This is Hamamoto from TIMEWELL.
Source: Stanford AI Index Report
Welcome to the 6th AI Index Report
The AI Index Report 6th edition contains more original data than any previous version. New additions include a chapter on AI public opinion, more granular technical performance analysis, coverage of large language models and multimodal models, trend data on AI legislation across 127 countries, and research on the environmental impact of AI systems.
The report's mission: to provide unbiased, rigorously validated, broadly sourced data to help policymakers, researchers, executives, journalists, and the public develop a more nuanced understanding of the complex AI landscape.
Co-directors Jack Clark and Ray Perrault note: "AI has entered a new era of development. From 2022 to early 2023, major new AI models were released every month—ChatGPT, Stable Diffusion, Whisper, DALL-E 2—capable of an increasingly wide range of tasks from text manipulation and analysis to image generation to unprecedented speech recognition. These systems demonstrate capabilities for generating text, images, and code—and answering questions—that would have been unimaginable ten years ago."
However, these systems are also prone to hallucinations, can be biased, and can be used for harmful purposes—highlighting complex ethical challenges.
Key Finding 1: Industry Has Overtaken Academia
Until 2014, the most significant machine learning models came primarily from academic institutions. Industry then took over. In 2022, academic institutions produced just 3 notable ML models; industry produced 32. Building state-of-the-art AI systems increasingly requires massive data, compute, and capital—resources industry has in far greater supply than nonprofits or universities.
Key Finding 2: LLMs Are Getting Much Larger and More Expensive
GPT-2, released in 2019 and considered the first broadly accessible large language model, had 1.5 billion parameters and cost approximately $50,000 to train. PaLM, a leading LLM released in 2022, has 540 billion parameters and cost approximately $8 million to train. PaLM is roughly 360x larger and 160x more expensive than GPT-2. This scaling trend is continuing.
Looking for AI training and consulting?
Learn about WARP training programs and consulting services in our materials.
Key Finding 3: AI Private Investment Declined for the First Time in a Decade
Global AI private investment in 2022 was $91.9 billion—a 26.7% decline from 2021. The total number of AI funding events and the number of newly funded AI companies also fell. Despite this, investment over the decade has grown dramatically: 2022 AI private investment was 18x the 2013 level.
The United States continues to lead global AI investment by a wide margin. US investment in 2022 was $47.4 billion—approximately 3.5x China's $13.4 billion. The US also leads in newly funded AI companies: 1.9x the EU and UK combined, and 3.4x China.
The top three AI investment categories in 2022: healthcare/medical ($6.1B), data management/processing/cloud ($5.9B), and fintech ($5.5B).
Key Finding 4: AI Adoption Rate Is Plateauing—But Adopters Still Lead
According to McKinsey's annual survey, the share of companies that have adopted AI roughly doubled from 2017 to 2022, but has recently leveled off at 50-60%. Critically, organizations that have adopted AI consistently report meaningful cost reductions and revenue gains.
The most commonly adopted AI functions: Robotic Process Automation (39%), Computer Vision (34%), Natural Language Text Understanding (33%), and Virtual Agents (33%).
GitHub's research on Copilot found that 88% of users became more productive, 74% could focus on more satisfying work, and 88% completed tasks faster.
Key Finding 5: Policy Activity Is Surging
The number of bills mentioning "artificial intelligence" that passed into law increased from 1 in 2016 to 37 in 2022 across 127 countries. AI mentions in parliamentary records across 81 countries have increased approximately 6.5x since 2016.
Key Finding 6: AI Has Environmental Costs—and Benefits
Training BLOOM (a large language model released in 2022) emitted 25x more carbon than a single air passenger flying one-way from New York to San Francisco. However, newer reinforcement learning models like BCOOLER are demonstrating AI's ability to optimize energy use—meaning AI can be both a cause of and a solution to environmental impact.
Key Finding 7: Public Attitudes Vary Widely by Country
In an IPSOS 2022 survey, 78% of Chinese respondents agreed that AI products and services have more benefits than drawbacks—the highest of any surveyed country. Saudi Arabia (76%) and India (71%) followed. Among Americans, only 35% agreed.
Male respondents were more likely than female to view AI positively. Only 27% of global survey respondents said they would feel safe in a self-driving car.
Key Finding 8: AI Misuse Incidents Are Accelerating
The AIAAIC database tracking AI ethics misuse incidents shows a 26x increase since 2012. Notable 2022 incidents included a deepfake video of Ukrainian President Zelensky appearing to surrender, and US prisons using call monitoring technology on inmates.
The Research Team
The AI Index Steering Committee is co-directed by Jack Clark (Anthropic, OECD) and Raymond Perrault (SRI International), with contributors from Stanford, Google, Hugging Face, and many other institutions.
Citation: Nestor Maslej et al., "The AI Index 2023 Annual Report," AI Index Steering Committee, Institute for Human-Centered AI, Stanford University, April 2023. Licensed under CC BY-ND 4.0.
TIMEWELL AI Consulting
TIMEWELL supports business transformation in the AI agent era.
Our Services
- ZEROCK: High-security AI agent running on domestic servers
- TIMEWELL Base: AI-native event management platform
- WARP: AI talent development program