Data Visualization · 2025
The
Boundary
Boundary
As AI crossed human-level benchmarks, jobs began to disappear. With CEO Sam Altman predicting AI will automate up to 40% of jobs globally by 2030, the question is no longer whether the boundary moves, but what it leaves behind. This project maps that crossing: the rise, the displacement, and what remains irreducibly human.
Data sources ·
Panel I: Epoch AI Notable AI Models · published model cards · Hugging Face Open LLM Leaderboard ·
Panel II: Frey & Osborne (Oxford, 2017) · Tomlinson et al. (Microsoft Research, 2025) · U.S. BLS OES 2023 ·
Panel III: Tomlinson et al. (Microsoft Research, 2025), Table 2 · 200K Bing Copilot conversations
Panel I · Cold / Analytical
The Rise
A test called MMLU asks AI questions across 57 subjects: history, math, medicine, law. Higher scores mean the AI got more right. Each dot here is one AI model. The white ring means the model's code is freely available (open-source). Hover any dot for details.
"Late 2023 was the moment AI models started scoring higher than the average human expert on this test. Within 14 months, open-source models caught up to the big tech labs. By 2025, the test had become too easy to tell models apart. The competition moved on to harder challenges."
Filter:
Org:
Sources: Epoch AI Notable AI Models Dataset · published model cards · Hugging Face Open LLM Leaderboard · Human expert average on MMLU: 89.8% (Hendrycks et al., 2021)
Panel II · Predicted vs. Observed
The Displacement
Each bubble is a U.S. job. Position from left to right shows how exposed that job is to AI. Position from bottom to top shows how much the job pays. Bigger bubbles mean more people work in that job. Hover any bubble for details. Use the View buttons below to compare what experts predicted in 2017 with what's actually happening in 2025.
"In 2017, two Oxford economists predicted AI would hit the lowest-paid jobs hardest. Cashiers, retail workers, and data entry clerks ended up in the dangerous bottom-right corner: most likely to be automated, least able to afford retraining."
View:
Highlight:
Sources: Frey & Osborne, "The Future of Employment" (Oxford, 2017) · U.S. Bureau of Labor Statistics OES (2023) median annual wage · Employment in thousands
Panel III · The Edge
The Asymmetry
Microsoft researchers read 200,000 real conversations between people and Bing Copilot. For each one, they noted two things: what the person was trying to do, and what the AI ended up doing. Those don't always match. Each bar below is a kind of work. The longer the bar, the bigger the gap.
"AI is great at explaining, teaching, and advising. But the moment work crosses into the physical, financial, or in-person world, the human still has to do it. The bars on the right show the gap. The bars on the left show where AI is now quietly doing the work itself."
Sources: Tomlinson, Jaffe, Wang, Counts, Suri (Microsoft Research, 2025), "Working with AI: Measuring the Applicability of Generative AI to Occupations," Table 2 · Work activity definitions from O*NET Intermediate Work Activities (IWAs) · Ratios computed from user-goal-share vs. AI-action-share across 200K anonymized Bing Copilot conversations, Jan–Sep 2024
Closing · What this means
The Boundary, Read Together
"Three panels, one story: AI got smarter than the average expert in 2023. The way it's actually being used in 2025 doesn't match what experts predicted ten years ago. And the gap between what AI can do and what it actually does is where human work still lives."
Finding 1 · The Rise
AI got smarter than experts in late 2023.
In November 2023, Google's Gemini Ultra was the first AI to beat the average human expert on a 57-subject knowledge test. By April 2025, OpenAI's o3 was scoring 95%. Open-source models (the kind anyone can download and run) caught up to the big labs in just over a year. The question now isn't whether AI is capable. It's where it gets used.
Finding 2 · The Displacement
The 2017 predictions were wrong.
Two Oxford economists predicted in 2017 that AI would hit cashiers, retail workers, and data entry clerks the hardest. By 2025, Microsoft's data on real AI use shows the opposite: AI is being used most by writers, translators, salespeople, and analysts, the educated knowledge workers everyone thought were safe.
Finding 3 · The Asymmetry
AI explains. Humans act.
In 200,000 real conversations, people asked AI to help them buy things, transact money, and do physical tasks 118 times more often than AI did those things itself. The moment work crosses into the real world (clicking purchase, signing the form, going to the appointment), humans have to step in.
What to watch next
"Agent" AI that can take actions. Right now AI mostly answers questions. Companies are racing to make AI that can actually buy things, fill out forms, and book appointments on your behalf. When that lands, the right-side bars in Panel 3 should shrink fast.
What happens to office wages. If AI helps most with knowledge work, those jobs get cheaper to do. Watch whether software, marketing, and finance pay growth slows down compared to healthcare and trades over the next few years.
Tests getting too easy. The MMLU test in Panel 1 has basically been beaten. Researchers are now using harder tests around reasoning and reliability. The chart's vertical axis will need a new measuring stick within a year.
Helping vs. replacing. Microsoft's 2026 reports show AI is starting to replace work more often than just helping with it. The split shown in Panel 3 is a snapshot, and it's already moving toward the left side.
Who's missing from this data. Microsoft's data only includes people using Copilot. Workers without access to AI (and there are many) don't appear at all. Some "low-AI" jobs may just be jobs where AI hasn't reached people yet.
Open-source coming from anywhere. DeepSeek, the Chinese AI lab in Panel 1, reached the top of the field from outside the U.S. tech giants. That's the future: powerful AI built and released by anyone, anywhere.
"The boundary between human work and AI work isn't a wall. It's a moving line. The honest answer to 'what will AI do to our jobs' is: we are watching it happen right now, and the picture from 2017 is already wrong."
Project: The Boundary · AI, Work & What Remains · 2020–2025