsky proposed that intelligent machines must build internal abstract models of their environments. A machine placed in a new environment would first construct a simplified representation, then explore solutions within that model before trying them in reality. As he wrote: "Because of this preliminary internal study, these external experiments would appear to be rather clever, and the behavior would have to be regarded as rather imaginative."

This is exactly how modern AI systems work. AlphaGo builds a neural model of the Go board and searches millions of moves within that model before placing a single stone. A planning agent simulates outcomes before choosing an action. Even a language model can be seen as maintaining a latent "world model" — learned from vast text — that it queries to answer questions.

McCarthy identified four properties of English that no formal language of the time could match:

• Conciseness — arguments in English can be brief and powerful.
• Universality — English can set up and use specialized sublanguages within itself.
• Self-reference — a speaker of English can refer to their own reasoning process.
• Rules of conjecture — beyond deduction, English supports guessing, speculation, and hypothesis formation.

He proposed building an artificial language with these properties so that a machine could "learn to play games well and do other tasks." Today's large language models — trained on vast corpora of English and other languages — are that artificial language. They can be concise, self-referential, and they reason by conjecture every time they generate a chain-of-thought.

The workshop ran for eight weeks at Dartmouth College, funded by a \$7,500 grant from the Rockefeller Foundation (about \\$85,000 in 2024 dollars). Rather than a fixed group of ten, it operated as a revolving seminar: participants came and went, some staying days, others weeks.

The workshop did not produce the "significant advance" its proposers hoped for — no general theory of intelligence emerged that summer. But it accomplished something more lasting: it defined the field. The attendees left with a shared vocabulary, shared ambitions, and shared frustrations that would drive research for decades. Allen Newell and Herbert Simon presented the Logic Theorist — arguably the first AI program. Arthur Samuel continued his checkers work, which became a landmark in machine learning.

The most remarkable thing about the Dartmouth proposal is how right it was. Every one of its seven pillars became a major research area. Its core conjecture — that intelligence is simulatable — remains the foundational assumption of the entire field. And its insistence on bringing diverse minds together to tackle the problem collaboratively set the pattern for AI research culture.

The proposal also reveals the founders' humility. They asked for one summer and \$7,500. They used the word "conjecture," not "theorem." They did not promise a thinking machine — they promised to try. The field they launched has not yet fulfilled their grandest ambitions. But it has transformed the world in ways none of them could have imagined.