AI Research in 1950–1980: The Three Decades That Founded the Field
Between 1950 and 1980, a handful of researchers with room-sized computers invented artificial intelligence. Turing asked whether machines can think, and a 1956 summer workshop gave the field its name. Programs proved theorems, learned checkers and found shortest paths. The first neural networks learned from examples, until a famous book showed their limits. There was a chatbot, a robot that planned its own actions, and expert systems that rivalled specialists. Then came the first AI winter. This guide covers twenty papers and milestones in eight themes, each explained in plain English with an animation, followed by what it found, why it matters and its sources.
01 Prizes for the Pioneers
| Prize | For work covered here |
|---|---|
| Turing Award 1969 | Marvin Minsky — "for his central role in creating, shaping, promoting, and advancing the field of Artificial Intelligence" |
| Turing Award 1971 | John McCarthy — artificial intelligence (Dartmouth, LISP) |
| Turing Award 1975 | Allen Newell and Herbert Simon — contributions to AI, the psychology of human cognition and list processing |
| Nobel Prize in Physiology or Medicine 1981 | David Hubel and Torsten Wiesel (shared with Roger Sperry) — information processing in the visual system |
| Turing Award 1994 | Edward Feigenbaum and Raj Reddy — large-scale AI systems (DENDRAL and expert systems for Feigenbaum) |
The Turing Award itself is named after Alan Turing, whose 1950 paper opens this story.
02 Paper Map
Twenty papers and milestones in seven themes, plus what survived. Filter and click to jump:
03 🌅 The Birth of a Field
A question, a chess plan and a summer workshop that gave AI its name.
Computing Machinery and Intelligence
Turing (Manchester) · Mind 1950Instead of arguing about what "thinking" means, Turing proposed a test: if a judge chatting by text cannot tell a machine from a person, we should credit the machine with intelligence. He also predicted learning machines — a "child machine" taught rather than programmed.
- Proposed the imitation game (now called the Turing test) as a practical stand-in for "can machines think?"
- Answered nine objections, from theology to "machines can only do what we tell them"
- Suggested building a child machine and educating it — an early vision of machine learning
Set AI's founding question and its behavioural yardstick — debated again with every chatbot, up to today's LLMs.
Programming a Computer for Playing Chess
Shannon (Bell Labs) · Philosophical Magazine 1950Look ahead a few moves, score the positions you could reach, and assume your opponent always picks the move that is worst for you. Shannon laid out this minimax plan for chess — and showed why you cannot search to the end: there are far too many games.
- Described minimax search with an evaluation function, and full-width ("type A") versus selective ("type B") search
- Estimated the number of possible chess games at around 10120 (now called the Shannon number)
The blueprint for game-playing programs through Deep Blue (1997); AlphaZero still searches a tree, guided by learned evaluations.
A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence
McCarthy, Minsky, Rochester & Shannon · proposal dated 31 Aug 1955; workshop summer 1956Four researchers asked for money to spend a summer on the idea that every aspect of learning or intelligence can be described precisely enough for a machine to simulate it. The proposal coined the phrase "artificial intelligence", and the 1956 workshop is treated as the founding of the field.
- Proposed a two-month, ten-person study at Dartmouth College in summer 1956
- Listed the problems to attack: automatic computers, programming a computer to use language, neuron nets, the size of a calculation, self-improvement, abstractions, and randomness and creativity
- Organisers Minsky (1969) and McCarthy (1971) later won Turing Awards
Named the field and set an agenda — language, neural nets, self-improvement — that reads like a list of today's research areas.
04 🔎 Search and Problem Solving
Programs that prove theorems, learn checkers and find shortest paths by searching cleverly.
The Logic Theory Machine → General Problem Solver
Newell & Simon (RAND, Carnegie Tech), with Shaw · IRE Trans. Information Theory 1956; GPS 1959A program that proves theorems in logic the way a person might: start from what you want to prove, try rules that could get you there, and use rules of thumb (heuristics) to avoid trying everything. Its successor, GPS, aimed to solve any problem by shrinking the gap between where you are and where you want to be.
- The Logic Theorist proved theorems from chapter 2 of Whitehead and Russell's Principia Mathematica, often reported as 38 of the first 52
- Introduced heuristic search and list processing; GPS introduced means-ends analysis
- Newell and Simon shared the 1975 Turing Award for contributions to AI, the psychology of human cognition and list processing
Established AI as search through a space of possibilities, guided by heuristics — the frame for decades of AI and, in modern form, for LLM reasoning with search.
Some Studies in Machine Learning Using the Game of Checkers
Samuel (IBM) · IBM Journal of R&D 1959Instead of hand-tuning a checkers player, Samuel let the program adjust its own scoring rules from experience — including by playing against itself — until it beat him. The paper helped popularise the term "machine learning".
- A checkers program that learned by rote (remembering positions) and by generalisation (tuning the weights of a scoring polynomial)
- Adjusted its evaluation so earlier positions' scores agreed with scores found by searching deeper — an early form of temporal-difference learning
- Learned to play better than its programmer
The first widely known self-improving program; its update idea resurfaced in TD learning and TD-Gammon.
A Formal Basis for the Heuristic Determination of Minimum Cost Paths (A*)
Hart, Nilsson & Raphael (SRI) · IEEE Trans. Systems Science and Cybernetics 1968To find the shortest route, explore first where distance so far + estimated distance to go is smallest. As long as the estimate never overestimates, you are guaranteed the best route — while exploring far less than a blind search.
- Defined A*: expand the node with the lowest f = g (cost so far) + h (heuristic estimate)
- Proved A* finds an optimal path when h never overestimates, and is optimally efficient among comparable algorithms
- Developed for path planning on the Shakey robot
Still the standard pathfinding algorithm in games, maps and robots — see it computed in the lab (section 11).
05 🧠 The First Neural Networks
Perceptrons and adaptive neurons, the book that chilled them, and the quiet origin of backpropagation's maths.
The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain
Rosenblatt (Cornell Aeronautical Laboratory) · Psychological Review 1958An artificial neuron adds up weighted inputs and fires if the total passes a threshold. When it gets an answer wrong, nudge the weights toward the right answer. Repeat, and it learns to separate two kinds of input — the first neural network that learned from examples.
- Proposed the perceptron and its error-correcting learning rule
- Built as hardware (the Mark I Perceptron) for recognising simple images
- The perceptron convergence theorem later showed it always finds a separating line if one exists
The unit that deep networks still stack by the billion — with smoother activations and backpropagation.
Adaptive Switching Circuits (ADALINE and the LMS rule)
Widrow & Hoff (Stanford) · IRE WESCON Convention Record 1960Rather than just counting mistakes, measure how wrong the neuron's raw output is and move each weight a little to shrink that squared error. This least-mean-squares rule is gradient descent — the same idea that trains every modern network.
- ADALINE: an adaptive linear neuron trained with the least-mean-squares (Widrow–Hoff, delta) rule
- Minimised mean squared error by stochastic gradient descent on each example
- LMS became a workhorse of adaptive signal processing (echo cancellation, equalisers)
Gradient descent on a loss, one example at a time — stochastic gradient descent, the optimiser behind modern deep learning, in its earliest practical form.
Perceptrons: An Introduction to Computational Geometry
Minsky & Papert (MIT) · MIT Press 1969A careful mathematical book showing what single-layer perceptrons cannot do — for example, tell whether exactly one of two inputs is on (XOR), or whether a shape is connected. It was widely read as a verdict against neural networks, and funding moved elsewhere.
- Proved limits of single-layer perceptrons on problems such as parity and connectedness
- Multi-layer networks could escape these limits, but there was no known way to train them
- Minsky won the 1969 Turing Award for his role in creating and shaping AI
Contributed to a long lull in neural-network research until backpropagation (1986) trained hidden layers — the lab in our 1980–2000 article solves XOR that way.
Reverse-mode automatic differentiation
Linnainmaa (Helsinki) · master's thesis 1970; BIT 1976Any calculation is a chain of small steps. Record them on the way forward, then walk the chain backwards to get the derivative of the result with respect to every input at once — for about the cost of one extra pass. Applied to neural networks, this is backpropagation.
- Described the general method of reverse-mode automatic differentiation (in the context of rounding errors), with code
- Werbos (1974 thesis) proposed using such gradients to train neural networks
The engine inside PyTorch, JAX and TensorFlow: every model trained today uses reverse-mode differentiation.
06 💬 Symbols and Language
A language for AI, the first chatbot, and a program that understood commands about a world of blocks.
Recursive Functions of Symbolic Expressions and Their Computation by Machine (LISP)
McCarthy (MIT) · Communications of the ACM 1960A programming language where programs and data are the same thing — nested lists of symbols. That made it natural to write programs that build and change other programs, and it became AI's main language for thirty years.
- Defined LISP from a handful of primitives, with recursion and conditional expressions
- Introduced ideas such as garbage collection and treating code as data
- McCarthy received the 1971 Turing Award for his work in AI
Its ideas — functional programming, garbage collection, interactive interpreters — live on in Python, JavaScript and the notebooks AI researchers use daily.
ELIZA — A Computer Program for the Study of Natural Language Communication Between Man and Machine
Weizenbaum (MIT) · Communications of the ACM 1966The first famous chatbot. It spotted keywords and turned your sentence back into a question, like a therapist. It understood nothing — yet some users confided in it, which alarmed its creator.
- Scripts of decomposition and reassembly rules keyed on words (the DOCTOR script imitated a psychotherapist)
- Showed how easily people attribute understanding to simple programs — now called the ELIZA effect
The ancestor of chatbots, and an early warning about anthropomorphising machines — a live issue with LLM chat assistants.
Understanding Natural Language (SHRDLU)
Winograd (MIT) · PhD 1971; Cognitive Psychology 1972Talk to the computer in plain English about a small world of coloured blocks — "pick up a big red block", "what is in the box?" — and it plans, acts, and answers questions, keeping track of what "it" refers to.
- Combined parsing, meaning, reasoning and planning in one program for a simulated blocks world
- Handled pronouns, questions about its own actions and new definitions — but only within its tiny world
Showed deep understanding was possible in a micro-world, and that hand-built knowledge did not scale beyond it — a lesson LLMs sidestep by learning from text.
07 🤖 Seeing and Acting
How the brain sees edges, and the first robot that reasoned about its own actions.
Receptive Fields of Single Neurones in the Cat's Striate Cortex
Hubel & Wiesel (Johns Hopkins, Harvard) · Journal of Physiology 1959Recording single brain cells, Hubel and Wiesel found neurons in the visual cortex that fire for an edge at one particular angle in one small patch of the view. Vision is built from simple local detectors, layered into more complex ones.
- Visual cortex cells respond to edges and bars at specific orientations and positions (receptive fields)
- Later work distinguished "simple" and "complex" cells arranged in a hierarchy
- Hubel and Wiesel shared the 1981 Nobel Prize in Physiology or Medicine
The direct inspiration for the Neocognitron and convolutional networks — whose first-layer filters look strikingly like these edge detectors.
Shakey the Robot and STRIPS
SRI International · project 1966–72; Fikes & Nilsson, Artificial Intelligence 1971Shakey was the first mobile robot to reason about its own actions: given a goal like "push the box into the other room", it planned a sequence of steps, then carried them out with a camera and wheels. Its planner, STRIPS, described each action by what must be true before and what changes after.
- Integrated vision, planning (STRIPS) and navigation (A*) on one robot
- STRIPS represented actions by preconditions, add lists and delete lists — a format still used in planning languages
The template for robot planning, and the origin of A*; modern agents still plan by chaining actions with preconditions and effects.
08 📚 Logic and Knowledge
Machine-friendly logic, and the first expert systems in chemistry and medicine.
A Machine-Oriented Logic Based on the Resolution Principle
Robinson (Argonne, Syracuse) · Journal of the ACM 1965One simple rule — resolution — is enough for a computer to prove anything that follows from a set of logical facts. Add the negation of what you want to prove, and keep combining clauses until you reach a contradiction.
- A single inference rule, with unification, that is complete for first-order logic (refutation-complete)
- Made automated theorem proving practical and led to logic programming (Prolog, 1972)
The basis of theorem provers and logic programming; formal verification tools and LLM-plus-prover systems still build on it.
DENDRAL — the first expert system
Feigenbaum, Buchanan, Lederberg, Djerassi (Stanford) · from 1965; review in Artificial Intelligence 1978Given a mass spectrum, DENDRAL worked out which molecule produced it. It generated candidate structures and used chemists' rules to prune the impossible ones — expert knowledge, not general cleverness, did the work.
- Plan-generate-test using chemists' knowledge to infer molecular structure from mass-spectrometry data
- Meta-DENDRAL learned new rules from data
- Feigenbaum shared the 1994 Turing Award (with Reddy) for large-scale AI systems
Launched the "knowledge is power" era and the 1980s expert-system boom.
MYCIN — rule-based diagnosis of infections
Shortliffe, Buchanan et al. (Stanford) · thesis 1974; blinded evaluation, JAMA 1979About 600 if-then rules from doctors helped choose antibiotics for serious infections. Each rule came with a certainty factor, so evidence could add up — and MYCIN could explain which rules led to its advice.
- Backward-chaining rules with certainty factors and an explanation facility
- In a 1979 blinded evaluation on ten meningitis cases, experts rated MYCIN's therapy choices acceptable 65% of the time, versus 42.5–62.5% for Stanford faculty specialists
- Never used in routine care, partly over liability and integration issues
Showed expert-level performance in a narrow domain, pioneered explainable AI — and exposed the difficulty of deploying AI in medicine.
09 ❄️ Winter and Alternatives
A damning report that froze funding, and evolution as an alternative way to search.
Artificial Intelligence: A General Survey (the Lighthill report)
Lighthill (Cambridge), for the UK Science Research Council · 1973Asked to judge AI research, mathematician James Lighthill concluded it had not delivered on its promises: programs worked on toy problems but hit a combinatorial explosion on real ones. Funding was cut, and the first AI winter followed.
- Divided AI into automation, brain modelling and a "bridge" of general robots — the part he judged a failure
- Identified combinatorial explosion as the core obstacle
- Led the UK government to end support for AI research in most British universities; US funding also tightened in the same period
A lasting lesson about hype cycles — and about combinatorial explosion, which learning (not bigger search) eventually tamed.
Adaptation in Natural and Artificial Systems (genetic algorithms)
Holland (Michigan) · University of Michigan Press 1975Solve a problem the way evolution does: keep a population of candidate solutions, let the fittest reproduce, mix their parts (crossover), add random mutations, and repeat.
- Formalised genetic algorithms: selection, crossover and mutation on strings
- The schema theorem offered an account of why they work
Evolutionary methods live on in optimisation, design and neural architecture search — and in evolutionary approaches to improving AI-written programs.
10 🌱 What Survived
The 1950s–70s ideas still inside today's AI.
The founders expected human-level AI within a generation; it took far longer, because intelligence turned out to need vast knowledge and vast computation. But almost every thread of today's AI starts here: the learning neuron and gradient descent, reverse-mode differentiation, heuristic search, self-play, planning with actions and effects, chatbots, and the lesson that hand-written rules do not scale.
11 A* Search, Computed
The script finds the shortest path across a grid with walls, first with no heuristic (Dijkstra's method), then with A*'s "distance still to go" estimate, and counts how many cells each has to examine:
"""A* search from scratch (Hart, Nilsson & Raphael, 1968).
Find the shortest path across a grid with walls. Dijkstra's method explores
outward in every direction; A* adds a heuristic - the straight-line (here
Manhattan) distance still to go - so it explores toward the goal first. Because
the heuristic never overestimates, A* still returns an optimal path.
Standard library only.
"""
import heapq
GRID = [
"S.........#.........",
"..........#.........",
"..........#.........",
"..........#.........",
"......#####.........",
"....................",
"..............#####.",
"..............#.....",
"..............#....G",
"..............#.....",
]
H, W = len(GRID), len(GRID[0])
find = lambda ch: next((r, c) for r in range(H) for c in range(W) if GRID[r][c] == ch)
start, goal = find("S"), find("G")
def search(use_heuristic):
h = (lambda p: abs(p[0] - goal[0]) + abs(p[1] - goal[1])) if use_heuristic else (lambda p: 0)
frontier, cost, expanded = [(h(start), 0, start)], {start: 0}, set()
while frontier:
_, g, p = heapq.heappop(frontier)
if p in expanded: continue
expanded.add(p)
if p == goal: return g, len(expanded)
for dr, dc in ((1, 0), (-1, 0), (0, 1), (0, -1)):
q = (p[0] + dr, p[1] + dc)
if 0 <= q[0] < H and 0 <= q[1] < W and GRID[q[0]][q[1]] != "#" and g + 1 < cost.get(q, 1e9):
cost[q] = g + 1
heapq.heappush(frontier, (g + 1 + h(q), g + 1, q))
open_cells = sum(ch != "#" for row in GRID for ch in row)
print(f"Grid {W}x{H}, {open_cells} open cells, start {start}, goal {goal}\n")
for name, use in (("Dijkstra (no heuristic)", False), ("A* (Manhattan heuristic)", True)):
length, n = search(use)
print(f"{name:<26} path length {length} cells expanded {n}")
d, a = search(False)[1], search(True)[1]
print(f"\nSame optimal path; A* expanded {a} cells instead of {d} ({100 * (1 - a / d):.0f}% fewer).")
print("The heuristic is 'admissible' (never overestimates), which is what the 1968 paper")
print("proved is enough to guarantee the path A* returns is the shortest.")Output:
Grid 20x10, 183 open cells, start (0, 0), goal (8, 19) Dijkstra (no heuristic) path length 27 cells expanded 168 A* (Manhattan heuristic) path length 27 cells expanded 111 Same optimal path; A* expanded 111 cells instead of 168 (34% fewer). The heuristic is 'admissible' (never overestimates), which is what the 1968 paper proved is enough to guarantee the path A* returns is the shortest.
Both methods find a path of the same, shortest length, but A* looks at about a third fewer cells because the heuristic pulls it toward the goal. On bigger maps the savings grow. The same trade-off — a good guess about what is promising, while never overestimating — is why heuristic search powered AI for decades.
12 The 1950 → 2026 Roadmap
The whole series in one line. Press play: