AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

AI now writes a large share of the code merged at Anthropic, but the company says engineers still direct and review that work. Hiring data and learning experiments point to a changing path into software development, while evidence that AI consistently makes developers faster remains limited.

A review published by Thorsten Meyer AI on November 18, 2026, argues that students should still learn to code as AI tools take on more software writing, but that the skills taught need to change, a question explored in rethinking education in the AI era. The review draws on company disclosures, hiring data and research experiments; its evidence suggests that directing, checking and debugging code remain important, while the scale of productivity gains and effects on early-career jobs are still uncertain.

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

Anthropic says that in May 2026, Claude authored more than 80% of the code merged into the company’s codebase. It reported that the share had been in the low single digits before Claude Code launched in February 2025. Anthropic says engineers direct and review the generated code. The figure describes one AI company’s own operations, not software development across the industry.

Anthropic also reported that its typical engineer merged eight times as much code per day in the second quarter of 2026 as in 2024. The company cautioned that lines of code measure quantity rather than quality and that the figure is “almost certainly an overstatement” of the productivity gain. In a March 2026 employee poll, the median estimate was about four times as much output with AI; Anthropic said the actual uplift was likely lower.

Other evidence in the review gives a more mixed picture. A METR randomized trial in early 2025 found that experienced developers took 19% longer to complete tasks with AI, even though they believed the tools had made them 20% faster. METR later said newer data offered “only very weak evidence” about the size of any speed-up by early 2026. The review also cites an Anthropic learning experiment in which mostly junior developers who used AI while learning a new library scored 50% on a follow-up quiz, compared with 67% among developers who coded by hand. Debugging showed the biggest gap.

At a glance
analysisWhen: Published November 18, 2026; the underl…
The developmentA November 2026 review of AI coding disclosures, hiring figures and developer studies argues that coding education should put more emphasis on specifying, reviewing, testing and debugging software.
Bar chart titled Developers felt faster with AI. They were slower. METR trial in early 2025 with 16 experienced open-source developers and 246 real tasks: before the study they expected AI to cut their time per task by 24%; afterwards they believed it had cut it by 20%; measured, tasks took 19% more time with AI allowed. A highlighted note: plus 19% time; METR thinks tools sped people up more by early 2026 but cannot yet say by how much
Expectation, belief and measurement in METR's developer trial. Sources: Becker et al. (METR), arXiv:2507.09089, p. 1; METR, February 24, 2026. Chart: Thorsten Meyer AI.

The Skills Behind AI-Written Code

The evidence points to a shift in where effort goes when an AI tool drafts more of a program. Someone still has to specify what the software should do, judge whether the result fits the existing system, test it and take responsibility for shipping it. Those tasks rely on understanding code, even when a person types fewer lines.

That has consequences for both education and employers. If learners use AI to produce working answers before they can explain or debug them, they may miss practice with the reasoning needed to catch mistakes. The quiz result in Anthropic’s experiment raises that concern, though a single experiment does not establish how every student learns with AI. Courses and workplace training may need to teach students to use these tools while still requiring them to explain, test and repair the code.

Hiring figures add another reason to watch how training pathways change. The review says 4.5% of US software development job postings on Indeed were entry-level in the first quarter of 2026, while 69.3% were senior. Stanford researchers cited in the review found employment among 22- to 25-year-olds in AI-exposed jobs 19% below where it would otherwise have been. These figures do not establish that AI caused the hiring pattern, but they sharpen questions about how new developers gain experience if fewer entry-level roles are available.

Products Worth Considering

Jackery Explorer 2000 v2 Portable Power Station, 2042Wh, 2200W

Jackery Explorer 2000 v2 Portable Power Station, 2042Wh, 2200W

DJI Mini 4K Drone Bundle with Strobe Light, Landing Gear, and Propeller Holder Strap

DJI Mini 4K Drone Bundle with Strobe Light, Landing Gear, and Propeller Holder Strap

As an affiliate, we earn on qualifying purchases.

From Code Forecasts to Hiring Data

On March 10, 2025, Anthropic chief executive Dario Amodei told the Council on Foreign Relations that AI would be writing 90% of code in three to six months and “essentially all of the code” within twelve months. He also said programmers would still need to specify what was being built, how it fit with other code and whether the design was secure. The subsequent Anthropic disclosure shows substantial use of Claude inside that company, but it does not verify Amodei’s forecast for the software industry as a whole.

The review also cites a statement by Google chief executive Sundar Pichai in October 2024 that AI generated more than a quarter of new code at Google, which engineers then reviewed and accepted. It notes a February 2026 METR estimate that Claude Code authored about 4% of GitHub commits. These measures cover different companies, tools and periods, so they cannot be treated as a direct comparison.

Indeed’s software development postings in the United States were about 23% higher year over year in November 2026, according to the review, after reaching a low in May 2025. That movement shows postings had risen from the stated low; by itself, it does not show that hiring had returned to an earlier level or explain the change. The article’s evidence combines company reporting and research rather than a single study of coding education.

““The programmer still needs to specify” what is being built, how it fits with other code and “whether this is a secure design or an insecure design.””

— Dario Amodei, Anthropic chief executive, at the Council on Foreign Relations, March 10, 2025

What the Evidence Cannot Settle

The review does not establish how widespread Anthropic’s code-writing figures are among employers. Anthropic builds Claude and reports on its own use, while Google’s statement covers a different company and period. The measures cited for GitHub commits, code merged and code generated are not interchangeable.

The effect on productivity also remains unresolved. METR’s 2025 trial involved 16 experienced developers, and its later update described its newer speed evidence as weak. The available figures do not show whether results apply to beginners, other kinds of software work or later versions of AI tools.

The education and labor findings have limits too. The learning experiment’s quiz scores do not show how long any difference lasts or whether other teaching approaches change the outcome. The cited employment and job-posting figures describe patterns, but the material does not establish that AI alone caused them. The review does not specify how the Stanford estimate’s counterfactual was constructed, so its “19% below where it would have been” figure should be read as the researchers’ estimate, not a directly observed count of jobs lost to AI.

How Coding Courses May Adapt

The review offers a direction for educators, rather than a confirmed policy change: keep teaching programming while giving students practice with reading, specifying, testing and debugging code. That includes asking learners to explain AI-generated work, find and correct errors, and assess how a change affects the rest of a system. The source does not identify a particular curriculum or institution that has adopted this approach.

For employers and students, the next useful evidence will include hiring data over time, controlled studies involving a wider range of developers and clearer measures of software quality alongside output. Newer AI tools may change the results, so findings should be tied to the dates and tools studied. Until then, the available evidence supports neither abandoning coding education nor assuming that AI use automatically makes developers faster or better.

Key Questions

Does the review say students should stop learning to code?

No. It argues that students should continue learning to code, with greater attention to directing, reviewing, testing and debugging AI-generated work.

How much code did Anthropic say Claude wrote?

Anthropic reported that Claude authored more than 80% of the code it merged in May 2026. This is the company’s figure for its own codebase, not an industry-wide estimate.

Has AI been shown to make software developers faster?

The evidence cited is mixed. In a METR trial in early 2025, experienced developers took 19% longer with AI, while later METR data offered only weak evidence about the size of any speed-up by early 2026.

What did the AI learning experiment find?

In an Anthropic experiment, mostly junior developers who used AI to learn a new library scored 50% on a follow-up quiz, compared with 67% for those who coded by hand. The review says the largest gap was in debugging.

Does the hiring data prove AI is removing entry-level jobs?

No. The review cites a low share of entry-level software postings and a Stanford estimate about employment among young workers in AI-exposed jobs. Those figures describe labor-market patterns; the material does not show that AI alone caused them.

Source: Thorsten Meyer AI

You May Also Like

Infrastructure Drift Is a Bigger Risk Than Most Teams Notice

AIThis post was created with the assistance of artificial intelligence (AI).Infrastructure drift…

Threlmark: Disk Is the Contract

Threlmark launches a new approach where the roadmap is a plain JSON file on disk, enabling open, interoperable, and durable planning tools.

Staging Parity Is the Deployment Fix Teams Keep Avoiding

With proper staging parity, teams can drastically reduce deployment failures, but many overlook this crucial step—discover why it’s worth prioritizing to ensure smooth releases.

KVM Switches Make Shared Desks Better Than Most Devs Expect

Discover how KVM switches can transform shared desks, boosting efficiency and organization—find out what makes them a game-changer for developers.