The chart that scared a lot of people — and what it's actually saying
A graphic from Anthropic's recent report on AI and the labor market started circulating widely this month. At first glance, it looks terrifying. It shows that AI has "theoretical capability" to perform over 80% of individual job tasks across categories as broad as legal work, finance, management, arts and media, and office administration.
If you took that at face value, the conclusion seems obvious: AI is going to eliminate most jobs. And fast.
But when you actually dig into how Anthropic arrived at those "theoretical capability" numbers, the picture changes significantly — and the methodology behind those blue bars reveals something important about the entire category of AI-replaces-jobs research.
What Anthropic actually measured — and what they didn't
Anthropic's report makes an important distinction between two things: "theoretical capability" (what AI could theoretically do, based on research) and "observed exposure" (what AI is actually being used for right now).
The observed exposure data — what's actually happening — is Anthropic's own contribution. That's real, based on how people actually use Claude and other LLMs in their work. The numbers here are much smaller than the theoretical numbers. For example, the Computer & Math category shows 94% theoretical exposure but only 33% actual observed coverage. Office & Administrative Support shows 90% theoretical but only 25% actual.
The "theoretical capability" numbers, though, are where it gets interesting. Anthropic didn't measure this themselves using current AI models. They borrowed the numbers from a 2023 research paper co-authored by OpenAI, OpenResearch, and the University of Pennsylvania — a paper based on GPT-4 performance back in 2023, not current models, and not Anthropic's own Claude.
How that 2023 paper arrived at its numbers
The 2023 study started with O*NET — a US government database that breaks down every job into extremely granular individual tasks. Researchers then went through those tasks and judged whether "the most powerful OpenAI large language model" at the time could reduce the time needed for each task by at least 50% with equivalent quality.
If current AI couldn't do it, they also asked: could "anticipated LLM-powered software" in the future achieve that time saving? That second question is where the speculative element enters. Many of the "theoretically capable" tasks are only theoretically possible with future AI tools that don't yet exist — not with anything you can use today.
The labelers who made these judgments — the people who decided whether AI could or couldn't perform each task — were not the professionals who actually do those jobs. They were workers hired through an annotation platform. People judging whether AI could perform complex legal research or nuanced financial analysis may not have had deep familiarity with what those tasks actually involve day to day.
When you add those caveats together — borrowed from a 2023 study, based on older GPT-4, includes speculative future tools, judged by annotators not domain experts — the scary blue bars start looking a lot less like a prediction and more like an educated (and significantly outdated) upper-bound estimate.
What the observed data actually shows
The more interesting part of Anthropic's report is the observed exposure data — what they can actually see happening in how people use AI at work right now.
Occupations with higher current AI usage are projected to see slower job growth through 2034, according to US Bureau of Labor Statistics forecasts. Workers in the most AI-exposed professions tend to be older, more educated, and better paid. Younger workers aged 22 to 25 are seeing slower hiring rates in AI-exposed fields — not more job losses once employed, but fewer new people being brought in to start with.
Crucially, Anthropic found no systematic increase in unemployment for workers in highly AI-exposed jobs since late 2022, when ChatGPT launched and the current AI wave began. Jobs aren't disappearing en masse. What seems to be happening is more subtle: teams are handling more work without hiring as many new people, and entry-level hiring in certain roles is slowing down.
What this means for Indian professionals — the honest assessment
India's tech industry employs millions in jobs that score high on AI exposure metrics: software development, data analysis, content writing, customer support, legal services, financial analysis, and back-office operations. Indian IT services companies built their business model on providing human labor for tasks that are increasingly being automated.
The honest answer to "will AI take your job?" is: probably not tomorrow, possibly not even in five years in the way the scary charts suggest. But the entry-level version of your job — the parts that new graduates used to do to build experience — those parts are under real pressure right now.
A junior developer who spends time writing boilerplate code? AI does that now. A junior analyst who spends time pulling standard reports and formatting data? AI does that too. The concern isn't that experienced professionals get replaced. It's that the traditional path — start junior, build skills, move up — gets disrupted because the junior work disappears or shrinks significantly.
For Indian students entering the job market, this means skills that AI cannot easily replicate become more valuable: client communication, complex problem definition, creative judgment, domain expertise combined with technical skill, project management, and understanding contexts that require cultural or institutional knowledge that isn't in training data.
The bigger lesson: how to read AI job displacement research
Every few months a new report drops claiming AI will replace X% of jobs by year Y. Most of these reports have the same issue as the 2023 paper Anthropic's theoretical numbers are based on: they measure whether AI could theoretically help with a task, not whether it will actually replace the human doing that job.
There's a big difference between "AI can help a lawyer research case precedents 40% faster" and "AI will replace lawyers." The first is already true. The second ignores that lawyers do dozens of things that require judgment, relationships, accountability, and contextual understanding that AI doesn't reliably provide.
The gap between theoretical AI capability and actual observed usage — which Anthropic's own report shows is enormous — is the space where human professionals currently live and work. That gap is narrowing. Slowly in some areas, faster in others. But it's not collapsing overnight the way alarming charts imply.
TamilTech's take
Anthropic deserves credit for being transparent about the limitations of their own data. Most companies would have just led with the scary theoretical numbers and let the viral attention do its work. Instead, they explicitly distinguished between what AI theoretically can do and what it's actually doing — and the gap is large. The more important finding in their report isn't the theoretical capability chart. It's the hiring data showing entry-level jobs in AI-exposed fields are slowing down. That's a real, present-tense effect you can point to with data. For Indian students and early-career professionals, that's the number to take seriously — not the scary 80% theoretical coverage that's based on three-year-old research and speculative future tools. Build skills that are genuinely difficult to automate. The mid-career and senior roles aren't going anywhere fast. The entry door is getting narrower.
Comments (0)
Be the first to comment!