RSS Amplifier

The Context Window with David Deming · Aug 5, 2026

I read the AI usage reports so you don't have to

Read inOriginalenesdefrptjaittr
0
Sign in to vote or save

David Deming · The Context Window with David Deming

“I had a front row seat because I helped write one of the papers, and I’ve closely read the others so you don’t have to. The reassuring part is that AI is right now more likely to be a helpful assistant than a job replacement or a killer robot. The less reassuring part is that chatbots, helpful as they are, may already be yesterday’s technology.” — David Deming

One in ten people on Earth used ChatGPT in the past week. So why has almost no one’s job changed?

In this solo episode, David Deming - labor economist and Dean of Harvard College - reads every usage study published by OpenAI, Anthropic, and Google, including the one he helped write, and reports what millions of conversations actually show.

Generative AI reached 56% of US working-age adults within three years, faster than the PC or the internet at the same age. But most people use it for ordinary help with ordinary problems, and the same seven work activities dominate AI use in nearly every occupation. The exception is arriving fast: AI agents that take delegated work and keep going.

Featuring the phone number that explains how researchers study private chats without reading them, and the government database that says podcasting is not part of David’s job. Also, Coke, Pepsi, and maybe Dr Pepper.

Listen on YouTube, Spotify, and Apple Podcasts

I’ve spent the last few solo episodes making arguments about what AI will do to work. This week I’m doing something different: reporting. OpenAI, Anthropic, and Google have each published research on what millions of people actually do with their chatbots, and the studies agree far more than they differ. I read all of them, appendices included, and this episode is my attempt to put the three labs on one page.

A disclosure, which I also make in the episode - I helped write OpenAI’s paper, “How People Use ChatGPT,” on a project I joined before becoming Dean. Nobody at any AI lab has ever paid me, and I have friends at all three.

Let me know what you think at david@thecontextwindow.com.

[00:00] One in ten, this week - ChatGPT has reached roughly 900 million weekly users, faster than any technology before it.

[00:46] A guided tour of the evidence - David previews the usage reports from OpenAI, Anthropic, and Google and promises a best guess at who is winning the AI race.

[02:28] A disclosure - David co-wrote OpenAI’s study, has taken no money from any AI lab, and is not playing favorites.

[03:10] What the OpenAI paper covers - “How People Use ChatGPT” studies consumer ChatGPT only, leaving out enterprise accounts, APIs, and Codex.

[04:02] OpenAI’s Signals website - A public site updates the paper’s analyses monthly and adds state-level and business data.

[04:29] Claude meets the Department of Labor - Anthropic mapped a million Claude conversations onto O*NET, the government’s taxonomy of job tasks.

[04:57] What do people do all day? - David looks up his own occupation code and finds that podcasting is not among his 22 official tasks.

[06:50] Anthropic’s follow-ups - Later reports add API usage and attempt to score conversations for autonomy, expertise, and success.

[07:42] Google builds an ATLAS - Google studies 15 million interactions across the Gemini app, the Gemini API, and AI Mode.

[08:35] Studying chats nobody reads - Automated classifiers strip identifying information and return category labels, so researchers never see raw conversations.

[10:20] Is the classifier right? - With no ground truth, the labs validate against human raters, whose agreement with the model was higher than their agreement with each other.

[12:04] What this data cannot tell us - Company records say little about demographics, non-users, or people who switch between models.

[12:39] Takeaway 1: the fastest adoption in history - Generative AI reached 56% of US working-age adults in three years, against 20% for the PC and 30% for the internet at the same age.

[16:45] Takeaway 2: ordinary help with ordinary problems - Practical guidance, information seeking, and writing account for about 75% of ChatGPT messages, while coding is under 5%.

[19:54] Where AI fits into daily life - Google maps conversations onto the American Time Use Survey and finds education and paperwork heavily overrepresented relative to the time people spend on them.

[23:30] Takeaway 3: broad, but not deep - More than half of doctors, teachers, and lawyers use AI, yet the same seven work activities dominate in nearly every occupation.

[26:30] 70% of jobs, 21% of tasks - Google finds AI useful somewhere in most occupations and close to replacing none of them.

[27:52] Takeaway 4: the medium is the message - The three rival chatbots are used almost identically, so the distinction that matters is chatbot versus agent.

[28:56] From assistance to delegation - In Claude Code, 79% of coding interactions run autonomously, and users with domain expertise get the longest runs.

[30:57] The agent boom - Codex use grew fivefold between January and June 2026, and agents already account for a majority of output tokens among business customers.

[32:58] Four conclusions - Historic adoption, practical help, shallow work integration, and the rise of agents.

Papers

  • Aaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan, and Kevin Wadman, “How People Use ChatGPT” (NBER Working Paper 34255, 2025) - the study of consumer ChatGPT behind the topic shares, the work/non-work split, and the seven work activities discussed in the episode.

  • Alexander Bick, Adam Blandin, and David J. Deming, “The Rapid Adoption of Generative AI” (NBER Working Paper 32966, 2024) - the nationally representative survey comparing generative AI adoption with the early spread of PCs and the internet; the November 2025 figures cited in the episode appear in the authors’ update for the St. Louis Fed.

Reports from the labs
Articles
Data
Further listening
Credits

Host: David Deming, Danoff Dean of Harvard College

Executive Producer: Denise Koller Consulting Producers: Tim Smith and Jonathan Palumbo

Leave a comment

Read the original on thecontextwindowpodcast.substack.com ↗

Comments

Nothing yet. Say the first thing.

    Sign in to join the conversation.