AI 已经摧毁了初级程序员的市场
2026 年 7 月 4 日
·
8 分钟阅读
2025 年初,我曾预测 AI 将催生大量、大量的程序员,并且新的编程工作会有所不同。三月份我进行了阶段性评估,发现初创公司正以前所未有的速度用算力替代人力,而新工作岗位的浪潮却毫无踪影。本文是下一次阶段性评估,我带来了好消息和坏消息。
坏消息是:AI 已经摧毁了初级程序员的市场。好消息是:我预测的新程序员长尾效应确实出现了,但有一个巨大的转折——他们并不自称程序员。让我给你看看数据,看看你是否同意我的看法。
年轻程序员的市场已经崩溃
下面这张关于 AI 与编程工作最重要的图表,基于 ADP 薪资数据,由斯坦福大学数字经济实验室制作。它追踪了按年龄划分的美国软件开发人员就业情况,以 2022 年 10 月为基准:
22 至 25 岁的开发人员数量较 2022 年末的峰值下降了 19%。同期,所有 30 岁以上年龄段的开发人员数量均有所增长,其中 41 至 49 岁人群增长了 14%。这并非个别公司的偶然现象:在控制了单个公司层面的冲击后,斯坦福团队仍然发现,在受 AI 影响的岗位中,年轻工人的相对就业率下降了 16%,而且这种下降主要集中在 AI 自动化工作而非增强工作的职业领域。软件开发只是其中最典型的例子。
其他数据也指向同一方向。初级软件岗位的招聘信息较 2022 年峰值下降了 28%。计算机科学毕业生的失业率目前为 6.1%,高于文科专业——这句话在 2019 年足以让你在任何职业咨询办公室被嘲笑。
该图表中值得注意的一个细节是:初级开发人员的曲线在 ChatGPT 推出时并未断崖式下跌。它在 ChatGPT 推出前几个月达到峰值,然后在 2023 年缓慢下降,接着在 2024 年和 2025 年初恶化速度最快——这正是编程助手从自动补全代码行转向自动完成整个任务的时间点。真正加剧这一趋势的是智能体编程,而非 ChatGPT。
当然,还有其他嫌疑因素。同一时期还经历了零利率政策(ZIRP)退出、第174条税法变更,以及疫情后的招聘回调,而2025年宣布的裁员中,只有约4.5%被实施裁员的公司归因于AI。但斯坦福大学的研究结果在控制了企业层面冲击和利率风险敞口后依然成立,而且这些干扰因素都无法解释为何损害如此精准地集中在22至25岁、从事AI可自动化职业的年轻人身上,而他们40岁的同事却发展良好。考虑到科技行业年龄歧视现象依然普遍存在,如果市场只是对程序员整体不利,你原本肯定会预期出现相反的情况。
然而其他指标并未下降
进一步证明这特指编程岗位的证据来自更广泛的经济层面——事实上,即使只看"计算机岗位"而不具体限定为编程岗位,也能发现这一点。
从2024年5月到2025年5月,美国总就业人数增长了0.8%。计算机和数学类职业增长了1.3%,快于整体经济增速。根据美国劳工统计局的数据,受雇的软件开发人员数量从2022年5月的153万人增至2025年5月的169万人,在整个AI时代增长了10%。美国、丹麦以及Anthropic公司自身的严谨研究均发现,AI接触度与总体就业之间没有关联;丹麦的研究利用政府工资记录,可以排除大于约1%的影响。
这两件事怎么可能同时成立?按各年龄段在劳动力中的占比进行加权,就能得到答案:
自2022年10月以来,开发人员总就业人数增长了4.4%。初级开发者(这里按年龄而非经验定义,这是一个重要前提)仅占开发人员劳动力的约8%,因此对他们而言的灾难几乎不会影响平均值。即使你将他们在劳动力中的占比翻倍,总体数据仍然保持正值。这就是为什么每项关注平均值的研究都一无所获,而每项关注初级开发者的研究都发现惨状——他们看的是同一组数据的不同部分。
消亡的是职位头衔,而非工作本身
当你观察哪些职位头衔在缩减时,情况变得更加有趣。同样是美国劳工统计局的数据,从2024年5月到2025年5月:
美国劳工统计局(BLS)将“计算机程序员”这一职业定义为按照他人规格编写代码的人,该职业人数在一年内下降了16%。BLS此前预测该职业每十年会减少6%。我所在的群体——网页开发者,人数下降了11%,质量保证测试员下降了6.5%。与此同时,数据科学家增长了12%,系统分析师增长了4.4%,而宽泛的“软件开发人员”类别则增长了2%。
正在消失的岗位,其工作产出是按规格编写的代码;而正在增长的岗位,其工作产出是关于应该编写什么代码的判断。AI正在吞噬一种非常特定的编程工作。
长尾效应出现了,只是它没有对应的职位名称。
早在2025年,我就写过,AI是一个新的抽象层,就像它之前的每一个抽象层一样,它会催生出数量多得多的开发者,去构建数量多得多的软件。我还写道,我们应该称这些新人为“软件开发人员”,因为给他们起别的名字,会在本不需要的地方制造门槛。
我相信我当时是对的——一大批全新的开发者已经涌现。但他们并没有使用那个头衔。
软件繁荣是真实且可衡量的。在最近一个Octoverse年度中,GitHub新增了3600万个账户,这是其有史以来最快的增长,每秒新增超过一名开发者,同时新增了1.21亿个仓库,这是该平台历史上创建仓库最多的一年,而这也体现在其基础设施的运转压力上。这些新用户中有80%在加入的第一周内就使用了Copilot。有史以来最大规模的单一开发者涌入,是以AI原生方式出现的,而恰在此时,付费初级岗位的招聘却崩溃了。
我最喜欢的证据是App Store,因为在iOS上发布应用是一项成本高昂且设有门槛的行为:99美元的开发者费用、审核流程、以及一个可运行的二进制文件。它衡量的是已交付的软件,而不是教程。
新应用提交量在2016年达到峰值后,连续八年下滑。2025年增长了24%,这是自峰值以来的首次真正增长;而2026年第一季度,iOS提交量同比飙升80%。增长幅度如此之大,以至于苹果的审核时间从两天延长到了数周。应用类别结构也发生了变化,生产力、工具和生活方式类应用占比上升——这完全符合预期:首次开发者们是在解决自己的问题,而非工作室追逐游戏收入。
这些人是谁?据Vercel称,63%的“氛围编码”用户是非开发者。Lovable表示其60%的用户是“非开发者”,而这些用户每天创建超过10万个新项目。Replit声称已有5000万人使用过其平台。他们是营销人员、创始人、教师、分析师和产品经理,他们正在编写软件——在我看来,这让他们成为了开发者。只是他们自己不这么认为,更重要的是,这不是他们的职位头衔,而劳动统计部门统计的正是职位头衔。
因此,新开发者的长尾如期大规模出现了。但它并非以单一职位头衔下的员工数量形式出现,而是作为一种能力渗透到了每一个职位头衔中。一位用“氛围编码”编写自己归因仪表盘的营销经理,在BLS数据中仍显示为营销经理。崩溃的是资质证书的市场,而开发活动本身正在蓬勃发展。
下一代资深开发者将从何而来?
因此,我对2025年的预测评分是:关于开发者的判断正确,关于职位头衔的判断错误。这听起来像是一个圆满的结局,直到你追问接下来会发生什么。
专业软件工程师的职业入门路径曾经是这样的:你被雇佣来编写平庸的代码,一位资深工程师审阅你的代码,你通过反复练习和纠错慢慢积累判断力,十年后你成了那位资深工程师。如今这条链条已经断裂。AI现在负责编写平庸的代码,因此没有人再雇佣初级开发者,也就没有人排队等着成为负责审阅代码的资深工程师。
与此同时,数百万新的开发者正在发布产品,却没有任何人进行审查。Veracode 的一项研究发现,45% 的 AI 生成代码未能通过基本的 OWASP 安全测试。一项对 vibe-coding 应用的审计发现,10% 的应用存在严重的行级安全漏洞,导致用户数据泄露。苹果公司正被大量其无法及时审查的提交内容所淹没。软件正在被构建出来,但判断层却跟不上节奏,而过去用来构建软件的机制——即雇佣关系下的学徒制——已经崩溃。
在初级开发者这片焦土上,出现了一些有希望的绿芽。IBM 正在将初级岗位的招聘数量增加两倍,其理论依据是,配备 AI 的初级开发者能够胜任以前高级开发者才能做的工作,并将初级岗位的职责围绕客户接触和需求规格说明而非编码进行重新设计。但另一方面,Salesforce 在上一财年没有招聘任何工程师。这是两种可能的未来,哪一种胜出将决定这个行业在 2036 年是否还有高级开发者。
转折点是否已经到来?
一个不招聘初级开发者的市场,其合乎逻辑的结果之一是我们会开始感受到痛苦并进行自我纠正。这种情况,或许,或许正在发生。Indeed 的职位发布数据实际上在 2025 年 5 月触底,此后连续十三个月上涨,同比增长 10%。
如果在斯坦福大学的下一份更新报告中,22 至 25 岁人群的就业曲线出现上升,那可能意味着市场已经找到了新的平衡点。要留意其他大型雇主是否会像 IBM 那样推出类似的计划。如果我们没有看到这些,那么我们就必须自己去创造它们,否则整个软件创造的繁荣都将化为泡影。
我们必须重建晋升阶梯
我们看到的不是编程的消亡。我们看到的是编程不再是一个职位头衔,而变成了一种能力,就像“打字员”在成为每个人都应掌握的技能后就不再是一个职位头衔一样。这种转变对所有人都还好,除了那些正准备开始攀爬旧阶梯,而我们却把它点燃了的人。正是这些人,我们欠他们一架新的阶梯,如果我们不为他们搭建,我们也将感受到痛苦。
资料来源
- Brynjolfsson、Chandar 与 Chen 合著,《煤矿中的金丝雀?》,斯坦福数字经济实验室,2025 年 11 月(年龄序列数据来自图 1 数字化结果)
- 美国劳工统计局职业就业与工资统计,全国数据文件,2022 年 5 月至 2025 年 5 月
- Indeed 招聘实验室职位发布追踪器(原始数据,截至 2026 年 6 月)
- GitHub Octoverse 2025
- Appfigures,App Store 发布数据:2025 年全年及 2026 年第一季度,数据来源 TechCrunch
- 耶鲁预算实验室,《评估人工智能对劳动力市场的影响》(持续更新)
- Humlum 与 Vestergaard 合著,《大语言模型,微小的劳动力市场影响》,NBER 工作论文 33777,2025 年
- Anthropic,《人工智能的劳动力市场影响》,2026 年
- Vercel、Lovable 和 Replit 的用户构成披露,由 Hostinger 和 Panto 整理;Lovable 安全审计数据来自 Taskade 的《Vibe Coding 现状报告》
- Veracode,《2025 年生成式 AI 代码安全报告》
- 纽约联邦储备银行,《应届大学毕业生的劳动力市场》;计算机科学毕业生失业率讨论来自 Stack Overflow 博客
- Oxford Economics / Challenger, Gray & Christmas 裁员归因数据,数据来源 Fortune
- IBM 入门级招聘及 Salesforce 工程冻结,数据来源 CNN Business
- 年龄组权重来自 Data USA 基于 ACS PUMS 2024 的数据;方法论及完整数据均可获取
人工智能
软件开发
编程岗位
劳动力市场
初级开发者
科技行业就业
人工智能影响
Laurie Voss 是一名开发者、写作者,也是 npm 的联合创始人(现已退出)。自 2001 年起撰写关于技术及其对人类意义的文章。
更多关于 Laurie 的信息 →
AI has torched the market for junior programmers
July 4, 2026
·
8 min read
In early 2025 I predicted that AI will create many, many more programmers, and that new programming jobs would look different. In March I checked in and found startups substituting compute for labor at record rates, with the wave of new jobs nowhere in sight. This post is the next check-in, and I have good news and bad news.
The bad news: AI has torched the market for junior programmers. The good news: the long tail of new programmers I predicted has materialized, but with a big twist: they don't call themselves programmers. Let me show you the data, and see if you believe me.
The market for young programmers has collapsed
Here's the single most important chart about AI and programming jobs, built from ADP payroll data by Stanford's Digital Economy Lab. It tracks employment of US software developers by age, indexed to October 2022:
Developers aged 22 to 25 are down 19% from their late-2022 peak. Every cohort over 30 grew over the same period, with 41-to-49-year-olds up 14%. This isn't a firm-level fluke: after controlling for shocks at the individual company level, the Stanford team still finds a 16% relative employment decline for young workers in AI-exposed jobs, and the decline concentrates specifically in occupations where AI automates work rather than augments it. Software development is just the poster child.
Other data points in the same direction. Entry-level software postings are down 28% from their 2022 peaks. Computer science graduates now have a 6.1% unemployment rate, higher than liberal arts majors, a sentence that would have gotten you laughed out of any career counseling office in 2019.
One detail worth noticing in that chart: the junior line doesn't fall off a cliff when ChatGPT launches. It peaks a couple of months before, drifts down through 2023, and then deteriorates fastest in 2024 and early 2025, which is when coding assistants stopped autocompleting lines and started completing tickets. Agentic programming is what really turned up the heat, not ChatGPT.
There are other suspects, of course. The same period saw the ZIRP unwind, the Section 174 tax change, and a post-pandemic hiring correction, and only about 4.5% of 2025's announced layoffs were actually attributed to AI by the companies doing the laying off. But the Stanford results survive controls for firm-level shocks and interest rate exposure, and none of those confounders explains why the damage is so precisely concentrated among 22-to-25-year-olds in AI-automatable occupations while their 40-year-old colleagues thrive. With ageism in tech being alive and well you would certainly have expected the opposite if the market were just tough for programmers in general.
And yet nothing else is down
Further evidence that this is specifically about programming jobs comes if you look at the wider economy, and in fact even if you look at only "computer jobs" without specifying programming specifically.
Total US employment grew 0.8% from May 2024 to May 2025. Computer and mathematical occupations grew 1.3%, faster than the economy. The count of employed software developers, per the BLS, went from 1.53 million in May 2022 to 1.69 million in May 2025, up 10% right through the AI era. Careful studies in the US, Denmark, and by Anthropic itself find no relationship between AI exposure and aggregate employment; the Danish study, using government payroll records, can rule out effects bigger than about 1%.
How can both things be true? Weight the age bands by their share of the workforce and you get your answer:
Total developer employment is up 4.4% since October 2022. Juniors (here defined by age rather than experience, a big caveat) are only about 8% of the developer workforce, so a catastrophe for them barely moves the average. Even if you double what percentage of the workforce you think they are, the aggregate stays positive. This is why every study that looks at averages finds nothing and every study that looks at juniors finds carnage. They’re looking at different parts of the same data.
The title is dying, not the work
It gets more interesting when you look at which job titles are shrinking. Same BLS data, May 2024 to May 2025:
The occupation "computer programmer," the BLS category for people who write code to someone else's specification, fell 16% in a single year. The BLS had projected that occupation to decline 6% per decade. My people, the web developers, fell 11%, and QA testers 6.5%. Meanwhile data scientists grew 12%, systems analysts 4.4%, and the broad "software developer" category grew 2%.
The jobs disappearing are the ones where the work product is code written to spec. The jobs growing are the ones where the work product is judgment about what code should exist. AI is eating a very specific kind of programming job.
The long tail showed up. It just doesn't have the job title.
Back in 2025 I wrote that AI is a new abstraction layer, and like every abstraction layer before it, it would create vastly more developers building vastly more software. I also wrote that we should call these new people "software developers," because giving them some other name would create gatekeeping where none needs to exist.
I believe I was right -- a huge new body of developers has turned up. But they don't use that title.
The software boom is real and it is measurable. GitHub added 36 million new accounts in the last Octoverse year, its fastest growth ever, more than one new developer per second, and 121 million new repositories, the biggest year for repository creation in the platform's history, and that’s shown up as their infrastructure creaking at the joints. Eighty percent of those new arrivals used Copilot within their first week. The single biggest developer influx ever recorded arrived AI-native, at exactly the moment paid junior hiring collapsed.
My favorite evidence is the App Store, because publishing an iOS app is a costly, gated act: a $99 developer fee, a review process, a working binary. It measures shipped software, not tutorials.
New App Store submissions declined for eight consecutive years after peaking in 2016. In 2025 they grew 24%, the first real growth since the peak, and in Q1 2026 iOS submissions were up 80% year over year. The surge is so large that Apple's review times have stretched from two days to weeks. And the category mix shifted toward productivity, utilities, and lifestyle apps, which is exactly what you'd expect from first-timers solving their own problems rather than studios chasing game revenue.
Who are these people? According to Vercel, 63% of vibe-coding users are non-developers. Lovable says 60% of its users are “non-developers”, and its users create over 100,000 new projects every day. Replit claims 50 million people have used its platform. These are marketers, founders, teachers, analysts, and product managers, and they are writing software, which in my book makes them developers. They just don't identify that way, and more importantly it's not their job title, and job titles are what labor statistics count.
So the long tail of new developers materialized, on schedule and at scale. But it materialized as a capability spreading through every job title instead of as headcount in one job title. A marketing manager who vibe-codes her own attribution dashboard shows up in the BLS data as a marketing manager. The market that collapsed is the market for the credential. The activity is booming.
Where does the next generation of senior devs come from?
So my 2025 prediction scores as: right about the developers, wrong about the title. Which sounds like a happy ending until you ask what happens next.
The career on-ramp for professional software engineers used to work like this: you got hired to write mediocre code, a senior engineer reviewed it, you slowly absorbed judgment through repetition and correction, and a decade later you were the senior engineer. That chain is now broken. AI now writes the mediocre code, so nobody hires the junior developer, so nobody is in the queue to become the senior who reviews things.
Meanwhile millions of new builders are shipping with no one reviewing anything. A Veracode study found 45% of AI-generated code fails basic OWASP security tests. An audit of vibe-coded apps found 10% with critical row-level security flaws exposing user data. Apple is drowning in submissions it can't review fast enough. The software is getting built. The judgment layer is not keeping up with it, and the mechanism that used to build it, apprenticeship inside employment, has collapsed.
There are a few promising green shoots in the scorched landscape for junior devs. IBM is tripling entry-level hiring on the theory that AI-equipped juniors can do formerly senior work, redesigning the junior role around customer contact and specification rather than typing. But on the other hand, Salesforce hired zero engineers last fiscal year. Those are the two candidate futures, and which one wins determines whether the profession has senior developers in 2036.
Is a turnaround already happening?
One logical outcome of a market that doesn’t hire junior developers is that we would start feeling the pain and correct ourselves. That maybe, maybe is already happening. Indeed's postings data actually bottomed in May 2025 and has risen for thirteen straight months, up 10% year over year.
If the 22-to-25 employment line turns upwards in Stanford's next update, it may be that the market has found a new equilibrium. Look for other major employers ramping up programs like IBM’s. If we don’t see them, we are going to have to create them, or this whole boom in software creation will turn to bust.
We have to rebuild the ladder
We are not watching the death of programming. We are watching programming stop being a job title and become a capability, the same way "typist" stopped being a job title when it became a thing everyone was expected to know. That transition is going fine for everyone except the people who were about to start climbing the old ladder when we set it on fire. They're the ones we owe a new ladder, and if we don’t build it for them, we will also feel the pain.
Sources
- Brynjolfsson, Chandar & Chen, "Canaries in the Coal Mine?", Stanford Digital Economy Lab, Nov 2025 (age series digitized from Figure 1)
- BLS Occupational Employment and Wage Statistics, national files, May 2022 through May 2025
- Indeed Hiring Lab job postings tracker (raw data, through June 2026)
- GitHub Octoverse 2025
- Appfigures, App Store release data: 2025 annual and Q1 2026 via TechCrunch
- Yale Budget Lab, "Evaluating the Impact of AI on the Labor Market" (rolling updates)
- Humlum & Vestergaard, "Large Language Models, Small Labor Market Effects", NBER Working Paper 33777, 2025
- Anthropic, "Labor market impacts of AI", 2026
- Vercel, Lovable, and Replit user-composition disclosures, compiled by Hostinger and Panto; Lovable security audit figures via Taskade's State of Vibe Coding
- Veracode, 2025 GenAI Code Security Report
- Federal Reserve Bank of New York, "The Labor Market for Recent College Graduates"; CS-grad unemployment discussion via Stack Overflow blog
- Oxford Economics / Challenger, Gray & Christmas layoff attribution data, via Fortune
- IBM entry-level hiring and Salesforce engineering freeze, via CNN Business
- Age-band weights derived from ACS PUMS 2024 via Data USA; methodology and full data are available
artificial intelligence
software development
programming jobs
labor market
junior developers
tech employment
ai impact
Laurie Voss is a developer, writer, and recovering npm co-founder. Writing about technology and what it means for humans since 2001.
More about Laurie →