🔍 Read the full analysis: What Should Students Learn About Coding As AI Evolves? on ThorstenMeyerAI.com
Get smart everyday buys 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
Evidence reported through November 2026 suggests AI can author a large share of code, but its effect on developer productivity is less certain and varies by task. Students should still learn programming, with particular attention to specifying requirements, reading and testing code, debugging, and evaluating security. Research also warns that relying on AI while learning can weaken understanding.
Students should still learn to code, but the evidence available in 2026 points to a change in emphasis: specifying software requirements, reading and testing code, debugging, and checking security are increasingly important as AI tools generate more of the code itself. That conclusion comes from a review of company disclosures, controlled experiments, developer surveys and job-posting data; it does not establish that AI makes all programmers faster or that coding jobs are disappearing.
Anthropic reported that in May 2026, Claude authored more than 80% of the code the company merged into its own codebase, compared with “the low single digits” before Claude Code launched in February 2025. The company said engineers directed and reviewed the code rather than typing it all themselves. It also reported that a typical engineer merged eight times as much code per day in the second quarter of 2026 as in 2024, while warning that lines of code measure quantity rather than quality and likely overstate productivity gains.
Controlled research gives a more mixed picture of speed. In a 2025 randomized trial, 16 experienced developers completed real tasks in familiar open-source projects; tasks assigned to allow AI took 19% longer on average. The developers had expected AI to save time and later believed it had made them faster. A separate 2022 experiment found developers completed a small, newly assigned coding task 55.8% faster with GitHub Copilot. The different results reflect different tasks, tools and settings, and neither gives a universal estimate of AI’s effect.
Learning outcomes also matter. In an Anthropic experiment, mostly junior developers using AI to learn an unfamiliar programming library scored 50% on a follow-up quiz, compared with 67% among developers who coded by hand; the largest gap involved debugging. At the same time, entry-level hiring appears difficult: Indeed data showed that only 4.5% of U.S. software-development postings were entry-level in the first quarter of 2026. But postings were up about 23% year over year in November, according to the source report, so the available figures do not show that demand has vanished.
The Skills Behind AI-Assisted Coding
For students, the distinction is between producing code and understanding whether it solves the right problem. AI can generate a plausible implementation, but someone still has to define requirements, fit changes into an existing system, test unusual cases, locate defects and judge whether a design is secure. Those responsibilities appear in the evidence as important even when a tool handles much of the typing.
The findings also caution against treating code volume as a proxy for value. Anthropic’s eightfold measure counts lines merged per engineer, not whether those lines improved software or reduced project time by the same amount. METR’s trial, by contrast, measured completion time for selected tasks and found slower results in its 2025 setting. Students and employers should distinguish AI-authored code from verified productivity, and productivity from demonstrated understanding.
That distinction affects training and hiring. If employers expect junior staff to review AI-generated work, new programmers need practice identifying errors and explaining decisions, not just prompting a tool. The reported drop in entry-level postings makes access to supervised practice a concern, although the data presented do not establish that AI caused the change or predict how hiring will develop.
Products Worth Considering

Apple iMac (2024) 24-inch All-in-One Desktop with M4 Chip, 16GB Unified Memory, 256GB SSD, Pink

Dell UltraSharp U2424H 24-Inch FHD Monitor

Apple Mac mini Desktop Computer with M5 Pro Chip (2026)
As an affiliate, we earn on qualifying purchases.
From Code Generation to Review
In March 2025, Anthropic chief executive Dario Amodei told the Council on Foreign Relations that AI could write 90% of code within three to six months and essentially all code within a year. He also said programmers would still need to specify what was being built, how it connected to other code and whether the design was secure. Anthropic’s later disclosure that Claude authored more than 80% of code merged at the company offers one company-specific measure of adoption, not a representative figure for the software industry.
Other reported figures are lower or cover different periods. Google chief executive Sundar Pichai said in October 2024 that AI generated more than a quarter of new Google code, which engineers then reviewed and accepted. METR cited an estimate in February 2026 that Claude Code authored about 4% of GitHub commits. These measures use different companies, tools and definitions, so they should not be treated as direct comparisons.
Evidence on speed is similarly dependent on context. The 2022 Copilot test involved writing a small HTTP server, while METR’s 2025 trial involved experienced developers working in codebases they already knew. METR said its second study, begun in August 2025, produced an unreliable signal: some developers were reluctant to take tasks without AI. Its estimates suggested possible time savings, but confidence intervals included no effect. The organization described its evidence for a larger early-2026 speed-up as “only very weak evidence”.
““More than 80% of the code we merge into Anthropic’s codebase was authored by Claude.””
— Anthropic, in its 2026 report on coding at the company
Where the Evidence Falls Short
The figures do not settle how much AI improves software productivity across workplaces. Anthropic’s code-share and output measures come from a company that builds AI coding tools and uses them heavily; they cannot establish typical results elsewhere. The studies also involve different tools, tasks and participants, while METR’s follow-up trial had methodological problems that left its estimates uncertain.
It is also unclear how much AI contributed to the drop in entry-level job postings or the reported employment decline among 22- to 25-year-olds in AI-exposed work. The source report cites a Stanford analysis finding employment 19% below where it would have been, but that comparison does not by itself establish AI as the cause. Posting counts, employment estimates and actual hiring are different measures.
The learning experiment raises a concern about reliance on AI, but one experiment does not show that every student learns less with these tools. The report does not provide enough detail to establish how results vary by course design, level of AI assistance or the way students are assessed. The best teaching practices and the long-term effects on skill development remain unsettled.
What Schools and Employers Can Track
The next useful evidence will show whether newer controlled studies can measure the effect of current AI tools without the participation problems METR reported. Researchers and employers will need to compare similar tasks, track both time and quality, and distinguish code generated from code that is reviewed, accepted and maintained.
For education, the source material points toward evaluating students on more than whether a program runs. Courses can test whether students can explain requirements, trace unfamiliar code, write and interpret tests, debug failures and identify security risks, including when AI assistance is permitted. The reported learning experiment makes it worth checking whether students retain those skills after using a coding assistant.
Employers and educators can also watch entry-level postings and hiring alongside broader software-development demand. Indeed’s November 2026 posting increase is a single indicator, not proof of a sustained recovery or a forecast. Further hiring data, replication of learning studies and more reliable productivity measurements will clarify how coding education and early-career work are changing.
Key Questions
Should students still learn to code?
Yes. The evidence suggests that students need programming knowledge to specify what software should do and to evaluate, test, debug and secure AI-generated code. It does not show that code-writing skills are no longer useful.
Does AI make developers faster?
Results vary by task and study. METR’s 2025 trial found experienced developers took 19% longer with AI on selected real-world tasks, while a 2022 Copilot experiment found faster completion on a small coding assignment. METR said its later evidence for improved speed was weak.
What coding skills should students prioritize?
The findings point to requirements writing, code reading, testing, debugging and security review, alongside the ability to write code. Students should also practice explaining why a solution works and checking AI output rather than accepting it without review.
Does AI explain the decline in entry-level software postings?
The reported figures do not establish that. Indeed data put entry-level roles at 4.5% of U.S. software-development postings in the first quarter of 2026, but that figure alone does not identify the cause. The source report also says postings were up about 23% year over year in November 2026.
Source: ThorstenMeyerAI.com
