A SHORT HISTORY OF ARTIFICIAL INTELLIGENCE

Can machines

think?

1950 → 2026 · from Turing to Claude

Discussion slides for AI and the Law. Watch the film first.

Seventy-six years in 21 moments

I · THE DREAM

1950

The imitation game

1956

A field gets a name

1958

The perceptron

1966

ELIZA

II · WINTERS AND RULES

1973

The first AI winter

1980s

Expert systems

1986

Learning from mistakes

1997

Deep Blue

III · LEARNING FROM DATA

2012

Deep learning takes off

2014

Machines that make

2016

Move 37

IV · THE GENERATIVE TURN

2017

Attention is all you need

2018–20

Bigger is better?

2022

Pictures from prompts

2022

ChatGPT

V · THE AGE OF ASSISTANTS

2023

Enter Claude

2023

The law responds

2024

Rules and recognition

2025

From chat to agents

2025

Tested in court

2026

A strategic technology

I · THE DREAM

1950

The imitation game

Alan Turing replaces the question “Can machines think?” with a test: can a machine’s typed answers pass for a person’s?

DISCUSS

Should the law care whether an AI is ‘intelligent’, or only about what it does?

THINK ABOUT

• Would anything change legally if a system passed Turing’s test?
• Where does law already judge conduct rather than what goes on inside?

I · THE DREAM

1956

A field gets a name

A summer workshop at Dartmouth College, proposed by John McCarthy and colleagues, names the field ‘artificial intelligence’ and expects rapid progress.

DISCUSS

Does calling it ‘intelligence’ shape how lawmakers and the public treat these systems?

THINK ABOUT

• What if the field had been called ‘complex information processing’?
• How do words like ‘learning’, ‘hallucination’ or ‘agent’ frame legal debates?

I · THE DREAM

1958

The perceptron

Frank Rosenblatt’s perceptron learns to recognise patterns by adjusting its connections, rather than by following rules someone wrote.

DISCUSS

When a system learns rather than being programmed, who is responsible for what it does?

THINK ABOUT

• The programmer, whoever supplied the data, or the user?
• Is ‘the machine learned it’ ever a defence?

I · THE DREAM

1966

ELIZA

Joseph Weizenbaum’s ELIZA mimics a therapist by turning users’ words back into questions. People confide in it anyway.

DISCUSS

What should a chatbot owe the people who trust it?

THINK ABOUT

• Should a bot always have to say it is a bot?
• Who answers for harm caused by a ‘companion’ or ‘therapy’ chatbot?

II · WINTERS AND RULES

1973

The first AI winter

Promises outrun results. Minsky and Papert show the limits of simple networks (1969), and in Britain the Lighthill Report (1973) leads to funding cuts.

DISCUSS

AI has boomed and bust before. Should that change how, and when, we regulate it?

THINK ABOUT

• Is it wiser to regulate early, or to wait and see?
• Who benefits from hype, and who pays when it collapses?

II · WINTERS AND RULES

1980s

Expert systems

AI returns as hand-written IF–THEN rules. Researchers even encode the British Nationality Act 1981 as a logic program.

DISCUSS

Could law itself be written as code? What would be gained, and what lost?

THINK ABOUT

• What happens to open-textured words like ‘reasonable’ or ‘good character’?
• Who checks that the code matches the statute?

II · WINTERS AND RULES

1986

Learning from mistakes

Rumelhart, Hinton and Williams popularise backpropagation: training many-layered neural networks by passing their errors backwards.

DISCUSS

Modern networks learn patterns nobody wrote down. Can we regulate what we can’t fully explain?

THINK ABOUT

• Is a right to an explanation realistic?
• Would you accept an unexplained decision that is more accurate than a human’s?

II · WINTERS AND RULES

1997

Deep Blue

IBM’s Deep Blue defeats world chess champion Garry Kasparov by searching around 200 million positions a second.

DISCUSS

Does it matter how a machine reaches a result, if the result is right?

THINK ABOUT

• Compare a judge, a doctor and an exam marker.
• When does the process matter as much as the outcome?

III · LEARNING FROM DATA

2012

Deep learning takes off

Trained on ImageNet’s 1.2 million labelled photos using gaming graphics chips, AlexNet wins a major image-recognition contest by a wide margin.

DISCUSS

The breakthrough was built on images taken from the web. Who should be asked, or paid, when their work trains AI?

THINK ABOUT

• The photographers, the people pictured, the workers who labelled them?
• Is ‘publicly available’ the same as ‘free to use’?

III · LEARNING FROM DATA

2014

Machines that make

Generative adversarial networks pit a forger against a detective until the fakes look real. In 2018 a GAN portrait sells at Christie’s for $432,500.

DISCUSS

Who, if anyone, is the author of a GAN-made portrait?

THINK ABOUT

• The coders, the people who chose the output, or no one?
• UK law: ‘computer-generated works’ under CDPA 1988 s 9(3).

III · LEARNING FROM DATA

2016

Move 37

DeepMind’s AlphaGo beats Go champion Lee Sedol 4–1. Its 37th move in game two stuns professional players.

DISCUSS

If a machine can make a ‘creative’ move, should its outputs be protected like human creativity?

THINK ABOUT

• What are copyright and patents for: rewarding people, or encouraging output?
• Does originality need a mind?

IV · THE GENERATIVE TURN

2017

Attention is all you need

Google researchers publish the Transformer, which learns how every word in a text relates to every other. The whole field builds on it.

DISCUSS

The key idea was published openly. Should AI models be open, or are some too risky to share?

THINK ABOUT

• Open-source licences versus ‘open weights’.
• Does openness improve safety through scrutiny, or make misuse easier?

IV · THE GENERATIVE TURN

2018–20

Bigger is better?

OpenAI’s GPT models grow from 117 million to 175 billion parameters. GPT-2 is at first held back over fears of misuse.

DISCUSS

Who should decide when an AI model is too dangerous to release?

THINK ABOUT

• The developer, a regulator, or an independent body?
• Should size or computing power trigger legal duties?

IV · THE GENERATIVE TURN

2022

Pictures from prompts

DALL·E 2, Midjourney and Stable Diffusion turn text into images. They are trained on billions of pictures scraped from the web.

DISCUSS

Is training on artists’ work more like learning from it, or copying it?

THINK ABOUT

• Does the model store the images it was trained on?
• Should artists be able to opt out, opt in, or be paid?

IV · THE GENERATIVE TURN

2022

ChatGPT

On 30 November, OpenAI releases a free chatbot as a ‘research preview’. An estimated 100 million people use it within two months.

DISCUSS

Should law regulate the technology, its uses, or the companies behind it?

THINK ABOUT

• Compare the EU’s single AI Act with the UK’s reliance on existing regulators.
• What changes when 100 million people adopt a tool at once?

V · THE AGE OF ASSISTANTS

2023

Enter Claude

Anthropic, founded in 2021 by former OpenAI researchers, releases Claude on the same day OpenAI launches GPT-4. Claude is trained using a written ‘constitution’.

DISCUSS

Can a company’s own principles do the job of regulation?

THINK ABOUT

• Who writes the constitution, and who enforces it?
• Compare codes of conduct in other industries.

V · THE AGE OF ASSISTANTS

2023

The law responds

Getty sues Stability AI in the UK, the UK Supreme Court rules on the DABUS AI, and the New York Times sues Microsoft and OpenAI.

DISCUSS

Should the law recognise AI inventors or authors, or leave AI output unprotected?

THINK ABOUT

• Who would benefit from each option?
• Is ‘no protection’ a problem, or a feature?

V · THE AGE OF ASSISTANTS

2024

Rules and recognition

The EU AI Act enters into force on 1 August, regulating AI by level of risk. Nobel Prizes go to AI pioneers in physics and chemistry.

DISCUSS

Pick an AI use you know. Which risk tier should it sit in, and why?

THINK ABOUT

• CV screening, exam proctoring, a customer-service chatbot, emotion recognition at work.
• Should the UK copy the EU’s approach?

V · THE AGE OF ASSISTANTS

2025

From chat to agents

AI moves from answering to acting: agents that browse, use computers and write code for hours. Anthropic releases Claude 4 and Claude Code.

DISCUSS

An AI agent works unsupervised for an hour and makes a costly mistake. Who is liable?

THINK ABOUT

• The user who set the task, the business that deployed it, or the developer?
• Can an AI agent make a binding contract?

V · THE AGE OF ASSISTANTS

2025

Tested in court

Thomson Reuters v Ross, Bartz v Anthropic and Getty v Stability AI bring the first major rulings on training AI with copyright works.

DISCUSS

Courts have reached different results. Should the UK change its law on text and data mining?

THINK ABOUT

• Keep the non-commercial exception, add an opt-out, or require licences?
• Who gains and who loses under each option?

V · THE AGE OF ASSISTANTS

2026

A strategic technology

Models now write software and find security flaws on their own. In June, a US export-control order briefly forces Anthropic to suspend its most powerful models.

DISCUSS

Should the most powerful AI be controlled like weapons technology, regulated like software, or treated as something new?

THINK ABOUT

• Who should hold the off-switch?
• What happens to users abroad when one government decides?

THE FILM YOU JUST WATCHED

So who is its author?

HOW IT WAS MADE

Claude researched, scripted, designed and animated it. A lecturer asked for it, set the brief and requested changes.

UK LAW

A film’s authors are its producer and principal director (CDPA 1988 s 9(2)(ab)). For a computer-generated script or artwork, s 9(3) names whoever made the necessary arrangements.

US LAW

Copyright needs a human author (Thaler v Perlmutter, DC Cir 2025). Only the human-authored parts of AI-assisted work are protected.

Is there a ‘principal director’ here? Did anyone ‘make the arrangements’? Should the film be protected at all?

Three more questions for this course

LIABILITY

Who is liable when AI gets things wrong?

Developer, deployer or user? Negligence, product liability, or something new?

TRAINING DATA

What may AI learn from?

Copyright, privacy and consent: whose works, and whose data?

REGULATION

Who writes the rules?

Legislators, courts, regulators, the companies, or the models’ own ‘constitutions’?

IN GROUPS

One moment, three questions

1. Pick one moment from the timeline.
2. What legal question did it raise at the time, if any?
3. What question does it raise now?
4. Which law, court or regulator would answer it today, and would you trust the answer?

TIME

10

minutes, then one sentence from each group.

Sources and further reading

THE HISTORY

• A. M. Turing, ‘Computing Machinery and Intelligence’ (1950) 59 Mind 433
• J. McCarthy et al., ‘A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence’ (1955)
• J. Lighthill, ‘Artificial Intelligence: A General Survey’ (Science Research Council, 1973)
• M. J. Sergot et al., ‘The British Nationality Act as a Logic Program’ (1986) 29 Communications of the ACM 370
• D. E. Rumelhart, G. E. Hinton and R. J. Williams, ‘Learning Representations by Back-propagating Errors’ (1986) 323 Nature 533
• A. Vaswani et al., ‘Attention Is All You Need’ (NeurIPS 2017)

THE LAW

• Copyright, Designs and Patents Act 1988, ss 9, 29A and 178
• Thaler v Comptroller-General of Patents, Designs and Trade Marks [2023] UKSC 49
• Getty Images v Stability AI [2025] EWHC 2863 (Ch) (on appeal)
• Thaler v Perlmutter (DC Cir, March 2025)
• Thomson Reuters v Ross Intelligence (D Del, February 2025) (on appeal)
• Bartz v Anthropic (ND Cal, June 2025); settlement approved 2026
• Regulation (EU) 2024/1689 (the EU AI Act)

CAN MACHINES THINK?

Welcome to

AI and the Law

MADE WITH CLAUDE · SEPTEMBER 2026