Field Guide
History of AI
Key moments in AI history. Case study entries open full example cards.
Turing's "On Computable Numbers"
Alan Turing proposes the theoretical foundations of computation, including the concept of a universal machine that can simulate any other machine. This paper establishes what computers can and cannot do – and raises the question of what it might mean for a machine to think.
The Turing Test
Turing proposes the "imitation game" as a test of machine intelligence: if a machine can converse indistinguishably from a human, should we say it thinks? The test frames intelligence as behavioral rather than mechanical – and invites anthropomorphism by design.
The Dartmouth Workshop
John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon convene a summer workshop at Dartmouth College and coin the term "artificial intelligence." Their funding proposal predicts that a small group could make significant progress toward machine intelligence in a single summer – the first in a long line of AI predictions that badly underestimate the problem.
The Perceptron
Frank Rosenblatt builds the Perceptron, an early neural network that can learn to classify inputs. It generates enormous excitement and the first major cycle of AI hype – followed by the first AI winter when its limitations become apparent.
ELIZA and the Illusion of Understanding
Joseph Weizenbaum builds a program that simulates a therapist by pattern-matching user input and reflecting it back as questions. Users attribute genuine understanding and empathy to a system that has none.
The Asilomar Conference on Recombinant DNA (1975)
Concerned about unknown risks of recombinant DNA research, a group of leading molecular biologists called a voluntary moratorium on the most dangerous experiments and convened an international conference to establish biosafety guidelines – before a documented incident had occurred.
Backpropagation
Rumelhart, Hinton, and Williams publish the backpropagation algorithm, making it practical to train multi-layer neural networks. This is the technical foundation for modern deep learning, though its full societal impact won't be visible for another two decades.
Actuarial Risk Tools in Criminal Sentencing
Statistical instruments that use group-level factors to predict individual recidivism appear in US sentencing and parole decisions – establishing the template for algorithmic decision-making in high-stakes institutional contexts, and the arguments about fairness that persist today.
Deep Blue Defeats Kasparov
IBM's Deep Blue defeats world chess champion Garry Kasparov. The event triggers widespread public debate about machine intelligence and human uniqueness – and demonstrates how a narrow, domain-specific system can produce apparently intelligent behavior without general understanding.
The USA PATRIOT Act and the Architecture of Data Surveillance
The PATRIOT Act dramatically expands government authority to access communications and financial records – establishing the legal architecture for large-scale data collection and demonstrating how infrastructure built for one purpose can be repurposed at scale.
The Netflix Prize
Netflix offers $1 million to improve its recommendation algorithm -- and releases 100 million user ratings as the training dataset. Netflix offers $1 million to improve its recommendation algorithm -- and releases 100 million user ratings as the training dataset. Researchers subsequently show the data can be de-anonymized, exposing sensitive attributes users never consented to share. The winning algorithm is never deployed. The competition accelerates machine learning research and generates significant press about Netflix's technical sophistication, while establishing a template platforms have used repeatedly since: release proprietary data as a public challenge, extract free research labor, and retain the data and reputational benefits regardless of outcome.
AlexNet and the Deep Learning Breakthrough
AlexNet wins the ImageNet competition by a massive margin, demonstrating that deep convolutional neural networks trained on large datasets dramatically outperform prior approaches. This moment marks the beginning of the current AI era – and the point at which neural network opacity became structural.
The NSA PRISM Surveillance Program
Edward Snowden reveals that the NSA has direct access to data held by major technology companies – demonstrating the gap between what users understood their data was used for and what it was actually used for.
Amazon's Algorithmic Hiring Tool
Amazon develops a machine learning tool to screen resumes. Trained on historical hiring data, it learns to penalize resumes associated with women. Amazon discovers the bias internally and abandons the tool in 2018.
ProPublica Publishes "Machine Bias"
ProPublica's investigation of the COMPAS recidivism prediction tool finds it nearly twice as likely to falsely flag Black defendants as high-risk compared to white defendants – sparking a major public and academic debate about what it means for a predictive system to be fair.
"Attention Is All You Need"
A team at Google introduces the Transformer architecture, replacing recurrent processing with a self-attention mechanism that lets a model weigh every part of an input against every other part in parallel. This single architectural choice becomes the technical foundation for every major large language model that follows, including the systems behind ChatGPT five years later.
Facial recognition accuracy gaps (Gender Shades)
Systematic testing of commercial face analysis systems found error rates up to 34 percentage points higher for darker-skinned women than for lighter-skinned men.
Facebook data and Cambridge Analytica
Data collected through a personality quiz for academic research was used to build political targeting profiles for 87 million users.
Clearview AI
Clearview AI scrapes billions of images from public websites to build a facial recognition database used by law enforcement – testing the claim that publicly accessible data is available for any use.
Sidewalk Toronto
Sidewalk Labs, an Alphabet subsidiary, proposed a sensor-saturated 'smart neighborhood' on Toronto's eastern waterfront. The project was cancelled in 2020 after sustained public opposition that contested not just the privacy implications but the imaginary itself – the idea that cities are optimization problems to be solved through data collection.
Face recognition match thresholds and wrongful arrests
Three men were wrongfully arrested in part because face recognition systems returned weak matches that analysts treated as identifications.
Facebook, Frances Haugen, and the Limits of Engagement Metrics
Frances Haugen, a former Facebook product manager, leaks internal documents and testifies before Congress about Facebook's awareness of its systems' effects – raising the question of what a system optimized for engagement is actually optimizing.
The ChatGPT Release
OpenAI releases ChatGPT, making large language model capabilities available to the general public at scale for the first time – triggering a wave of adoption and public debate about work, creativity, education, and truth.
SAG-AFTRA and WGA Strikes
Hollywood writers and actors strike, with AI among the central issues – the first major labor actions explicitly about AI's potential to displace creative workers and use their likenesses and work without consent or compensation.
The EU AI Act
The European Union's AI Act enters into force – establishing the first comprehensive regulatory framework for AI systems, and becoming the reference point for global debates about AI governance.