TL;DR. “Learn to code” and “coding is dead” are both wrong because they treat four decisions as one. The U.S. Bureau of Labor Statistics projects software developers to grow 10% from 2025 to 2035 (174,700 more jobs) while computer programmers shrink 7%. Recent computer science graduates had a 7.0% unemployment rate in the New York Fed’s latest by-major table, against 4.2% for all majors, yet one of the lowest underemployment rates and an $87,000 median early-career wage. Of six AI coding studies lined up below, two show clear speed or output gains, one shows experienced developers taking 19% longer, one shows less secure code, and one shows a 17% lower skills-quiz score for people who learned a new library with AI. The CEO move: decide which of four levels you need and what you will own there. The student move: learn to read, test and verify code before you learn to type it.
This article belongs to the CEOtudent AI literacy cluster. The hub piece, Prompt Engineering Is Not Enough, maps the full skill stack; The New Technical Literacy covers what non-technical staff must understand about AI systems; and the 4-Factor Decision Matrix is the general tool for choosing a next skill. This piece is general information, not personalised career or financial advice.
Why is “should you learn to code” the wrong question?
The question hides three variables.
Which job. Official statistics have no “coder” category. They have software developers, who design applications and systems; computer programmers, who in the Bureau of Labor Statistics’ words “turn the designs created by software developers and engineers into instructions that a computer can follow”; and web developers. The three are projected to move in different directions.
Which person. Entry-level and experienced workers are having different years.
Which purpose. Writing software that customers depend on, building a small tool for your own team, and reading what an AI agent produced are three skills with three costs. Stack Overflow’s 2025 survey uses the definition “the process of generating software from LLM prompts” for vibe coding. That practice has opened the second and third purposes to almost anyone. It has not made the first easier to enter.
A useful answer is a sorting test, not a yes or a no.
What do the official job numbers say about people who write code?
The Occupational Outlook Handbook now carries May 2025 wages and the 2025-2035 projections released on August 27, 2026.
Table 1. Code-writing occupations in the BLS Occupational Outlook Handbook (figures as printed)
| Occupation (SOC code) | Employment, 2025 | Projected employment, 2035 | Change, 2025-2035 | Median annual wage, May 2025 |
|---|---|---|---|---|
| Software developers (15-1252) | 1,717,800 | 1,892,600 | +10% (+174,700) | $135,980 |
| Software quality assurance analysts and testers (15-1253) | 187,600 | 198,200 | +6% (+10,600) | $104,300 |
| Computer programmers (15-1251) | 110,800 | 102,700 | -7% (-8,100) | $100,390 |
| Web developers (15-1254) | 87,400 | 90,600 | +4% (+3,300) | $92,650 |
| Web and digital interface designers (15-1255) | 132,700 | 140,800 | +6% (+8,000) | $104,000 |
Source: U.S. Bureau of Labor Statistics, Occupational Outlook Handbook. BLS rounds employment to the nearest 100, so a printed change can differ by 100 from simple subtraction. For comparison, total U.S. employment is projected to grow 3.5%, and the May 2025 median annual wage for all workers was $50,980.
The Bureau’s narrative matters as much as the numbers. For programmers it says that “many companies are expected to leverage technologies, including artificial intelligence (AI), to automate repetitive programming tasks” and that some higher-skilled programming tasks will likely shift to software developers. For developers it expects strong demand from software for AI, the Internet of Things, robotics and other automation. For web developers it warns that AI tools “may soften the employment growth” and may “allow some workers in other occupations to do basic web development tasks”. That last sentence is the official version of the vibe coding story.
Table 2. What the three occupations look like side by side (CEOtudent analysis of BLS Occupational Outlook Handbook data)
| Measure | Software developers | Computer programmers | Web developers |
|---|---|---|---|
| Median wage as a multiple of the all-worker median ($50,980) | 2.67x | 1.97x | 1.82x |
| Wage gap to software developers | – | $35,590 lower | $43,330 lower |
| Projected growth compared with all occupations (3.5%) | 2.9 times the average | declining | 4%, slightly above the average |
| Developers per programmer, 2025 | 15.5 | – | – |
| Developers per programmer, 2035 (projected) | 18.4 | – | – |
| Net projected change across the three | +169,900 jobs (+174,700, -8,100 and +3,300) |
Source: CEOtudent calculations from Table 1.
The market already pays a 35.5% premium for the job that decides what to build over the job that translates a design into code, and the projections widen the gap from 15.5 developers per programmer to 18.4. A shrinking occupation is not a closed one: about 4,400 programmer openings a year are still projected, all from replacing workers who leave. The Bureau’s AI exposure categories and their caveats are covered in the CEOtudent piece on layoff-resilient careers.
What is happening to new computer science graduates?
The Federal Reserve Bank of New York publishes outcomes for 73 college majors from the American Community Survey. The current table, released on February 4, 2026, uses 2024 data.
Table 3. Recent graduates by major (New York Fed, figures rounded to one decimal from the published data file)
| Major | Unemployment rate | Underemployment rate | Median wage, early career | Median wage, mid-career |
|---|---|---|---|---|
| Computer science | 7.0% | 19.1% | $87,000 | $120,000 |
| Computer engineering | 7.8% | 15.8% | $90,000 | $131,000 |
| All majors | 4.2% | 39.4% | $58,000 | $87,000 |
Source: Federal Reserve Bank of New York, The Labor Market for Recent College Graduates. Rates are for graduates aged 22 to 27; wages are for full-time workers with a bachelor’s degree only (mid-career: ages 35 to 45). Underemployment is the share working in jobs that typically do not require a degree.
Table 4. Where the two computing majors rank among 73 majors (CEOtudent analysis of New York Fed data)
| Measure | Computer science | Computer engineering |
|---|---|---|
| Rank by unemployment rate (1 = highest) | 4th highest | 2nd highest |
| Rank by underemployment rate (1 = lowest) | 9th lowest | 4th lowest |
| Rank by median early-career wage (1 = highest) | 2nd highest | 1st |
| Unemployment gap to all majors | +2.8 points | +3.6 points |
| Underemployment gap to all majors | -20.2 points | -23.5 points |
Source: CEOtudent calculations from the New York Fed data file; gaps use the unrounded values.
The pattern is a queue, not a collapse. Graduates of these majors are among the most likely to be unemployed. Those who are employed are far less likely to be in a job that does not need a degree, and they earn the highest starting wages in the table. A plausible reading is that they hold out for degree-level technical roles while the entry door to those roles has narrowed. The table cannot prove that mechanism on its own.
Payroll data points the same way. The August 2026 update of “Canaries in the Coal Mine?” by Brynjolfsson, Chandar and Chen at Stanford uses ADP payroll records through June 2026. It finds no evidence of economy-wide job displacement. It also finds that employment of workers aged 22 to 25 in AI-exposed occupations stands 19% below where it would be had it kept pace with their less-exposed peers, up from 15% at the July 2025 data vintage, while experienced workers show no comparable gap. By June 2026, employment of 22 to 25 year-olds in the panel was about 2% below its November 2022 level, while workers aged 35 to 40 were up 11%. The divergence works mainly through reduced hiring, and it persists when computer occupations and technology firms are excluded.
The authors call these facts “early, descriptive indicators”, not causal estimates, and note that the patterns attenuate when controlling for education. For this decision the meaning is narrow but important: the risk in a software career has moved to the entry point.
Do AI coding tools make people faster or slower?
Both, depending on who and on what. Table 5 lines up six studies, including those where the effect goes the wrong way.
Table 5. Six studies of AI coding assistance, lined up (CEOtudent analysis of the cited papers; effect sizes as printed by the authors)
| Study | Design and sample | Setting | Reported effect | Direction |
|---|---|---|---|---|
| Peng et al., 2023 (Microsoft, GitHub, MIT) | Controlled experiment; 95 professional programmers from a freelancing platform | One task: an HTTP server in JavaScript | 55.8% faster (95% confidence interval 21% to 89%); less experienced developers gained most | Faster, on a self-contained task |
| Cui et al., 2025 working paper | Three field experiments; 4,867 developers at Microsoft, Accenture and a Fortune 100 firm | Ordinary work with a code-completion assistant | 26.08% more completed tasks (standard error 10.3%); short-tenure staff gained 27% to 39%, long-tenure 8% to 13% (split not statistically significant) | More output, larger for newer staff |
| METR, 2025 | Randomized trial; 16 experienced open-source developers, 246 tasks | Their own mature repositories | 19% longer with AI allowed; they had forecast 24% less time and afterwards believed they saved 20% | Slower; perception gap of 39 to 43 points |
| METR, February 2026 update | 57 developers, more than 800 tasks | Open-source projects, later tools | Returning developers: 18% less time (interval from 38% less to 9% more); new developers: 4% less (15% less to 9% more) | Leaning faster, interval spans zero; METR calls the signal unreliable |
| Perry et al., 2023 (Stanford) | User study; 47 participants, 33 with an AI assistant, 14 without | Five security-related tasks | SQL task: 36% of the AI group wrote code open to SQL injection against 7% of controls; AI users were more likely to believe their code was secure | Less secure, more confident |
| Shen and Tamkin, 2026 (Anthropic, preprint) | Randomized experiment; 52 participants, 26 per group | Learning an unfamiliar Python library | 17% lower score on a follow-up skills quiz (Cohen’s d 0.738); no statistically significant time saving | Less learning, no reliable speed gain |
Source: the papers named in the Sources list. The perception gap is our arithmetic: forecast (24) or belief (20) plus the measured slowdown (19).
Gains are largest where the task is self-contained and the person is new. The 55.8% figure comes from one bounded exercise. In ordinary company work the gain was 26.08% more completed tasks, leaning toward short-tenure staff.
Familiarity can reverse the sign. The 2025 METR developers knew their repositories deeply, and early-2025 tools cost them time. The follow-up suggests later tools changed the picture, but METR reports that 30% to 50% of developers withheld tasks they did not want to do without AI, so it treats the newer estimate as weak evidence and is redesigning the study.
People misjudge their own speed. Developers in the 2025 trial expected to save 24%, lost 19%, and still believed they had saved 20%. A good afternoon with an AI tool is not a measurement.
Speed is not the only outcome. In the Stanford study, 67% of AI users solved the encryption task correctly against 79% of controls, and AI users were more confident about security while being less secure. In the Anthropic study the cost showed up in learning.
Developers appear to have absorbed this. Stack Overflow’s 2025 Developer Survey, a self-selected company survey, reported that 84% of respondents were using or planning to use AI tools, while more distrusted the accuracy of the output (46%) than trusted it (33%). The top frustration, cited by 66%, was “AI solutions that are almost right, but not quite”. In the 2026 edition, 73% use coding assistants and agents daily, and 48% say they trust AI when they can easily validate the answers. Heavy use coexists with conditional trust, and the condition is the ability to check.
Does learning to program make you a better thinker?
This is the oldest argument for learning to code regardless of career. One solid source exists: a meta-analysis by Scherer, Siddiq and Sanchez Viveros in the Journal of Educational Psychology covering 105 studies and 539 effect sizes. It reports a moderate overall transfer effect (g = 0.49), a strong effect for near transfer (g = 0.75) and a moderate effect for far transfer (g = 0.47), with school achievement and literacy benefiting least. Crossref lists no correction as of November 2026.
Three limits apply. Effects were significantly larger in studies with untreated control groups than with active controls, and larger in published studies than in grey literature, so the headline probably flatters the true effect. The abstract frames the question around students in schools, not adult professionals. And the paper predates AI assistants. Learning to program probably improves structured problem solving somewhat. That supports it as mental training, not as a career decision.
Which level do you need? The CEOtudent Code Level Test
The evidence sorts readers into four levels. The table is a CEOtudent editorial framework, not a validated instrument, and the level names are ours.
Table 6. The CEOtudent Code Level Test (CEOtudent editorial framework)
| Level | You are here if | What you learn | What you own | Evidence it rests on |
|---|---|---|---|---|
| 0. Operator | Code-shaped problems rarely appear in your work | AI literacy, not programming | The quality of your own outputs | BLS: basic building is moving into other occupations, but it is optional |
| 1. Reader-director | You specify, commission or approve software, automations or analyses that others (people or agents) build | To read code and data flows well enough to ask what it does, what it touches and how it was tested | The specification and the acceptance decision | Stack Overflow 2026: trust depends on being able to validate; Perry et al.: confidence rose as security fell |
| 2. Tool builder | A recurring problem in your own work would be solved by a small script or internal app that you will run and maintain | One language at working level, version control, testing, basic security | The tool, its data and its failures | Peng et al. and Cui et al.: assisted beginners gain most on bounded tasks; Shen and Tamkin: only if you stay engaged |
| 3. Professional engineer | You want to be paid for, and accountable for, software other people depend on | Computer science fundamentals, system design, production practice | Systems, reliability and other people’s data | BLS: developers +10% with a 35.5% wage premium over programmers; New York Fed and Stanford: a narrow entry door |
Run the test in order and stop at the first level that fits.
- Recurrence. In a normal month, does your work produce a problem that software would solve: repetitive data handling, a report rebuilt by hand, two systems that do not talk to each other? If no, Level 0 for now.
- Accountability. When such a problem appears, is your role to define it and accept the result while someone else builds it? If yes, Level 1 is a requirement, not an option.
- Blast radius. Would the thing you build touch only your own or your team’s low-risk data, with mistakes that are visible and reversible? If yes, and you will maintain it, Level 2.
- Income. Do you want building software to be the main thing you are paid for, with responsibility for systems other people rely on? If yes, Level 3, with the conditions below.
When should you NOT take each level?
- Not Level 0 if you approve budgets, vendors or AI agents that produce code. Signing off on work you cannot read is the pattern the security study warns about.
- Not Level 1 alone if you are already pasting AI-generated scripts into live spreadsheets, databases or websites. That is Level 2 work done without Level 2 skills.
- Not Level 2 for anything that handles customer personal data, payments, credentials or regulated records, or that other teams will depend on without an owner. Hand that to Level 3 people. The guide to vibe coding as a solo founder covers the build process; this is the boundary where it stops being safe.
- Not Level 3 if the plan depends on a short course leading straight to a first job at the Table 1 median. The medians describe everyone in the occupation, Tables 3 and 4 describe the door, and the Handbook lists a bachelor’s degree as the typical entry education for all three occupation groups. Treat Level 3 as a multi-year decision.
No time budgets are attached because none of these sources measured them. CEOtudent’s review of time-to-competence data covers hour estimates and their limits.
How should you learn to code when AI can write it?
The Shen and Tamkin experiment is small, comes from an AI company and tested one chat-based tool on one library. It is still the most relevant design available, because it measured learning and not output.
Participants who used AI fell into six interaction patterns. Three averaged 65% or more on the quiz: generating code and then asking follow-up questions to understand it, asking for code together with explanations, and asking only conceptual questions while resolving errors independently. Three averaged below 40%: delegating all code writing, sliding from questions into full delegation, and repeatedly asking the AI to debug. One group had only two people, so the percentages carry little weight. The direction matches the evidence in CEOtudent’s project-based learning ladder: effort that feels slower tends to stick.
Table 7. Learning rules and pre-run checks (CEOtudent editorial framework)
| Rule | In practice | Why |
|---|---|---|
| Read before you write | Explain each generated block in plain words before running it | The high-scoring patterns all involved comprehension work |
| Debug it yourself first | Read the error, form a guess, test it, then ask | Repeated AI debugging was the lowest-scoring pattern |
| Test what you cannot read | Run it on a harmless sample; keep a check that would reveal a wrong result | Developer trust is conditional on validation |
| Assume security is missing | Know what the code can read, change, delete or send, and keep keys and personal data out of it | AI users in the Stanford study were less secure and more confident |
| Measure, do not feel | Time a task with and without assistance | Experienced developers misjudged their speed by about 40 points |
For Level 1, rows one, three and four are the curriculum, together with CEOtudent’s guide to agent literacy. For Level 2, all five apply. For Level 3, they are habits on top of a formal curriculum. If two or more rows cannot be met for something that matters, the task belongs one level up.
What does this mean for the CEO and the student in you?
The CEO move is a scoping decision. Pick the level, write down what you own at that level, and write down what you will not touch. At Level 1 that is the specification and the decision to accept. At Level 2 it is the script, its data and its failures. Nobody should approve or run what they cannot check.
The student move is a method decision. At every level above zero, the scarce skill is reading and verification. Typing syntax is what the tools do well on bounded tasks. Judging whether the result is correct, secure and maintainable is what the studies show people overestimate in themselves.
What does this evidence not show?
- It is U.S. data, and projections are model outputs, not guarantees.
- The tools move faster than the studies. Any speed figure in Table 5 can age quickly.
- Samples are small or selected. Studies with 16, 47 and 52 people cannot be generalised, and the developer surveys are self-selected.
- The graduate data is from 2024. The New York Fed’s quarterly series, not broken out by major, put unemployment for all recent graduates at 5.2% in the third quarter of 2026.
- The test sorts typical situations, not individuals.
Frequently asked questions
Is learning to code still worth it in 2026?
For people who approve software, automations or AI agent output, learning to read and direct code (Level 1) is worth it. Building small tools is worth it when a recurring problem exists. Professional engineering remains well paid, with a harder entry point.
Will AI replace software developers?
The official projections do not say so. The Bureau of Labor Statistics projects 10% growth for software developers from 2025 to 2035 and a 7% decline for computer programmers, and names AI as a factor in both.
Should a non-technical professional learn Python or just use AI tools?
The two are not alternatives. Anyone who builds should learn enough of one language to read, test and limit what the AI wrote.
Is a computer science degree still a good choice?
The New York Fed data shows both sides: the fourth-highest unemployment rate of 73 majors (7.0%) and the second-highest early-career median wage ($87,000). The data cannot say what will hold for a student starting now.
Does using AI while learning to code hurt learning?
In the one randomized experiment that measured it, the AI group scored 17% lower on a follow-up quiz, with the damage concentrated among people who delegated. The study is small and awaits replication.
Sources
- U.S. Bureau of Labor Statistics. Occupational Outlook Handbook: Software Developers, Quality Assurance Analysts, and Testers. May 2025 wages, 2025-35 projections.
- U.S. Bureau of Labor Statistics. Occupational Outlook Handbook: Computer Programmers. May 2025 wages, 2025-35 projections.
- U.S. Bureau of Labor Statistics. Occupational Outlook Handbook: Web Developers and Digital Designers. May 2025 wages, 2025-35 projections.
- U.S. Bureau of Labor Statistics. Employment Projections: 2025-2035 Summary. News release USDL-26-1422, August 27, 2026.
- Federal Reserve Bank of New York. The Labor Market for Recent College Graduates: outcomes by major (released February 4, 2026, 2024 data) and 2026:Q3 highlights (released November 2, 2026).
- Brynjolfsson, E., Chandar, B., and Chen, R. Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab, August 2026 version (ADP data through June 2026).
- Peng, S., Kalliamvakou, E., Cihon, P., and Demirer, M. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. arXiv 2302.06590, 2023. Authors include Microsoft and GitHub employees.
- Cui, K. Z., Demirer, M., Jaffe, S., Musolff, L., Peng, S., and Salz, T. The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. Working paper, February 2025. Authors include Microsoft employees.
- Becker, J., Rush, N., Barnes, B., and Rein, D. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR, arXiv 2507.09089, July 2025.
- Becker, J., Rush, N., Cunningham, T., Rein, D., and Mahamud, K. We are Changing our Developer Productivity Experiment Design. METR, February 24, 2026.
- Perry, N., Srivastava, M., Kumar, D., and Boneh, D. Do Users Write More Insecure Code with AI Assistants? ACM CCS 2023; arXiv version 3.
- Shen, J. H., and Tamkin, A. How AI Impacts Skill Formation. arXiv 2601.20245, February 2026. Authors affiliated with Anthropic, an AI company.
- Scherer, R., Siddiq, F., and Sanchez Viveros, B. The cognitive benefits of learning computer programming: A meta-analysis of transfer effects. Journal of Educational Psychology, 111(5), 764-792, 2019 (abstract only).
- Stack Overflow. 2025 Developer Survey, AI section. Self-selected company survey.
- Stack Overflow. 2026 Developer Survey, AI section. Self-selected company survey.
Tables 2 and 4 are CEOtudent analyses recomputed from BLS and New York Fed figures. Table 5 is a CEOtudent line-up of published results; the perception gap is our arithmetic. Tables 6 and 7 are CEOtudent editorial frameworks. Deliberately not used: the widely repeated claim that employment of young software developers fell about 20% from its late-2022 peak, because the August 2026 edition of the Stanford paper that we opened does not print an occupation-specific figure in its main text; the 13% headline from that paper’s first edition, which the authors replaced with the 19% measure; hour estimates for each level; and vendor-reported productivity figures that could not be opened in full.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:















