🔍 Read the full analysis: How AI Affects The Choice To Learn Coding on ThorstenMeyerAI.com
TL;DR
AI is producing a growing share of code at some companies, but research does not yet establish that it reliably makes developers faster or removes the need for coding skills. Hiring data point to a narrower entry-level market, while an experiment found that learners using AI scored lower on a follow-up quiz than those who coded by hand.
AI now writes a large share of some companies’ code, but the evidence available in 2026 does not establish that learning to program is no longer worthwhile. Company figures, controlled studies and job-posting data instead point to a changing skill mix: developers may spend less time typing code and more time specifying, reviewing, testing and debugging it, while entry-level hiring remains a concern.
Anthropic reported that in May 2026, more than 80% of the code it merged into its codebase had been authored by Claude, compared with what it described as the low single digits before Claude Code launched in February 2025. The company also said its typical engineer merged eight times as much code per day in the second quarter of 2026 as in 2024. It cautioned that lines of code measure quantity rather than quality and that this increase was “almost certainly an overstatement” of the productivity gain. Anthropic is an AI developer with unusually strong reasons to use its own coding tools, so its figures should not be treated as representative of all employers.
Whether AI makes developers faster is less settled. In a randomized METR study conducted in early 2025, 16 experienced open-source developers completed 246 tasks in codebases they already knew. They took 19% longer when allowed to use the AI tools tested, despite expecting a 24% time reduction and later believing they had become 20% faster. METR said a later study offered only very weak evidence about the size of any speed-up by early 2026; its estimates included no effect.
Other indicators show pressure on people seeking their first software job, but not a collapse in overall postings. The source reports that 4.5% of US software-development job postings on Indeed were entry-level in the first quarter of 2026, while 69.3% were senior. It also reports that US software-development postings were about 23% higher year over year in November 2026 after reaching a low in May 2025. These figures describe postings, not confirmed hires, and do not by themselves establish that AI caused changes in demand.
Coding Skills Beyond Code Generation
For students and career changers, the practical issue is not simply whether an AI tool can generate working code. Someone still has to define the task, check whether the result fits a larger system, find faults and judge security risks. Those responsibilities matter when generated code is incomplete, unsuitable or unsafe, and they require enough technical understanding to evaluate what a tool produces.
The source also describes an Anthropic learning experiment in which mostly junior developers used AI to learn a new library. On a follow-up quiz, that group scored 50%, compared with 67% for participants who coded by hand; the largest gap involved debugging. This result suggests that outsourcing practice may reduce learning in at least some settings. It does not show that every use of AI harms learning, or that the same result applies to every learner, tool or task.
Together, the findings make the choice more specific than “learn coding or do not.” Coding instruction may need to include reading, testing and debugging generated code, as well as writing it. For employers, the narrow share of entry-level postings raises a separate question: how people can gain the practical experience needed to take on work that listings classify as senior.
Popular right now

Sony WH-1000XM6 Wireless Noise-Canceling Headphones, Black

Dyson V8 Cyclone Cordless Vacuum

Ninja Foodi 8-Quart DualZone 6-in-1 Air Fryer, Grey (DZ201)
As an affiliate, we earn on qualifying purchases.
From Copilot Tests to Workplace Data
The evidence comes from different kinds of work, so the results are not directly interchangeable. A 2022 experiment with 95 professional developers found that those using GitHub Copilot completed a small JavaScript HTTP-server task 55.8% faster on average: 71.17 minutes versus 160.89 minutes. Less experienced programmers benefited most in that test. METR’s later trial involved real tasks in large, familiar open-source projects, not a single new programming exercise. Task size, familiarity and the tools available can all affect the result.
Company reports also differ in scope. 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. Anthropic’s much higher 2026 figure describes its own codebase and its own tools. In February 2026, METR cited an estimate that about 4% of GitHub commits were authored by Claude Code. These measures cover different periods and populations; they do not form a single comparable measure of AI’s share of all software development.
Hiring figures need similar care. The source attributes to Stanford researchers a finding that employment among 22- to 25-year-olds in jobs exposed to AI was 19% below a projected level. That is a reported estimate, not proof that AI alone caused the difference. Indeed’s posting shares and year-over-year totals are also indicators of advertised demand, rather than a count of jobs filled or evidence about why employers changed their listings.
„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, speaking at the Council on Foreign Relations on March 10, 2025
How Broad Are the Effects?
It remains unclear how much AI changes productivity across the software industry. Anthropic’s code-output figures come from one company, while METR’s early trial found slower completion on the tasks studied and its later estimates were too uncertain to establish a reliable effect. The source does not provide a single comparable, industry-wide measure of time saved or code quality.
The hiring picture is also incomplete. A low entry-level share of postings may make the first step into software work harder, but postings are not hires, and the figures given do not isolate AI’s role from other market forces. The source’s reported Stanford estimate concerns employment in AI-exposed jobs among a particular age group; it does not establish a general outcome for all new programmers.
Finally, the learning experiment was limited to a particular group and task. Its quiz results raise a concern about relying on AI while learning, especially for debugging, but they do not show which teaching approaches best combine AI assistance with durable skill development.
Evidence to Watch in 2027
The next useful evidence will come from studies that test newer tools on a wider range of real development tasks and report both completion time and the quality, reliability and security of the resulting code. METR has already warned that its later study was difficult to interpret because some developers were reluctant to do tasks without AI; future research will need to account for that reliance as well as tool improvements.
Job-posting and hiring data will also need to be followed over time. A continued rise in total postings alongside a small entry-level share would indicate a different challenge from a broad decline in software demand. Researchers and employers will need to distinguish advertised roles from actual hiring and examine what experience employers expect from junior candidates.
For learners, the evidence supports a measured approach rather than abandoning programming study: use AI as an aid, but keep practicing the work needed to judge its output. Whether schools and employers adopt that balance—and whether it improves learning and job access—remains to be tested.
Key Questions
Does AI mean it is no longer worth learning to code?
No such conclusion is established by the evidence here. AI can generate substantial amounts of code in some workplaces, but developers still need to specify tasks and review, test and debug the output. The job market and the best way to teach programming are changing, but the source does not show that coding skills are obsolete.
Does AI make software developers faster?
The results are mixed. METR’s early-2025 randomized trial found experienced developers took 19% longer with the tools tested, while a 2022 Copilot experiment found participants completed a small coding task 55.8% faster. METR later said its newer evidence was too weak to establish the size of any speed-up.
Are entry-level software jobs disappearing?
The source reports that entry-level roles made up 4.5% of US software-development postings on Indeed in the first quarter of 2026. That points to a limited share of advertised roles, but it does not show that entry-level jobs have disappeared or establish why the share was low.
Can using AI make it harder to learn programming?
In one Anthropic experiment, mostly junior developers who used AI to learn a new library scored 50% on a follow-up quiz, compared with 67% among those who coded by hand. The result raises a concern, particularly around debugging, but it does not prove that every form of AI-assisted learning reduces skill.
Source: ThorstenMeyerAI.com