Anthropic just released an "AI Job Theft Report": the higher the education level, the more likely to be "stolen"

Author: Xin Zhiyuan

Your work's “gold content” is being drained by AI. Anthropic's latest report reveals an counterintuitive truth: the more complex the task, measured by years of education, the faster AI accelerates. Compared to direct replacement, a more frightening trend is “deskilling”—AI takes away the joy of thinking, leaving you with only menial tasks. But the data also points to a single solution: understanding human-AI collaboration can increase success rates tenfold. In this era of excess computing power, this is a survival guide you must understand.

Anthropic just released the “Economic Index Report” on their official website yesterday.

The report not only focuses on what people are doing with AI but also on the extent to which AI truly replaces human thinking.

This time, they introduced a new dimension called “Economic Primitives,” attempting to quantify task complexity, required education level, and AI autonomy.

The future of the workplace reflected behind the data is far more complex than simple “unemployment theory” or “utopianism.”

The harder the task, the faster AI completes it

In our traditional understanding, machines excel at repetitive, simple labor, and appear clumsy in fields involving advanced knowledge.

But Anthropic's data presents a completely opposite conclusion: the more complex the task, the more astonishing the “acceleration” brought by AI.

The report shows that for tasks understandable with a high school diploma, Claude can increase work speed by 9 times;

Once the difficulty reaches the college level, this acceleration jumps to 12 times.

This means that tasks that once required hours of human contemplation—white-collar work—are now among the fields where AI “harvests” most efficiently.

Even considering AI's occasional hallucinations and failure rates, the conclusion remains unchanged: the efficiency surge AI brings to complex tasks is enough to offset the repair costs caused by errors.

This explains why current programmers and financial analysts are more dependent on Claude than data entry clerks—because in these high-intelligence-density fields, AI's leverage effect is the strongest.

19 hours: The “New Moore's Law” of Human-AI Collaboration

The most shocking data in this report concerns AI's “durability” (task duration, measured by 50% success rate).

Standard benchmarks like METR (Model Evaluation & Threat Research) suggest that top models (such as Claude Sonnet 4.5) have success rates below 50% when handling tasks requiring 2 hours of human effort.

But in Anthropic's real user data, this time limit is significantly extended.

In commercial API scenarios, Claude can maintain over half success rate on tasks involving 3.5 hours of work.

In the Claude.ai chat interface, this number astonishingly extends to 19 hours.

Why such a huge gap? The secret lies in “human” involvement.

Benchmark tests are AI facing exams alone, but in reality, users break down complex projects into countless small steps, continuously providing feedback to correct the AI's course.

This human-AI collaboration workflow pushes the (measured by 50% success rate) task duration limit from 2 hours to about 19 hours—nearly 10 times.

This may be the future of work: not AI doing everything independently, but humans learning how to steer it through a marathon.

The folding of the world map: the poor learn knowledge, the rich do production

If we elevate our perspective to a global level, we see a clear and somewhat ironic “adoption curve.”

In developed countries with higher per capita GDP, AI has deeply integrated into productivity and personal life.

People use it to code, generate reports, even plan travel itineraries.

But in lower-GDP countries, Claude's main role is as a “teacher,” with many uses focused on homework and educational tutoring.

Beyond wealth disparities, this also reflects a technological gap.

Anthropic mentions they are working with the Rwandan government to help people move beyond mere “learning” and into broader applications.

Because without intervention, AI could become a new barrier: wealthy regions use it to exponentially boost output, while less developed areas still rely on it for basic education.

Workplace concerns: The ghost of “deskilling”

The most controversial and warning-worthy part of the report is the discussion on “deskilling.”

Data shows that the tasks covered by Claude currently require an average of 14.4 years of education (equivalent to a college diploma), higher than the 13.2 years needed for overall economic activity.

AI is systematically removing the “high-intelligence” parts of jobs.

For technical writers or travel agents, this could be disastrous.

AI takes over analyzing industry trends and planning complex itineraries—tasks that require “brainpower,” leaving humans possibly only with sketching and administrative work like invoicing.

Your job still exists, but its “gold content” is being drained.

Of course, there are beneficiaries too.

For example, real estate managers, after AI handles tedious administrative tasks like bookkeeping and contract comparison, can focus on high-emotional-intelligence activities like client negotiations and stakeholder management—this is a form of “upskilling.”

Anthropic cautiously states that this is just a projection based on current conditions, not an inevitable prophecy.

But the alarm is real.

If your core competitive advantage is merely handling complex information, you are at the eye of the storm.

Is productivity returning to its “golden age”?

Finally, let's look at the macro perspective.

Anthropic revised their forecast for US labor productivity.

After accounting for possible AI errors and failures, they estimate that AI will drive annual productivity growth of 1.0% to 1.2% over the next decade.

This is about one-third less than their previous optimistic estimate of 1.8%, but do not underestimate this 1 percentage point.

It is enough to bring US productivity growth back to the levels seen during the late 1990s internet boom.

Moreover, this projection is based on the capabilities of models as of November 2025. With the arrival of Claude Opus 4.5 and the gradual dominance of “enhanced mode” (where people collaborate more intelligently with AI rather than fully delegating work), this figure has enormous upward potential.

Conclusion

Reviewing the entire report, what is most impressive is not how powerful AI has become, but how quickly humans are adapting.

We are experiencing a migration from “passive automation” to “active reinforcement.”

In this transformation, AI is like a mirror—it takes over tasks that require high education but can be completed through logical deduction, pushing us to find those values that cannot be quantified by algorithms.

In this era of excess computing power, humanity's most scarce ability is no longer finding answers, but defining questions.

References: