Key Takeaways
- AI tool use among developers climbed from roughly 70% in 2023 to 84% in 2025, with about half of professionals now using AI daily.
- Trust moved the opposite way. Only 29% of developers trusted AI output accuracy in 2025, down from 40% in 2024.
- TypeScript overtook Python and JavaScript on GitHub in August 2025, the largest language shift in more than a decade, driven by how well typed code pairs with AI tools.
- GitHub crossed 180 million developers, adding about 36 million in a year, roughly one new account every second.
- Nearly 80% of new GitHub developers used Copilot within their first week, making AI a baseline expectation rather than an advanced skill.
- More than 1.1 million public repositories now use an LLM SDK, up 178% year over year.
- “Vibe coding” entered the vocabulary in early 2025, yet roughly three-quarters of developers say it is not part of their professional work.
- Python adoption jumped 7 points in a year, anchored by its role in AI and data science.
Generative AI coding systems – artistic impression. Image credit: Alius Noreika / AI
Over the past five years, the AI boom turned coding assistants from a curiosity into the default starting point for software work, while leaving developers more skeptical of the output than when they began. The clearest measures come from GitHub and Stack Overflow. AI tool use rose from around 70% of developers in 2023 to 84% in 2025, daily use became routine, and on GitHub nearly 80% of new developers reach for Copilot within their first week. AI is no longer an advanced tool that veterans adopt late; it is baseline infrastructure that beginners expect from day one, the way they expect syntax highlighting or version control.
The second half of the story is the wariness that came with the adoption. As use climbed, trust fell. By 2025, only 29% of developers trusted AI output to be accurate, down from 40% a year earlier, and more developers actively distrusted accuracy than trusted it. That gap between heavy use and low confidence is the defining shift of the period: developers built AI into daily work, then learned exactly where it fails, and adjusted their habits around verification rather than blind reliance.
The five-year arc, in brief
The timeline runs from tool to teammate. GitHub’s Copilot launched in June 2021, built on OpenAI’s Codex, putting AI autocomplete in front of mainstream developers for the first time. ChatGPT’s late-2022 arrival made conversational coding ordinary. Through 2023 and 2024, model quality climbed and adoption widened. In February 2025, former OpenAI researcher Andrej Karpathy coined “vibe coding” to describe building software by chatting with a model. By late 2025 and into 2026, agentic tools like Claude Code and OpenAI’s Codex agent could read an entire codebase, plan changes, and run for hours, moving AI from a prompt-and-wait helper toward something closer to a junior engineer. The release of a free Copilot tier in late 2024 coincided with a surge in sign-ups that exceeded prior projections.
How AI rewired language choice
The most concrete fingerprint of the boom is in which languages developers reach for. In August 2025, TypeScript overtook both Python and JavaScript to become the most-used language on GitHub by monthly contributors, reaching about 2.64 million, a 66% year-over-year jump and the biggest ranking change in over a decade. The reason is not fashion. Strongly typed languages give AI tighter constraints, so a declared type eliminates whole classes of wrong suggestions, which produces more reliable generated code. GitHub describes a “convenience loop”: when AI makes a technology feel frictionless, more developers use it, which creates more training data, which makes the AI better at it still.
That loop shows up in unexpected places. Shell scripting saw a 206% rise in AI-generated projects, because an agent can write the tedious parts few developers enjoy. Python, meanwhile, accelerated its own growth with a 7-point adoption jump, holding its place as the language of AI research and data science even as TypeScript took the overall crown. The pattern is consistent: AI is changing not only how fast code is written but which tools developers choose to write it with.
| Metric | Then | 2025–2026 |
|---|---|---|
| Developers using/planning AI tools | ~70% (2023) | 84% |
| Trust in AI output accuracy | 40% (2024) | 29% |
| Most-used GitHub language | Python | TypeScript |
| Public repos using an LLM SDK | — | 1.1 million+ (+178% YoY) |
| New GitHub developers using Copilot in week one | — | ~80% |
| GitHub developers total | 100 million (2023) | 180 million+ |
What developers now want to learn
The boom redirected learning effort. In the most recent Stack Overflow survey, 69% of developers had picked up a new technique or language in the past year, 44% with help from AI tools, and 36% learned to code specifically for AI work. Interest concentrated on large language models, retrieval-augmented generation, and tooling around agents. Among models, Claude Sonnet ranked as the most admired large language model after Gemini Reasoning, while ChatGPT remained the most widely used assistant and GitHub Copilot the most common coding companion. Roughly 41% of all new code is now written with AI assistance, a figure that captures how far the practice has spread even where developers stay cautious.
Tool choice became its own discipline. Developers increasingly match the assistant to the task rather than forcing everything through one, choosing between Claude Code for deep multi-file work and Codex for fast routine jobs. A growing market of dedicated coding tools emerged to serve both professionals and newcomers.
The limits developers keep drawing
For all the adoption, developers have set firm boundaries. Roughly three-quarters say vibe coding is not part of their professional work, treating prompt-to-app generation as useful for prototypes rather than production systems. Agentic AI remains niche: a majority either avoid agents or stick to simpler assistants, and many have no plans to adopt them. The resistance is sharpest for high-stakes, systemic tasks like deployment and project planning, where most developers say they will not hand control to AI.
The friction is practical. About 66% of developers report that AI answers are almost right but not quite, and 45% lose meaningful time debugging AI-generated code. When the stakes rise, most still ask another person for help rather than trust the model. The emerging consensus is not anti-AI; it is “trust but verify,” with humans owning the output and applying extra scrutiny to anything a model wrote.
Where this leaves the developer
Five years in, the boom produced a developer who reaches for AI by reflex, chooses languages and tools partly on how well they pair with that AI, and keeps a careful hand on quality because experience has taught the limits. The center of gravity is shifting from writing every line to directing, reviewing, and verifying work that a model drafts. The tools will keep improving, and agentic systems are early in their impact, but the habit that defines the era is the balance developers have struck: rapid adoption paired with hard-won skepticism. That balance, more than any single tool, is what the AI boom has built into the craft.
If you are interested in this topic, we suggest you check our articles:
- 5 Best Vibe Coding Tools in 2026
- Best AI LLM for Vibe Coding: Complete 2026 Tier List
- Best Tasks for Claude Code vs OpenAI Codex
- AI Models and Their Features: ChatGPT vs Grok vs Claude
- xAI vs OpenAI vs Anthropic: Which AI Lab Wins in 2026?
Sources: Stack Overflow Blog, 2025 Stack Overflow Developer Survey, GitHub Octoverse, InfoQ, GitHub Blog, Stack Overflow for Leaders
Written by Alius Noreika
