Will AI actually replace jobs?

This question now has internal data from a company worth examining closely.

Anthropic has shared research on how their work and hiring have changed after heavily adopting AI Agents: AI Agent work hours have greatly exceeded employee hours, yet the company continues to grow; the share of entry-level tech hires has noticeably dropped.

At the same time, Anthropic is reportedly pushing a new frontier AI safety bill in the US Congress, while the latest IMF meeting statement highlights AI’s uncertain impacts on finance and the economy as potential risks.

These three updates collectively address one question:

When AI is no longer just a chat tool but actively participates in work and reshapes businesses and the economy, what new rules must humans establish?

1. Anthropic’s Internal Employment Study: AI Agent Work Hours Six Times that of Employees—What Does It Mean?

On October 16 (US time), Anthropic released a study titled How employment is changing at Anthropic, analyzing changes in work style, recruitment, and job roles after heavy AI adoption.

The study used internal employment records, AI usage data, work output, employee surveys, meeting notes, and recruitment data, supported by interviews with 17 employees.

The most notable finding is the scale of AI Agent usage.

By August 2026, the total interactive AI Agent hours logged by employees had reached roughly six times the estimated employee work hours.

However, this figure should not be interpreted as "AI replacing six times the number of human employees."

The AI hours are estimated based on Agent execution logs; human hours are based on employee count times 40 hours per week.

These are different units of work output and cannot be directly converted into labor or salary savings.

What truly stands out is the shift in hiring structure

Since early 2025:

  • The share of entry-level tech roles among tech staff dropped from 21% to 14%.
  • Among new tech hires, entry-level roles decreased from 27% to 9%.
  • Some engineering and research teams have deprioritized hiring for entry-level tech positions.

What does this mean?

Previously, companies often needed numerous entry-level engineers to write simple code, find bugs, maintain tools, and handle basic development tasks.

With some tasks handed over to AI Agents, companies now prioritize problem-solving, interdisciplinary collaboration, and autonomy when recruiting new talent.

That said, Anthropic has not stopped hiring engineers.

The report notes technical hiring continues to increase, though other departments are growing faster, causing the entry-level tech share in recruitment to decline.

This is an internal insight from a highly AI-driven company and does not imply identical hiring trends across all firms or countries.

Are people busier the more AI does?

The study found another interesting trend.

As employees widely use Claude Agents, generating more code and documents, human attention becomes a new bottleneck.

Some employees no longer do every task themselves but spend more time planning Agent work, reviewing outputs, coordinating Agents, and tackling decisions only humans can make.

The study estimates that even in non-technical teams, reviewing all interactive Claude Agent outputs might consume over half a workweek.

Thus, new work managing Agents is emerging.

For example, building monitoring dashboards, writing rules for Agents, and organizing result reviews by other Agents.

AI hasn’t removed management work but has changed what is being managed.

What does this mean for regular workers?

If you are a newcomer to the workforce, the key focus isn’t "Will my job title disappear?"

It’s whether basic tasks traditionally assigned to newcomers are increasingly handled by AI.

Simply following fixed instructions for single tasks will likely become insufficient.

More valuable skills include understanding real problems, judging AI output reliability, and translating work results into actionable decisions.

Still, Anthropic’s study does not prove AI necessarily causes overall job losses or that any job’s efficiency will improve sixfold.

2. Anthropic Reportedly Proposes AI Safety Bill—Can the Government Stop Model Releases?

Another key update comes from the US Congress.

On October 16, Reuters, citing Politico, reported that Anthropic representatives have submitted a draft frontier AI safety and cybersecurity regulation to US lawmakers.

This draft is not yet law or an announced regulation.

It reveals an important direction in current AI developer discussions:

If a future AI model can cause extremely serious harm, under what conditions should the government intervene?

What thresholds does the draft set?

According to the draft obtained by the media, government powers to block release or use of certain frontier models will be tightly limited.

One threshold includes risks that are foreseeable, imminent, and could cause:

  • Over $1 billion in damages; or
  • More than 50 deaths or serious injuries.

The risks must also involve major specific harms, such as biological weapons, models escaping developer control, or cyberattacks.

This does not mean any faulty AI could be blocked.

The draft targets a few frontier models that might cause massive-scale harm.

Who decides if a model is dangerous?

The draft requires frontier model developers to undergo independent assessments and register these assessment bodies with the US Department of Commerce.

If serious risks are confirmed, these bodies must report to authorities following draft procedures.

An Anthropic spokesperson told Reuters these ideas align with the company’s previously proposed AI policy framework, aiming to empower government intervention in extremely high-risk scenarios.

However, formal legislation, negotiation, congressional votes, and implementation details remain uncertain.

Thus, it’s incorrect to say "The US government already has the power to stop AI models."

Why is this important?

Large AI companies traditionally decide when and how to launch models and conduct safety testing themselves.

Future clear external assessments, reporting, and legal processes may impose more outside checks on developers’ release decisions.

Yet balancing prevention of major risks with avoiding overreach in regulation remains a key issue.

This is why draft limits, assessor independence, appeal processes, and whether to establish a regulatory body are equally important.

3. IMF’s Latest Statement: AI Can Boost Productivity, But Why Might It Amplify Financial Risks?

The third update comes from the global economic level.

On October 17, the International Monetary and Financial Committee (IMFC), the IMF’s main policy advisory panel, issued a chairman’s statement after meeting in Bangkok.

The statement highlights that advanced intelligent technologies are driving investment and trade and could raise productivity in the future.

However, AI and similar technologies’ impacts on economies and financial systems remain uncertain.

With already high public debt, refinancing pressures, and rising government bond yields in major economies, such uncertainties could increase the overall financial system’s fragility.

Is the IMF predicting an AI bubble burst soon?

No.

The statement does not forecast a financial crisis on a specific day nor declare the AI investment boom will necessarily collapse.

It provides a risk assessment.

Rapid AI advancement might spur new investments, industries, and productivity growth, but discrepancies between investment scale, financing costs, and actual returns could pressure financial markets.

Especially amid high debt, rising interest rates, and geopolitical uncertainty, financial markets may be less resilient to disappointments.

The statement also notes varying impacts of technological change across countries, with low-income and vulnerable economies likely falling further behind.

How should the average person understand this?

For regular workers, ongoing AI industry growth may create new jobs and business opportunities.

But for business owners and investors, it’s unwise to assume AI-linked investments will always sustain growth.

More important is whether companies can prove AI delivers real productivity gains and whether large-scale investments rest on sustainable financial foundations.

The IMF’s message is that technological progress brings both growth opportunities and new systemic risks.

Risks should be managed, but warnings are not proof a financial crisis has occurred.

4. What’s the Common Significance of These Three AI Updates?

Viewed together, these stories show the speed at which AI is entering various systems.

Within companies, AI Agents are reshaping work division and hiring practices.

On the policy front, developers and lawmakers are discussing external oversight for extremely high-risk models.

At the global economy level, AI is now a critical factor in investment, productivity, and financial stability discussions.

They are different issues and don’t directly prove that "AI will cause massive unemployment" or "the AI bubble is about to burst."

But these developments show:

We can no longer evaluate AI solely by model test scores, generation speed, or demo videos.

We must also consider:

  • How AI is changing job content and hiring criteria;
  • Who is responsible for evaluating and supervising extremely high-risk models;
  • Whether AI-driven economic gains support the costs of investment.

Conclusion: The Next Competition in AI Isn’t Just Capability But Taking Real-World Responsibility

Perhaps the most important takeaway today is the shift revealed by Anthropic’s internal research.

As AI can produce vast amounts of work, human value no longer lies just in coding faster or calculating more accurately than AI.

What matters more is knowing which work deserves doing, how to allocate tasks, and which results merit adoption.

Likewise, the AI industry’s growth depends not only on showing stronger models but also on trusted safety systems and verifiable real-world outcomes.

How much AI can do is a technical question; what AI should do and who validates its results are challenges everyone must now face.

Recommended Reading