🔍 Read the full analysis: What AI Means For Coding Education And Professional Skills on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
AI is taking on more code-writing, but current evidence does not show that coding education has become unnecessary. Studies find uncertain productivity gains, a smaller share of entry-level software postings, and weaker learning outcomes in one experiment when junior developers relied on AI.
AI is writing a growing share of software code, but evidence available in 2026 does not show that learning to program is no longer worthwhile. Anthropic says Claude authored more than 80% of code merged into its own codebase in May 2026, while studies of developer productivity and AI-assisted learning report mixed or cautionary results. The findings matter to students and employers deciding which skills to teach and hire for as entry-level opportunities narrow.
Anthropic’s figures come from its own software operation, not a representative sample of employers. The company said the share of merged code authored by Claude rose from the “low single digits” before its Claude Code tool launched in February 2025 to more than 80% in May 2026. It also reported that its typical engineer merged eight times as much code per day in the second quarter of 2026 as in 2024. Anthropic cautioned that lines of code measure quantity rather than quality, making the output figure “almost certainly an overstatement” of productivity gains.
Independent evidence on speed is less conclusive. In a randomized METR trial conducted in early 2025, 16 experienced open-source developers completed real tasks in projects they knew. They took 19% longer when allowed to use the AI tools tested, despite expecting the tools to reduce their time and later believing they had made them faster. METR’s follow-up data, reported in February 2026, suggested possible improvements but had confidence intervals that included no effect. The group said the newer results were only “very weak evidence” about the size of any speed-up.
Hiring and education data raise separate concerns. Indeed data cited in the report show that 4.5% of U.S. software-development job postings were entry-level in the first quarter of 2026, while 69.3% were senior. Stanford researchers found employment among 22- to 25-year-olds in AI-exposed jobs was 19% below the level they estimated it otherwise would have reached. Those measures do not establish that AI alone caused the change. Meanwhile, Indeed postings for U.S. software-development jobs were up about 23% year over year in November 2026, after reaching a low in May 2025.
The Skills Behind AI-Written Code
The shift changes what coding education and professional development may need to emphasize. If AI generates more of the first draft, workers still need to define requirements, understand how a change fits into a larger system, review output, test behavior and identify security problems. These responsibilities are not proof that every employer will hire the same number of programmers; they indicate that writing code is only one part of software work.
For new entrants, the combination of a smaller share of entry-level postings and uncertain productivity gains makes the route into the field harder to read. Employers may expect candidates to contribute with AI tools, but the evidence does not establish how much this has changed hiring decisions. Students who learn only to prompt a code generator may also miss the fundamentals needed to spot mistakes. The practical issue is not simply whether AI can produce code, but whether a person can judge what it produces and take responsibility for the result.
One Anthropic learning experiment offers a specific warning, not a final verdict. Mostly junior developers learning a new library scored 50% on a follow-up quiz after using AI, compared with 67% among those who coded by hand; the largest gap concerned debugging. The result suggests that immediate task completion and durable understanding can diverge. It does not show that all AI-assisted instruction harms learning or that hand-coding is always superior.
Products Worth Considering

Sonos Era 300 Wireless Smart Speaker with Dolby Atmos, Black

Bose QuietComfort Wireless Bluetooth Noise-Cancelling Headphones, Cypress Green

Logitech MX Mechanical Mini for Mac Wireless Keyboard, Space Gray
As an affiliate, we earn on qualifying purchases.
From Code Generation to Review
In March 2025, Anthropic chief executive Dario Amodei told a Council on Foreign Relations audience that AI might write 90% of code within three to six months and essentially all of it within a year. The timeline was a prediction, not a measured industry-wide outcome. Anthropic’s later company-specific figures show extensive use of Claude inside Anthropic, but cannot establish that the same share applies across software companies.
Other reported figures offer a narrower comparison. Google chief executive Sundar Pichai said in October 2024 that AI generated more than a quarter of new code at Google, which engineers then reviewed and accepted. METR cited an estimate in February 2026 that Claude Code authored about 4% of GitHub commits. These figures concern different organizations, periods and measures, so they should not be treated as directly comparable estimates of AI’s share of all software development.
Productivity studies also test different kinds of work. A 2022 experiment involving 95 professional developers found that participants using GitHub Copilot completed a small JavaScript HTTP-server task 55.8% faster on average. METR’s later trial involved work in large, familiar codebases, where understanding project conventions and constraints may matter more. The differing results underline why a fast result on a bounded exercise does not necessarily predict performance on complex workplace tasks.
“More than 80% of the code we merge into Anthropic’s codebase was authored by Claude.”
— Anthropic
How Far the Evidence Reaches
The available figures do not settle how much AI improves software productivity across the industry. Anthropic’s adoption and output numbers describe one AI company, while METR’s experiments cover limited groups, tasks and tool versions. METR said its later study was difficult to interpret in part because some developers increasingly declined tasks they did not want to do without AI. Results may also vary with experience, codebase complexity and the quality of review.
Hiring data show a low share of entry-level postings and weaker employment among young workers in AI-exposed roles, but neither measure by itself proves AI caused the change. The cited learning experiment also involved a particular group and a particular library-learning exercise. It is not yet clear how these patterns will develop, which training methods best preserve learning, or whether software employment will expand or contract over the longer term.
What Schools and Employers Track
The next useful evidence will come from larger, repeated studies that measure completed work, quality, defects and security—not just how much code a tool generates. Researchers will also need to distinguish effects by task, developer experience and access to AI. For educators, the learning question is whether students can explain, test and debug code after using assistance, rather than only whether they finish an assignment quickly.
Employers and students can watch whether entry-level postings recover, how job descriptions change, and whether companies provide structured routes for beginners to gain experience. The November 2026 increase in software-development postings is one signal, but it does not show what share are entry-level or whether the increase will continue. For now, the evidence supports teaching programming fundamentals alongside AI use, with particular attention to specification, code review, debugging, testing and secure design.
Key Questions
Does AI mean students no longer need to learn programming?
No. The evidence described here shows AI generating substantial amounts of code in some settings, not that programming knowledge is unnecessary. Understanding requirements, reviewing output, debugging and testing remain relevant skills.
Has AI been proven to make software developers faster?
No single conclusion applies across tasks. METR’s early-2025 randomized trial found experienced developers took 19% longer with the AI tools tested, while a 2022 Copilot experiment found faster completion on a small coding task. METR said its newer evidence was too weak to establish the size of any later speed-up.
What does the entry-level job data show?
Indeed data cited in the report put entry-level roles at 4.5% of U.S. software-development postings in the first quarter of 2026, compared with 69.3% for senior roles. The figures describe posting shares; they do not prove AI caused the imbalance.
Can using AI interfere with learning to code?
One Anthropic experiment found mostly junior developers who used AI while learning a new library scored 50% on a follow-up quiz, compared with 67% among those who coded by hand. That result is limited to the experiment and does not establish that every form of AI-assisted learning reduces understanding.
Source: ThorstenMeyerAI.com
