🔍 Read the full analysis: What AI Means For The Future Of Learning To Code on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
AI now writes a large share of code at some companies, but evidence that it consistently makes developers faster is mixed. Hiring data point to a narrower entry-level market, while a learning experiment found that developers using AI scored lower on a follow-up quiz than those coding by hand.
AI is writing much of the code at some technology companies, but evidence on whether it reliably speeds developers up or helps beginners learn is mixed. A report published November 18 by ThorstenMeyerAI.com draws on 2026 hiring data, controlled studies and developer surveys to argue that learning to code still matters, though learners may need to focus more on directing, testing and understanding code than on typing it.
Anthropic said that more than 80% of code merged into its codebase in May 2026 was authored by Claude, compared with what the company described as the low single digits before Claude Code launched in February 2025. 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 track quantity rather than quality and that the eightfold figure is “almost certainly an overstatement” of the productivity gain.
Controlled evidence offers a more complicated picture. In a 2025 randomized trial, METR found that 16 experienced developers took 19% longer to complete 246 real tasks in familiar open-source projects when using the AI tools tested. Participants had expected a speed-up and afterward believed they had worked faster. In a later study, METR said its estimates were too uncertain to establish the size of any improvement, although it considered developers likely to be more sped up in early 2026 than a year earlier.
A separate Anthropic experiment points to a learning trade-off: 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. The report says the largest gap concerned debugging. These findings sit alongside employment signals: Indeed data show entry-level roles made up 4.5% of US software development postings in the first quarter of 2026, while senior roles accounted for 69.3%. The figures describe postings, not the number of people hired.
Skills Beyond Writing Code
The findings matter to people entering the field because AI adoption and a tight entry-level market are happening at the same time. If employers expect new hires to work with AI-generated code, beginners still need enough understanding to detect errors, explain what software should do and recognize when a design may be insecure. Faster code production alone does not establish that a project is correct or safe.
For schools and training programs, the learning experiment raises a practical question: whether AI assistance helps students complete an exercise while leaving them less able to explain or debug the result later. The quiz result is evidence from one experiment, not proof that all AI-supported learning weakens skills. It does suggest that training should assess understanding, not only whether a working program was produced.
For employers, the data also argue against treating code volume as a direct measure of productivity. Anthropic itself says its output measure overstates gains, and METR’s task-based studies show why results can depend on the work, the developers and the tools. Hiring and training decisions based on broad claims about AI replacing programming work would go beyond the evidence presented.
Popular right now

Bose QuietComfort Ultra Earbuds (2nd Gen), Wireless Noise-Cancelling Earbuds, Black

Bose QuietComfort Wireless Noise Cancelling Headphones – Twilight Blue Limited Edition

KitchenAid 5.5 Qt Bowl-Lift Stand Mixer – Porcelain White (KSM55SXXX)
As an affiliate, we earn on qualifying purchases.
From Code Generation to Review
In March 2025, Anthropic chief executive Dario Amodei predicted that AI would be writing 90% of code within three to six months and essentially all code within 12 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. Anthropic’s later internal figures indicate substantial AI authorship at the company, but do not establish that the same share applies across the industry.
Results vary with the task. 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 instead examined real tasks in large codebases developers already knew. Those different settings help explain why a result from a short coding exercise cannot by itself predict performance on everyday software work.
Other figures in the report offer a broader but incomplete picture. 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. A February 2026 estimate cited by METR put Claude Code’s share at about 4% of GitHub commits. These measures cover different organizations, tools and periods, so they should not be read as a single industry-wide rate.
“The eightfold increase is “almost certainly an overstatement of the true productivity gain.””
— Anthropic, on its reported code output per engineer
Limits in the Available Evidence
How widely Anthropic’s figures apply is unknown. It is an AI developer reporting on its own workforce, which may use its products more intensively than other employers. Its measure of merged lines does not establish whether code is better, safer or cheaper to maintain.
Productivity estimates also remain unsettled. METR’s 2025 trial covered 16 experienced developers and tasks in projects they knew; its later study had an unreliable signal as some participants resisted working without AI. METR said the newer evidence could not firmly determine the size of any speed-up. The 2022 Copilot experiment used a different, smaller task, so results cannot be directly compared as though they measured the same work.
The employment figures do not prove that AI caused fewer opportunities for young workers. The report cites Stanford researchers’ finding that employment of 22- to 25-year-olds in AI-exposed jobs was 19% below where it would have been, but the supplied material does not establish that AI alone caused the difference. Indeed postings are a measure of advertised jobs, not completed hires or the full labor market. The learning result is likewise one experiment, and its long-term effects are not established.
Evidence to Watch in 2027
Further controlled studies and updated hiring data will help show whether AI tools produce durable productivity gains and how entry-level demand develops. METR’s follow-up work may offer clearer estimates as developers and tools change, while job-posting data can show whether the share of junior roles recovers or continues to lag.
Schools, employers and learners can also test training approaches that pair AI assistance with independent practice, code review and debugging. The evidence available in November 2026 does not settle which balance works best. For now, the report’s practical conclusion is that learning to code remains relevant, but fluency should include judging and maintaining code as well as generating it.
Key Questions
Does AI make learning to code pointless?
No. The evidence described here suggests AI can generate substantial amounts of code, but developers still need to specify requirements, review output, debug problems and evaluate security. The right mix of skills may be changing, but the report does not show that programming knowledge is unnecessary.
Are developers consistently faster with AI?
The available studies do not establish that. METR’s 2025 trial found experienced developers took 19% longer on the tested tasks with AI, while its later evidence was too weak to measure the size of a possible speed-up. A separate 2022 Copilot experiment found faster completion on a small coding task.
What did the learning experiment find?
Mostly junior developers who used AI to learn a new library scored 50% on a follow-up quiz, compared with 67% for participants who coded by hand. The report says the largest difference was in debugging. The result comes from one experiment and does not establish effects for every learner or teaching method.
Are entry-level programming jobs disappearing?
The cited Indeed data show that entry-level roles were 4.5% of US software development postings in the first quarter of 2026, while senior roles were 69.3%. Those shares indicate a difficult market for beginners, but they do not prove AI caused the pattern or show how many jobs were ultimately filled.
Source: ThorstenMeyerAI.com
