Artificial Intelligence

Artificial intelligence (AI) is a branch of computer science focused on creating systems capable of performing tasks that typically require human intelligence, including learning, reasoning, problem-solving, perception, and natural language understanding.^[1] The field was formally established at the Dartmouth Workshop in 1956, where the term "artificial intelligence" was first coined by John McCarthy.^[2]

Overview

Artificial intelligence encompasses a broad range of technologies and approaches designed to simulate or replicate aspects of human cognition in machines. Modern AI systems power applications ranging from virtual assistants and autonomous vehicles to medical diagnosis tools and scientific research platforms. As of 2024, AI has become one of the most transformative technologies of the 21st century, with private U.S. investment reaching $109.1 billion—nearly 12 times China's $9.3 billion and 24 times the U.K.'s $4.5 billion.^[3]

The field has experienced multiple periods of intense optimism followed by "AI winters" of reduced funding and interest, but has undergone a sustained renaissance since the deep learning revolution of 2012. In recognition of AI's scientific importance, the 2024 Nobel Prize in Physics was awarded to John J. Hopfield and Geoffrey Hinton "for foundational discoveries and inventions that enable machine learning with artificial neural networks," while the 2024 Nobel Prize in Chemistry was awarded to Demis Hassabis, John M. Jumper, and David Baker for work on protein structure prediction using AI.^[4]

History

Foundational Ideas (1940s–1950s)

The theoretical foundations of AI emerged from the work of mathematicians and computer scientists in the mid-20th century. In 1950, English mathematician Alan Turing published his landmark paper "Computing Machinery and Intelligence," which posed the question "can machines think?" and introduced what became known as the Turing Test—a method for determining whether a machine could exhibit intelligent behavior indistinguishable from a human.^[1]

The concept of a universal computing machine, developed by Turing and John von Neumann, provided the theoretical underpinnings for the field. Though both died before AI had advanced significantly, their ideas profoundly influenced the next generation of researchers, particularly Marvin Minsky and John McCarthy.^[2]

The Dartmouth Workshop (1956)

The field of AI research was formally founded at a workshop held on the campus of Dartmouth College in the summer of 1956.^[1] In 1955, John McCarthy, then a young Assistant Professor of Mathematics at Dartmouth, organized the gathering to "clarify and develop ideas about thinking machines."^[2]

The workshop proposal stated that researchers intended to test the assertion that "every aspect of learning or any other feature of intelligence can be so precisely described that a machine can be made to simulate it."^[2] McCarthy chose the name "Artificial Intelligence" for its neutrality, avoiding a narrow focus on automata theory and steering clear of the term "cybernetics," which was associated with analog feedback systems.^[2]

Key Participants

At the workshop, Newell and Simon presented the "Logic Theorist," which is considered the first AI program. The workshop is widely regarded as the birth of AI as an academic discipline.^[2]

Early Optimism (1956–1974)

Following the Dartmouth Workshop, AI researchers achieved results that seemed remarkable to observers at the time. Computers were solving algebra word problems, proving theorems in geometry, and learning to process natural language. Researchers expressed intense optimism, with some predicting that a fully intelligent machine would be built in less than 20 years.^[2]

ELIZA (1966)

One of the most influential early AI programs was ELIZA, developed by Joseph Weizenbaum at MIT between 1964 and 1967. ELIZA was the first chatbot in the history of computer science, using pattern matching and substitution methodology to simulate conversation.^[5]

The most famous ELIZA script, called DOCTOR, simulated a Rogerian psychotherapist by reflecting users' words back to them in the form of questions. Weizenbaum was shocked when users took the program seriously, opening their hearts to it despite its simple underlying mechanics. This phenomenon became known as the "ELIZA effect"—the tendency of humans to attribute understanding to computer programs.^[5]

Expert Systems: DENDRAL and MYCIN

DENDRAL, begun in 1965 at Stanford University, is considered the first expert system. Developed by Joshua Lederberg, Edward Feigenbaum, and Carl Djerassi, it automated the decision-making process of organic chemists to identify unknown compounds using mass spectrometry data.^[6]

MYCIN, developed at Stanford in the early 1970s by Edward Shortliffe under Bruce Buchanan's direction, was an expert system that used approximately 600 production rules to identify bacteria causing severe infections and recommend antibiotic treatments. In evaluations, MYCIN demonstrated accuracy comparable to infectious disease specialists, outperforming general practitioners in selecting proper antibiotics and dosages.^[7] Though never deployed clinically due to liability concerns and physician reluctance, MYCIN's approach inspired dozens of commercially available expert system tools.^[7]

The First AI Winter (1974–1980)

The optimism of the early years gave way to disappointment as AI systems failed to live up to their initial promise. From 1974 to 1980, AI funding declined drastically in what became known as the first "AI winter."^[8]

The Lighthill Report (1973)

A pivotal moment came in 1973 when British mathematician Sir James Lighthill was commissioned by the UK Parliament to evaluate the state of AI research. His report harshly criticized the "utter failure of AI to achieve its grandiose objectives" and questioned the viability of continued funding.^[8]

Funding Collapse

DARPA funding for AI research plummeted from approximately $30 million annually in the early 1970s to almost nothing by 1974. The Mansfield Amendment, passed by Congress, limited military funding for research lacking direct military applications, contributing to the decline.^[8]

The impact was severe:

• Academic publications in AI-focused journals dropped by nearly 48% between 1974 and 1980
• Enrollment in specialized AI graduate programs declined by over 60%
• Many laboratories closed or drastically reduced operations^[8]

Revival and the Expert Systems Boom (1980–1987)

The 1980s saw a rebirth of AI, driven largely by the commercial success of expert systems.^[8] Japan's Fifth Generation Computer Systems project, launched in 1982 with $850 million in funding, reinvigorated global AI research. By 1985, corporations were investing over $1 billion annually in AI, primarily focused on expert systems.^[8]

The Second AI Winter (Late 1980s–1990s)

The second AI winter began in the late 1980s and extended into the 1990s. This downturn was even more severe than the first, following a period of intense optimism and investment.^[8] The limitations of expert systems became apparent as they struggled to scale and adapt to new problems. The collapse of the specialized AI hardware market further damaged the field's commercial prospects.^[8]

The Backpropagation Revolution (1986)

Despite the broader decline, foundational work continued. In 1986, David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams published their landmark paper "Learning representations by back-propagating errors" in Nature.^[9]

The paper described how the backpropagation algorithm could train multi-layer neural networks by repeatedly adjusting connection weights to minimize the difference between actual and desired outputs. This enabled "hidden" internal units to learn to represent important features of the task domain—a capability that distinguished backpropagation from earlier methods like the perceptron.^[9]

While the mathematical foundations had been developed earlier by researchers like Seppo Linnainmaa (1970) and Paul Werbos (1970s–1980s), the 1986 paper brought the algorithm to widespread attention and demonstrated its practical effectiveness.^[9]

Convolutional Neural Networks (1989)

In 1989, Yann LeCun at AT&T Bell Laboratories introduced the convolutional neural network (CNN) architecture.^[10] Working with team members, LeCun applied backpropagation to train CNNs to read handwritten numbers, successfully applying the technology to recognize zip codes for the US Postal Service.^[10]

The LeNet architecture was eventually adapted for ATM check recognition. By the late 1990s and early 2000s, systems based on LeCun's work were reading over 10% of all checks in the United States.^[10]

Deep Blue Defeats Kasparov (1997)

On May 11, 1997, IBM's Deep Blue became the first computer system to defeat a reigning world chess champion under standard tournament conditions.^[11]

Deep Blue was a customized IBM RS/6000 SP supercomputer designed by computer scientist Feng-hsiung Hsu. Development had begun in 1985 at Carnegie Mellon University under the name ChipTest, later renamed Deep Thought, and finally Deep Blue in 1989.^[11]

The Matches

In the 1996 match, Deep Blue won the first game—the first time a computer had beaten a reigning world champion under regular time controls—but Kasparov recovered to win the match.^[11]

In the 1997 rematch, Kasparov won Game 1, Deep Blue won Game 2, Games 3–5 were draws, and Deep Blue won the decisive Game 6.^[11]

Technical Specifications

Deep Blue achieved its prowess through brute-force computing power:

• 32 processors performing coordinated high-speed computations in parallel
• Capability to evaluate 200 million chess positions per second
• Processing speed of 11.38 billion floating-point operations per second^[11]

After its victory, Deep Blue was retired to the Smithsonian Museum in Washington, D.C. IBM refused Kasparov's request for a rematch.^[11]

Long Short-Term Memory (1997)

In 1997, Sepp Hochreiter and Jürgen Schmidhuber introduced Long Short-Term Memory (LSTM), a breakthrough architecture for recurrent neural networks.^[12]

LSTMs addressed the vanishing gradient problem that had plagued training of neural networks on sequential data. By introducing "constant error carousels" and multiplicative gate units, LSTMs could learn dependencies across more than 1,000 discrete time steps.^[12]

An LSTM unit consists of:

• A cell that remembers values over arbitrary time intervals
• An input gate controlling what information enters the cell
• An output gate controlling what information leaves the cell
• A forget gate deciding what information to discard^[12]

In 2016, Google implemented LSTMs in Google Translate, reducing translation errors by 60%.^[12]

IBM Watson Wins Jeopardy! (2011)

On February 16, 2011, IBM's Watson became the first computer to win the game show Jeopardy! against human champions.^[13]

Watson faced Ken Jennings (holder of the record for 74 consecutive Jeopardy! wins) and Brad Rutter (the show's all-time highest money winner). Watson won decisively with $77,147, compared to Jennings' $24,000 and Rutter's $21,600.^[13]

Technical Architecture

The original Watson consisted of:

• 10 racks holding 90 servers
• 2,880 processor cores
• 200 million pages of content including the full text of Wikipedia (2011 edition)
• 4 terabytes of disk storage
• No internet connection during the competition^[13]

Watson used IBM's DeepQA software and hundreds of algorithms to analyze questions and generate weighted lists of possible answers.^[13]

Notable moment: Ken Jennings' Final Jeopardy answer included the phrase "I for one welcome our new computer overlords."^[13]

The Deep Learning Revolution (2012)

The modern AI era began in 2012 when AlexNet, a deep convolutional neural network, achieved breakthrough results in the ImageNet Large Scale Visual Recognition Challenge.^[14]

AlexNet

Developed by Alex Krizhevsky in collaboration with Ilya Sutskever and Geoffrey Hinton at the University of Toronto, AlexNet achieved a top-5 error rate of 15.3%, dramatically outperforming the second-best entry's 26.2%.^[14]

Key specifications:

• 60 million parameters
• 650,000 neurons
• 8 layers (5 convolutional, 3 fully connected)
• Trained on 2 Nvidia GTX 580 GPUs^[14]

At the 2012 European Conference on Computer Vision, researcher Yann LeCun described AlexNet as "an unequivocal turning point in the history of computer vision."^[14]

Enabling Factors

Two developments paved the way for AlexNet:

1. Large-scale training data: ImageNet, created by Fei-Fei Li's team, provided over 14 million hand-annotated images across more than 20,000 categories^[14]
2. GPU computing: Graphics processing units enabled the parallel processing necessary for training deep networks^[14]

Commercial Impact

Within months of AlexNet's victory, Hinton and his students formed a shell company that was acquired by Google for $44 million.^[14] The technology led to subsequent breakthroughs including VGG, GoogLeNet, and ResNet.^[14]

AlphaGo Defeats Lee Sedol (2016)

In March 2016, DeepMind's AlphaGo defeated Lee Sedol, one of the world's top Go players, in a five-game match with a score of 4–1.^[15]

Go's complexity—with more possible board positions than atoms in the observable universe—had long made it a grand challenge for AI. AlphaGo's victory "shocked Go and AI experts alike—and changed the world's perception of what AI can do."^[15]

DeepMind later developed AlphaZero, which surpassed specialized AI systems in multiple games. In chess, AlphaZero defeated Stockfish after only a few hours of self-play training.^[15]

The Transformer Architecture (2017)

In 2017, researchers at Google published "Attention Is All You Need," introducing the transformer architecture that would revolutionize AI.^[16]

The paper was authored by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan Gomez, Łukasz Kaiser, and Illia Polosukhin. All eight were "equal contributors" to the paper, with the listed order randomized.^[16]

Key Innovations

The transformer architecture:

• Replaced recurrence with self-attention mechanisms
• Enabled parallel processing of all tokens in a sequence
• Dramatically reduced training times
• Allowed scaling to much larger models^[16]

The architecture uses multi-head self-attention with an encoder composed of 6 identical layers, each with two sub-layers: a multi-head self-attention mechanism and a position-wise feed-forward network. The model uses a dimension of d_model = 512.^[16]

Performance

The transformer achieved:

• 28.4 BLEU on WMT 2014 English-to-German translation (2+ BLEU improvement over previous best)
• 41.8 BLEU on WMT 2014 English-to-French (new single-model state-of-the-art)
• Training in as little as 12 hours on 8 P100 GPUs^[16]

Impact

As of 2025, the paper has been cited more than 173,000 times, placing it among the top ten most-cited papers of the 21st century.^[16] It is considered a foundational paper in modern AI and a main contributor to the current AI boom, as transformers have become the primary architecture for large language models.^[16]

The paper's title is a reference to "All You Need Is Love" by The Beatles.^[16]

BERT (2018)

In 2018, Google released BERT (Bidirectional Encoder Representations from Transformers), which introduced bidirectional attention to transformer models.^[17]

Unlike previous models that processed text in one direction, BERT analyzed context from both directions simultaneously. It was pre-trained on a large corpus and then fine-tuned for specific tasks including natural language inference and sentence similarity.^[17]

BERT quickly became "ubiquitous" and was used to improve query understanding in Google Search by 2019.^[17]

GPT Series (2018–Present)

OpenAI's Generative Pre-trained Transformer (GPT) series has driven much of the recent progress in large language models.

ChatGPT (November 30, 2022)

ChatGPT was released on November 30, 2022, as a free research preview.^[19] It became the fastest-growing consumer application in history:

• 1 million users within 5 days
• 100 million users by January 2023 (less than 3 months)^[19]

For comparison:

• TikTok took 9 months to reach 100 million users
• Instagram took 2.5 years
• Netflix took 3.5 years to reach 1 million users^[19]

As of September 2025, ChatGPT has more than 750 million weekly active users, with users sending more than 2.6 billion messages per day (over 30,000 messages per second).^[19]

GPT-4 (March 14, 2023)

GPT-4 was released on March 14, 2023, as OpenAI's most capable model at the time.^[18]

Capabilities

• Multimodal: Accepts both image and text inputs^[18]
• Academic performance: Passes simulated bar exam at top 10% of test takers (GPT-3.5 scored in bottom 10%)^[18]
• Improved reliability: More capable of handling nuanced instructions^[18]

Early adopters included Microsoft (Bing Chat), Stripe, Duolingo, Morgan Stanley, and Khan Academy.^[18]

AlphaFold and Scientific Discovery (2020–2024)

In 2020, DeepMind's AlphaFold 2 solved the 50-year-old grand challenge of protein structure prediction.^[20]

AlphaFold 2 could predict protein structures to within the width of an atom, matching the accuracy of laboratory techniques but returning results in hours instead of months.^[20]

Impact

• Over 3 million researchers from more than 190 countries use AlphaFold^[20]
• Applications include malaria vaccine development and liver cancer treatments^[20]
• In partnership with EMBL-EBI, DeepMind released predicted structures for over 200 million proteins^[20]
• The AlphaFold paper has been cited over 40,000 times^[20]

In 2024, Demis Hassabis and John Jumper were awarded the Nobel Prize in Chemistry for AlphaFold.^[4]

2024 Nobel Prizes for AI

The 2024 Nobel Prizes marked historic recognition of AI's scientific contributions.^[4]

Physics Prize

John J. Hopfield (born 1933, Princeton University) and Geoffrey Hinton (born 1947, University of Toronto) received the Nobel Prize in Physics "for foundational discoveries and inventions that enable machine learning with artificial neural networks."^[4]

• Hopfield invented the Hopfield network, an associative memory structure
• Hinton developed the Boltzmann machine for autonomous property discovery in data^[4]

Chemistry Prize

Demis Hassabis and John M. Jumper (both of Google DeepMind) shared the Chemistry Prize with David Baker "for computational protein design and protein structure prediction."^[4]

Technical Foundations

Types of Machine Learning

Machine learning, the primary approach to modern AI, encompasses several paradigms:^[21]

Supervised Learning

Algorithms train on labeled datasets where input data is paired with correct outputs. The model learns relationships between inputs and outputs to make predictions on new data.^[21]

Common algorithms: Linear regression, decision trees, support vector machines, neural networks

Applications: Spam detection, sentiment analysis, medical diagnosis^[21]

Unsupervised Learning

Algorithms find patterns in unlabeled data without predefined outputs. The model identifies structure through grouping similar data points or detecting patterns.^[21]

Common algorithms: K-means clustering, principal component analysis, hierarchical clustering

Applications: Customer segmentation, recommendation systems, anomaly detection^[21]

Reinforcement Learning

Agents learn by interacting with environments, receiving rewards or penalties for actions. The goal is to maximize long-term cumulative reward through trial and error.^[21]

Common algorithms: Q-learning, SARSA, Deep Q-Networks (DQN)

Applications: Game playing, robotics, autonomous vehicles^[21]

Semi-Supervised Learning

Combines small amounts of labeled data with large amounts of unlabeled data. Particularly useful when labeling is expensive or time-consuming, such as medical imaging.^[21]

Neural Network Architectures

Feedforward Neural Networks

The simplest architecture, where information flows in one direction from input to output through hidden layers.

Convolutional Neural Networks (CNNs)

Specialized for processing grid-like data such as images. CNNs use convolutional layers with learnable filters that detect features like edges, textures, and complex patterns.^[10]

Recurrent Neural Networks (RNNs)

Designed for sequential data, with connections that form directed cycles. Information can persist across time steps, making RNNs suitable for language modeling and time series analysis.

Long Short-Term Memory (LSTM)

A specialized RNN architecture that can learn long-range dependencies through gated memory cells.^[12]

Transformers

The dominant architecture for modern large language models, using self-attention mechanisms to process sequences in parallel and capture relationships between all positions simultaneously.^[16]

Large Language Models

Large language models (LLMs) are AI systems trained on vast text corpora to generate and understand natural language. Key developments include:^[17]

Reinforcement Learning from Human Feedback (RLHF)

RLHF fine-tunes LLMs for desired behaviors using reward signals derived from human preferences. This technique is critical for:

• Aligning model outputs with user expectations
• Improving factuality
• Reducing harmful responses^[17]

Generative AI

Generative AI creates new content—text, images, audio, video—from training data and user prompts.

Text-to-Image Models

Models generate images from natural language descriptions:^[22]

These systems use diffusion techniques, progressively adding and removing noise from images during training to learn how to generate coherent images from random pixels.^[22]

More than 34 million images are generated by AI per day.^[22]

Key Organizations

OpenAI

OpenAI was founded in December 2015 by Sam Altman, Elon Musk, Greg Brockman, Ilya Sutskever, and others as a nonprofit research lab with the mission to develop AI beneficial to humanity.^[23]

The organization transitioned toward commercialization following ChatGPT's 2022 launch, becoming "one of the fastest-growing commercial entities on the planet."^[23]

OpenAI's annualized revenue exceeded $1.6 billion by December 2023.^[19]

Google DeepMind

DeepMind was founded as a startup before being acquired by Google in January 2014 for a reported $400–650 million.^[23]

Investors included Horizons Ventures, Founders Fund, Peter Thiel, and Elon Musk.^[23]

In April 2023, DeepMind merged with Google Brain to form Google DeepMind as part of Google's response to ChatGPT.^[23]

Anthropic

Anthropic was founded in 2021 by seven former OpenAI employees, including siblings Dario Amodei (formerly OpenAI's Vice President of Research) and Daniela Amodei.^[23]

Other founders include Tom Brown, Chris Olah, Sam McCandlish, Jack Clark, and Jared Kaplan.^[23]

Anthropic develops Claude, an AI assistant focused on being "helpful, harmless, and honest" through constitutional AI principles.^[17]

As of November 2025, Anthropic was valued in the range of $350 billion following investment deals with Microsoft and Nvidia.^[23]

Notable Researchers

The "Godfathers of AI"

Three researchers received the 2018 ACM A.M. Turing Award for their foundational work on deep learning:^[24]

Geoffrey Hinton

• Birth: 1947, London, UK
• PhD: 1978, University of Edinburgh
• Positions: Professor emeritus at University of Toronto, former VP and Engineering Fellow at Google, Chief Scientific Adviser at Vector Institute
• Key contributions: Backpropagation, Boltzmann machines, deep belief networks
• Nobel Prize: 2024 Physics^[24]

Yann LeCun

• Positions: VP and Chief AI Scientist at Meta, Professor at New York University, Co-director of CIFAR's Learning in Machines & Brains program
• Key contributions: Convolutional neural networks (LeNet), image recognition systems^[24]

Yoshua Bengio

• Positions: Professor at University of Montreal, Scientific Director at Mila (Quebec AI Institute), Canada-CIFAR AI Chair
• Key contributions: Sequence learning, neural language models^[24]

Other Pioneers

Applications

Healthcare

AI has achieved significant medical breakthroughs:^[25]

• Medical device approvals: The FDA approved 223 AI-enabled medical devices in 2023, up from 6 in 2015^[25]
• Cancer detection: Google's AI system for breast cancer detection shows 94.5% accuracy, exceeding human radiologists^[25]
• Surgical robotics: AI-powered robots assist in surgeries with precision^[25]
• Drug discovery: AlphaFold accelerates understanding of protein structures for therapeutic development^[20]

Finance

AI applications in finance include:^[25]

• Algorithmic trading with real-time data analysis
• Fraud detection and prevention
• Risk assessment and credit scoring
• Automated customer service

Autonomous Vehicles

As of 2024:^[25]

• Waymo provides over 150,000 autonomous rides per week in the United States
• Baidu's Apollo Go robotaxi fleet serves numerous cities across China

Robotics and Manufacturing

AI integration in manufacturing includes:^[25]

• Predictive maintenance
• AI-driven quality assurance
• Collaborative robots (cobots)
• Automated logistics

Ethics, Safety, and Societal Impact

Algorithmic Bias

AI systems can perpetuate and amplify societal biases:^[26]

Documented Cases

• Amazon hiring (2018): AI-powered hiring tool systematically discriminated against female applicants^[26]
• Lending discrimination: AI credit scoring models have shown bias favoring White borrowers over Black and Hispanic applicants with similar financial profiles^[26]
• Dutch child benefit scandal: An algorithm targeting dual nationals and ethnic minorities falsely flagged 26,000 parents for benefit fraud, forcing repayments without appeal rights^[26]
• Facial recognition: Studies reveal higher error rates for people with darker skin tones and women^[26]

Mitigation Approaches

Experts recommend sociotechnical frameworks combining:^[26]

• Technical debiasing methods
• Human-in-the-loop oversight
• Regulatory compliance
• Continuous evaluation
• Stakeholder engagement

AI Safety and Alignment

AI safety research addresses risks from advanced AI systems:^[27]

The Alignment Problem

"If we build an AI system that's significantly more competent than human experts but it pursues goals that conflict with our best interests, the consequences could be dire."^[27]

Research Priorities

Anthropic's alignment research focuses on six areas:^[27]

1. Mechanistic Interpretability
2. Scalable Oversight
3. Process-Oriented Learning
4. Understanding Generalization
5. Testing for Dangerous Failure Modes
6. Evaluating Societal Impact

Industry Collaboration

In 2025, Anthropic and OpenAI conducted joint alignment evaluation exercises to raise industry standards for safety testing.^[27]

Expert Concerns

According to a 2023 report by the Center for AI Safety, over 70% of AI researchers believe advanced AI could pose existential risks if not properly managed.^[27]

The 2025 AI Safety Index from the Future of Life Institute found that while companies claim they will achieve AGI within the decade, none scored above D in Existential Safety planning. One reviewer called this "deeply disturbing," noting that "none of the companies has anything like a coherent, actionable plan" for ensuring such systems remain safe and controllable.^[27]

Regulation

EU AI Act

The EU AI Act (Regulation (EU) 2024/1689) is the first comprehensive legal framework on AI worldwide, finalized on February 2, 2024.^[28]

Key features:

• Risk-based approach: Obligations vary based on risk level (minimal, limited, high, unacceptable)
• Banned applications: Real-time biometric surveillance in public spaces, social scoring, manipulative behavioral targeting
• Penalties: Up to 35 million euros or 7% of global turnover^[28]

Timeline:

• Entered into force: August 1, 2024
• Prohibited practices effective: February 2, 2025
• Full application: August 2, 2026^[28]

United States

The U.S. approach remains fragmented:^[28]

• No comprehensive federal AI legislation as of 2024
• Biden Administration's October 2023 Executive Order focused on federal agency guidance
• Trump Administration's Executive Order 14179 emphasized "America's global AI dominance"^[28]

State-level activity includes:

• Colorado AI Act (May 2024): First comprehensive state AI legislation^[28]
• Utah AI Policy Act (May 2024): Requires disclosure of GenAI use in consumer communications^[28]
• California: 18 AI-related bills signed into law in 2024^[28]

Impact on Employment

Job Displacement Projections

• 83 million jobs could be lost globally between 2023 and 2027 (20% more than jobs created)^[29]
• Goldman Sachs predicts AI may replace 300 million jobs worldwide (9.1% of global employment)^[29]
• By 2030, 30% of current U.S. jobs could be fully automated^[29]

Occupations at High Risk

Gender Disparities

In high-income countries, 9.6% of female jobs are in the highest-risk category for AI automation, compared to 3.2% of male jobs.^[29]

Countervailing Factors

• World Economic Forum predicts 97 million new jobs created by the "AI Revolution"^[29]
• 91% of companies using AI plan to hire new employees in 2025^[29]
• Goldman Sachs estimates unemployment will increase by only 0.5 percentage points during the AI transition^[29]
• PwC's 2025 AI Jobs Barometer finds AI can make people "more valuable, not less"^[29]

Timeline of Major Events

See Also

• The Internet
• DNA