🔍 Read the full analysis: Learning Programming In The AI Era: Evidence And Considerations 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 now authors a large share of code at some companies, but evidence that it consistently makes developers faster is mixed. Job postings suggest a tougher entry-level market, while a learning experiment found that junior developers using AI scored lower on a follow-up quiz than those who coded by hand.
A November 18 review of AI coding data, developer experiments and hiring indicators finds that learning programming still has practical value, but the work is shifting toward specifying, reviewing, testing and debugging code. The evidence is mixed: AI writes much of the code at some companies, yet studies have not established that it reliably speeds up complex software work, and an experiment found that junior developers learned less from a new library when they relied on AI.
The clearest company-level figure comes from Anthropic, which reported that Claude authored more than 80% of the code merged into its codebase in May 2026, compared with a share in 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. Those are company-reported measures, not an independent estimate of software quality or industry-wide productivity.
Anthropic cautioned that lines of code measure quantity rather than quality, making the eightfold output figure an overstatement of the true productivity gain. It described engineers as directing and reviewing AI-written code. That distinction also echoes a statement by CEO Dario Amodei at the Council on Foreign Relations on March 10, 2025: even if AI writes most code, programmers still need to specify what is being built, how it fits with existing software and whether its design is secure.
Controlled tests give a less uniform picture of speed. In a 2025 randomized trial, the nonprofit METR assigned 16 experienced open-source developers 246 tasks in projects they knew well. Developers using AI tools took 19% longer to complete the tasks, despite reporting that they believed AI had made them 20% faster. A separate 2022 experiment found developers completed a small JavaScript server task 55.8% faster with Copilot. METR said its newer study offered only very weak evidence about how much tools may have improved by early 2026.
The Skills Behind AI-Written Code
The findings matter to people choosing courses, changing careers and deciding how to train software teams. If AI handles more typing, learners still need to determine whether its output matches a specification, works with existing systems and behaves safely. Those tasks require enough programming knowledge to spot errors rather than simply accept a plausible-looking answer.
A learning experiment points to a trade-off. In an Anthropic study, 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 largest gap was in debugging. This does not prove that AI always impairs learning: the result concerns a particular experiment, a particular learning task and a particular group. It does suggest that producing working code with assistance and understanding that code are not the same outcome.
For employers, the evidence argues against treating code volume as a direct measure of productivity. AI-written code may reduce some typing while creating review, testing and maintenance work. For learners, the practical implication is not to avoid AI altogether, but to pair its use with deliberate practice in reading code, debugging and explaining why a solution works.
Popular right now

Samsung 65-Inch 4K OLED Smart TV with Dolby Atmos, 120Hz Refresh Rate, and Amber Protection Bundle

Philips 5500 Series Fully Automatic Espresso Machine with LatteGo Milk System

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
As an affiliate, we earn on qualifying purchases.
Hiring Signals and Earlier Tests
Employment indicators paint a mixed picture rather than showing that demand for programmers has disappeared. According to Indeed data cited in the November review, 4.5% of US 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 jobs exposed to AI was 19% below where it would otherwise have been. That finding covers AI-exposed jobs, not software roles alone, and does not by itself establish that AI caused the gap.
At the same time, Indeed’s US software-development postings were up about 23% year over year in November 2026, after reaching a low in May 2025. The figures indicate a difficult entry point alongside a recent increase in postings; they do not reveal how many jobs were filled, what skills employers sought or how much of the hiring pattern was caused by AI.
Results from coding experiments also depend on what developers are asked to do. The 2022 Copilot test involved writing a relatively small, fresh HTTP server task. METR’s trial involved experienced contributors working within large open-source projects they already knew. Google CEO Sundar Pichai said in October 2024 that more than a quarter of new code at Google was AI-generated and then reviewed and accepted by engineers. These company statements and experiments offer useful examples, but they are not interchangeable measures of the whole industry.
“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 CEO, speaking at the Council on Foreign Relations on March 10, 2025
What the Evidence Cannot Yet Settle
There is no single reliable measure here of AI’s effect on software productivity. Anthropic’s code share and output-per-engineer figures come from a company that builds AI tools and uses them extensively; its own report warns that code volume does not equal quality or productivity. METR’s early-2025 trial covered a small group of experienced developers and particular tasks, while its later data were too uncertain to establish the size of any speed-up.
The hiring figures also leave key questions open. The cited data do not show whether AI caused the narrower share of entry-level postings or the employment gap among young workers in AI-exposed roles. Nor does the learning experiment establish how outcomes change with different tools, teaching methods or longer-term practice. The evidence supports caution about assuming AI makes everyone faster or learning unnecessary; it does not settle what programming employment or education will look like over time.
Signals to Watch in 2027
The next useful evidence will come from larger, independently conducted studies that test current AI tools on varied software tasks and measure more than lines of code or completion time. Researchers will need to track code quality, defects, security, review effort and whether developers retain what they learn after using AI. Updated hiring data can show whether entry-level postings recover, but those figures should be read alongside broader economic and industry changes.
For now, learners can use AI while preserving opportunities to solve problems unaided, explain generated code and find faults through testing and debugging. Employers can compare AI-assisted work with the full costs of review and maintenance rather than counting generated output alone. Those steps respond to what current studies measure without treating uncertain findings as a final verdict on programming careers.
Key Questions
Is learning programming still worthwhile?
The evidence does not show that programming has become pointless. It suggests that learners should develop skills in reading, testing, debugging and specifying software alongside writing code, including with AI assistance.
Does AI make software developers faster?
Results vary by task and study. METR’s 2025 trial found experienced developers took 19% longer with AI on its tasks, while a 2022 Copilot experiment found faster completion of a small coding exercise. METR said its later evidence was too weak to measure the newer tools’ speed-up reliably.
Does the Anthropic figure mean AI writes 80% of all software?
No. Anthropic reported that Claude authored more than 80% of code merged into Anthropic’s own codebase in May 2026. It is not an estimate for the software industry as a whole.
Did AI reduce entry-level programming jobs?
The cited data show that 4.5% of US software-development postings were entry-level in the first quarter of 2026, and Stanford researchers reported an employment gap among young people in AI-exposed jobs. These figures do not establish that AI caused either pattern.
Source: ThorstenMeyerAI.com
