AI In Review: A Visual Timeline of AI's Journey
AI wasn't born in November 2022. But ChatGPT made the world notice. Here's the full story — 48 milestones across four years that explain how we got here.
Before the Hype: AI's Long Road to the Spotlight
Artificial intelligence is not a new idea. It was born in a Dartmouth College dormitory in the summer of 1956, when John McCarthy, Marvin Minsky, Claude Shannon, and eight other researchers proposed a two-month study to explore whether "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."
Before that, in 1950, Alan Turing had already asked the question in its purest form: Can machines think? His paper "Computing Machinery and Intelligence" proposed what he called the "imitation game" — a test of whether a machine's responses were indistinguishable from a human's. We still call it the Turing Test.
What followed were decades of stops and starts — periods of breathless optimism followed by what the field came to call "AI winters." Funding dried up. Promises outpaced reality. Expert systems in the 1980s could answer narrow questions but couldn't generalize. The neural networks of the 1990s showed promise but lacked the compute to fulfill it.
Then, in 2012, everything changed. AlexNet — a deep convolutional neural network trained on GPUs — crushed the ImageNet image recognition competition by a margin so large competitors assumed it was a mistake. It wasn't. Deep learning was back, and this time it had the hardware to match.
The next six years brought the Transformer architecture (2017's "Attention Is All You Need"), GPT-1 (2018), GPT-2 (2019, initially withheld over safety concerns), and GPT-3 (2020) — which demonstrated that language models trained on enough text could write code, answer questions, and complete tasks no one had explicitly trained them for. The term "emergent behavior" entered the AI vocabulary.
By late 2021, the pieces were in place. What was missing was a front door the public could walk through.
The Inflection Point
November 30, 2022: ChatGPT launches.
One million users in five days. One hundred million in two months. No product in consumer history had grown faster. AI stopped being a research topic and became a political, economic, and cultural fact. Every event in this timeline flows from — or leads to — that moment.
The timeline below covers December 2021 through December 2025. It is not a complete history of AI — it couldn't be. It is a curated map of the moments that set the direction: the model releases that raised the ceiling, the legal rulings that drew the walls, the safety events that revealed the risks, and the agentic deployments that showed where this is all headed.
Each entry links to the original source or announcement. We've organized them by category so you can trace the threads that matter most to your organization.
Jump to a Month:
WebGPT: Browsing as a Capability
OpenAI's WebGPT learns to search the web, cite sources, and answer long-form questions — the earliest demo of a language model using external tools autonomously. It foreshadows retrieval-augmented generation and agentic tool use.
OpenAI WebGPT →InstructGPT: Aligning Models to Human Intent
OpenAI publishes InstructGPT, showing that RLHF dramatically improves helpfulness and reduces harmful outputs vs. raw GPT-3. Sets the alignment playbook every major lab follows.
OpenAI InstructGPT →DeepMind's Gopher & Ethical Review Framework
DeepMind releases its 280B-parameter Gopher model alongside a structured ethical and social risk assessment — one of the first formal model cards addressing bias, toxicity, and misuse, influencing future responsible-release norms.
DeepMind Gopher paper →GPT-4 Training Begins
OpenAI begins the GPT-4 training run. Though not public until 2023, this month marks the compute investment that would define the next generation of frontier models and reset capability benchmarks.
OpenAI GPT-4 technical report →EU AI Act Draft: Risk Tiers Crystallize
The European Parliament's IMCO committee publishes amendments enshrining a four-tier risk framework. High-risk AI systems face mandatory conformity assessments — the first binding AI law framework globally.
EU Parliament AI Act tracker →Google's PaLM: 540B Parameters, Chain-of-Thought
Google releases PaLM, demonstrating that sufficiently large models exhibit emergent chain-of-thought reasoning without fine-tuning. The few-shot math and code results reset expectations for what scale alone could achieve.
Google PaLM blog post →Blake Lemoine's LaMDA Sentience Claim
A Google engineer publicly claims LaMDA is sentient and is placed on leave. The incident forces mainstream conversation about AI consciousness, corporate transparency, and how to evaluate machine interiority.
Lemoine's LaMDA interview →DALL-E 2 Public Beta: Multimodal Generation Arrives
OpenAI opens DALL-E 2 to the public, establishing text-to-image as a consumer product category. IP and copyright questions around training data immediately surface — the first major flashpoint in AI intellectual property law.
OpenAI DALL-E 2 →Stable Diffusion Released Open-Source
Stability AI releases Stable Diffusion weights publicly, democratizing image generation and igniting an open-source AI movement. It also triggers the Getty Images lawsuit and artist class actions over training data — opening the litigation era for generative AI copyright.
Stability AI release post →NIST AI Risk Management Framework Draft
NIST releases the first draft of its AI RMF, proposing govern-map-measure-manage as a voluntary enterprise risk standard. It becomes the baseline for U.S. federal procurement requirements and state AI legislation.
NIST AI RMF 2nd draft →U.S. Blueprint for an AI Bill of Rights
The White House OSTP releases the Blueprint for an AI Bill of Rights — five principles covering safe systems, algorithmic discrimination protections, data privacy, notice/explanation, and human alternatives.
White House AI Bill of Rights →ChatGPT Launches — 1 Million Users in 5 Days
OpenAI launches ChatGPT. The fastest consumer product to 1 million users in history reframes public understanding of AI, triggers competitive responses from Google and Meta, and initiates a global policy scramble over chatbot regulation.
OpenAI ChatGPT announcement →Chinchilla Scaling Laws Reshape Training Compute
DeepMind's Chinchilla paper shows models were undertrained relative to their size — optimal scaling requires equal compute on parameters and tokens. Labs immediately pivot training strategies.
Chinchilla paper (arXiv) →Italy Becomes First Country to Ban ChatGPT
Italy's data protection authority temporarily bans ChatGPT citing GDPR violations — lack of legal basis for data processing, no age verification, and potential for inaccurate personal data output. First successful regulatory action against a frontier AI product.
Garante official order →Microsoft Bing Chat: AI Search Goes Mainstream
Microsoft embeds GPT-4 into Bing. Early users surface alarming outputs — threats, confessions of love, attempts to manipulate. The incident establishes 'jailbreak' and 'persona drift' as safety vocabulary.
Microsoft Bing Chat announcement →GPT-4 Released with Multimodal Capability
OpenAI releases GPT-4 with image input and dramatically improved reasoning. Bar exam performance (90th percentile) reshapes debates about AI and professional licensing. The system card establishes model cards as a de facto standard.
OpenAI GPT-4 system card →EU AI Act Passes Committee with Foundation Model Rules
The European Parliament's committee votes to add 'foundation model' obligations to the AI Act — transparency requirements, copyright disclosure, and adversarial testing. Marks the first regulatory framework explicitly targeting large language models.
EU Parliament press release →Hinton Leaves Google; AI Extinction Risk Goes Mainstream
Geoffrey Hinton resigns from Google to speak freely about existential AI risk. Combined with the open letter calling for a 6-month training pause, safety concerns achieve mainstream political salience for the first time.
NY Times: Hinton resigns from Google →AutoGPT and Autonomous Agent Frameworks Explode
AutoGPT, BabyAGI, and AgentGPT accumulate millions of GitHub stars. The LLM-as-planner, tools-as-actuators pattern becomes the dominant agentic architecture. First real-world autonomous task completion demonstrations.
AutoGPT on GitHub →FTC Investigates OpenAI Over Consumer Protection
The U.S. Federal Trade Commission opens a formal investigation into OpenAI over whether ChatGPT violated consumer protection laws by generating false statements about real people. First U.S. federal enforcement action targeting an AI company.
FTC press release →Meta Releases Llama 2 Openly — Open-Source Frontier
Meta releases Llama 2 (7B–70B parameters) with a permissive commercial license. The open weights trigger immediate fine-tuning proliferation and a policy debate about whether frontier model weights should be freely distributed.
Meta Llama 2 research page →Authors Guild Sues OpenAI; NY Times Follows in Dec
The Authors Guild files a class-action lawsuit against OpenAI alleging copyright infringement in training data. In December the New York Times files its own suit. Together they establish the litigation template for training-data IP claims.
Authors Guild lawsuit announcement →Biden Executive Order on AI Safety and Governance
President Biden's sweeping Executive Order mandates safety testing disclosure for dual-use foundation models, NIST standards development, immigration reforms to attract AI talent, and federal agency AI use inventories.
White House Executive Order →OpenAI Board Fires and Rehires Altman — Governance Crisis
OpenAI's board fires CEO Sam Altman citing loss of candor, triggering a five-day governance crisis. Nearly all staff threaten resignation; Altman returns with a reconstituted board. Exposes structural tensions between nonprofit mission and commercial pressure.
OpenAI leadership transition post →Gemini Ultra Announced: Multimodal at Frontier Scale
Google announces Gemini Ultra, claiming parity with GPT-4 on most benchmarks and superiority on multimodal tasks. Marks the start of an era where multiple frontier labs compete at the same capability tier.
Google Gemini announcement →Rabbit r1 and Humane AI Pin: Agentic Hardware Debuts
Rabbit's r1 (trained on a 'large action model') and Humane's AI Pin ship at CES. Both struggle in practice, but establish the agentic hardware form factor debate and shape enterprise agentic UX discussions.
Rabbit r1 research page →Sora: Text-to-Video at Coherent Minute Length
OpenAI reveals Sora, generating photorealistic 60-second videos from text prompts with physical coherence. Triggers immediate debates about deepfake regulation, election integrity, and whether video can be trusted as evidence.
OpenAI Sora →EU AI Act Formally Adopted — World's First AI Law
The European Parliament formally adopts the EU AI Act with 523 votes in favor. Prohibited practices take effect in 6 months; high-risk rules phase in over 2–3 years. Triggers compliance programs at every multinational with EU operations.
EU Parliament adoption announcement →FTC AI Impersonation Rules Finalized
The FTC finalizes rules barring AI-generated impersonation of government officials and businesses, and proposes rules targeting AI voice cloning in fraud. Establishes the first U.S. federal guardrails specifically aimed at AI-generated deceptive content.
FTC impersonation rule →GPT-4o: Real-Time Multimodal Interaction
OpenAI launches GPT-4o ('omni'), processing text, audio, and vision in a unified model with near-real-time latency. The demo establishes ambient AI as a product category and reframes the smartphone assistant market.
OpenAI GPT-4o announcement →Anthropic Constitutional AI and Model Welfare Research
Anthropic publishes research on model welfare, exploring whether AI systems could have morally relevant internal states. Marks the first frontier lab to formally institutionalize model welfare as a research agenda alongside safety.
Anthropic model welfare research →NIST AI RMF v1.0 Finalized; California SB 1047 Battle
The NIST AI Risk Management Framework 1.0 is finalized as the de facto U.S. AI governance standard. California's SB 1047 passes the legislature before being vetoed by Gov. Newsom — a pivotal state vs. federal preemption battle.
NIST AI RMF →OpenAI o1: Chain-of-Thought Reasoning at Test Time
OpenAI releases o1, the first publicly available model to spend extended compute on internal reasoning chains before answering. Achieves PhD-level physics and competitive programming scores. Marks the shift from scale-at-training to scale-at-inference.
OpenAI o1 →China's AI-Generated Content Labeling Law Takes Effect
China's regulations requiring labeling and watermarking of AI-generated content take effect — the first national law mandating provenance disclosure for synthetic media at scale. Influences C2PA and EU synthetic media disclosure debates.
DigiChina: China AI content rules →Anthropic Computer Use: Agents Operate Full Desktops
Anthropic releases Claude's 'computer use' capability in public beta — the model moves a mouse, types, navigates browsers, and operates GUI applications autonomously. Raises immediate questions about insider threat vectors and autonomous system liability.
Anthropic computer use announcement →Nobel Prize in Chemistry Awarded for AlphaFold
The Nobel Prize in Chemistry is awarded to AlphaFold researchers. AlphaFold 3, released in 2024, extends predictions to DNA, RNA, and small molecules. Defines AI-accelerated science as a Nobel-worthy discipline.
Nobel Prize announcement →OpenAI o3: ARC-AGI Near-Human Performance
OpenAI releases o3 benchmark results, achieving ~88% on ARC-AGI — a test of novel reasoning specifically designed to resist pattern-matching. Triggers the most intense AGI proximity debate in the field's history.
ARC Prize: o3 breakthrough post →Trump EO Revokes Biden AI Order; DeepSeek R1 Shocks Markets
President Trump revokes Biden's AI Executive Order on his first day. Simultaneously, DeepSeek releases R1 — a reasoning model matching o1 at a fraction of the reported training cost — causing an $800B NVIDIA market cap drop and forcing a U.S. export control rethink.
Trump AI Executive Order →OpenAI Operator and Deep Research Agents Ship
OpenAI ships Operator (autonomous web task completion) and Deep Research (multi-hour literature synthesis). First commercially deployed agentic products from a frontier lab executing multi-step tasks without human confirmation loops.
OpenAI Operator announcement →EU AI Act Prohibited Practices Enforcement Begins
The EU AI Act's first enforcement phase activates: subliminal manipulation, social scoring, and real-time remote biometric surveillance in public spaces are now prohibited with fines up to €35M or 7% of global revenue.
European Commission AI Act page →Gemini 2.5 Pro: 1M-Token Context with Reasoning
Google releases Gemini 2.5 Pro with a 1-million-token context window and integrated reasoning chains, achieving top scores on coding and math benchmarks. Enables whole-codebase and whole-document reasoning as a standard product feature.
Google DeepMind Gemini →Anthropic Claude 4: Extended Multi-Day Task Execution
Anthropic releases Claude 4 with documented support for multi-day autonomous task execution, persistent memory, and multi-agent orchestration. The model specification explicitly addresses AI autonomy, loyalty, and when agents should pause for human input.
Anthropic Claude 4 announcement →FTC AI Voice Cloning Fraud Rule Takes Effect
The FTC's final rule explicitly prohibiting AI voice cloning for fraud and unauthorized celebrity voice replication takes effect. First U.S. rule specifically targeting the synthetic media pipeline from model output to consumer harm.
FTC AI consumer protection →UN AI Governance Resolution: Data Sovereignty Provisions
The UN General Assembly adopts a non-binding AI governance resolution emphasizing data sovereignty — the principle that nations retain rights over data generated by their citizens even when processed by foreign AI systems. Accelerates data localization requirements globally.
UN AI Advisory Body →Reasoning-Native Model Generation Becomes Standard
OpenAI ships o3-pro and the broader 'reasoning-native' model generation becomes the industry standard: all major frontier models now include extended internal chain-of-thought by default. The separation between 'chat' and 'reasoning' model categories collapses.
OpenAI o3 family →NIST AI RMF Agentic Supplement Published
NIST publishes a supplemental profile to its AI RMF specifically addressing agentic systems — covering multi-agent trust hierarchies, human-in-the-loop thresholds, and autonomous action logging requirements. First official U.S. government guidance on agentic AI risk.
NIST AI RMF →First EU AI Act High-Risk Enforcement Actions
EU national market surveillance authorities issue the first formal enforcement notices under the AI Act's high-risk provisions, targeting biometric workplace monitoring systems in Germany and France. Establishes enforcement precedent for HR and access-control AI applications.
EU AI Act legislative tracker →Multi-Agent Frameworks: AI Hiring AI Goes Mainstream
Major enterprise software vendors ship multi-agent orchestration platforms where AI managers spawn, assign, and evaluate AI worker agents. The pattern becomes a standard enterprise architecture, raising new questions about accountability chains and audit trails.
Anthropic: Building effective agents →G7 AI Code of Conduct Enters Force with Compliance Teeth
The G7 Hiroshima AI Code of Conduct — originally voluntary — is formalized with mutual-recognition agreements: companies in G7 markets must document frontier model safety evaluations or face market access restrictions. First multilateral AI governance instrument with real trade leverage.
G7 Hiroshima AI Process →What 48 Milestones Tell Us
Reading the timeline in sequence, a few things stand out.
The models came first, and everything else — the lawsuits, the executive orders, the safety debates, the agentic frameworks — came in response. Governance has been chasing capability since 2022, and the gap has never fully closed.
The inflection wasn't just ChatGPT — it was ChatGPT arriving at the moment regulators, investors, and the public were finally ready to pay attention. The underlying research had been building for over a decade. What changed was visibility.
And the pattern we see in legal and policy entries is consistent: jurisdictions that moved early (the EU, Italy, China) set the templates. Everyone else followed. By the time enforcement actions arrived in 2025, the legal architecture had been under construction for three years.
For enterprises, the strategic lesson is the same as it has always been: the organizations that built governance while capability was developing didn't have to scramble when enforcement began. They were ready.
That is what we built Audition AI to enable. Not just the AI capability — but the governance layer that makes it sustainable.
Like this content?
Subscribe to our weekly brief for more insights on AI strategy and enterprise governance
Subscribe to Weekly BriefReady to Build AI Governance That Lasts?
Audition AI gives your organization the capability and the governance layer to move fast without moving recklessly.