The State of AI-Generated Markdown: What the 2026 Data Shows

I started this post wanting to open with a big number: the share of software documentation now written by AI. Every stats page quotes it: "60 to 80% of docs and comments are AI-written." I couldn't trace it to a single primary source. It hops from one SEO blog to the next, methodology never attached. So I dropped it.

What's left is smaller and, I think, more useful: the numbers I can actually stand behind about AI-generated Markdown, a shift that's real even without the inflated headline. AI writes an enormous and growing share of what developers produce, and it hands most of it back as Markdown. The plans, the specs, the config files, the reports, the reviews. You didn't write them. You still have to read them.

TL;DR — the 2026 picture, sourced:

1. AI writes most of the new code. 75% of new code at Google is AI-generated (Pichai, April 2026); Copilot writes ~46% where enabled.
2. The output format converged on Markdown. Agent instruction files, plans, and specs are .md; Cloudflare now serves Markdown to agents at the edge.
3. A config file became a standard. AGENTS.md is in 60,000+ repos and lives under the Linux Foundation as of December 2025.
4. Humans still read every line. 66% of engineers won't merge AI code without a manual review (SonarSource, 2026); only 32% trust AI output (Stack Overflow, 2025).
5. "Markdown is the new source code" stopped being a metaphor (Hartley Brody, March 2026).

Key statistics at a glance

Every number below is pulled from the sections that follow and attributed to the primary source named there. External links are rel="nofollow".

Statistic	Value	Source
New code at Google that is AI-generated	75% (up from ~50% in fall 2025)	Google / Sundar Pichai, Apr 2026
Code written by GitHub Copilot where enabled	~46% (up to 61% in Java)	GitHub / Accenture
Developers using AI tools daily	50.6%	Stack Overflow Survey 2025
Developers using or planning to use AI	84% (up from 76%)	Stack Overflow Survey 2025
Developers who distrust AI output	46% (only 32% trust it)	Stack Overflow Survey 2025
Engineers who won't merge AI code without manual review	66%	SonarSource, 2026
Repos containing an AGENTS.md file	60,000+	Linux Foundation, Dec 2025
AI-assisted code reviews on GitHub	60M+ (now 1 in 5 reviews)	GitHub, Mar 2026
Token cut when serving Markdown vs HTML to agents	~80% (16,180 → 3,150 on a sample page)	Cloudflare, Feb 2026

How much of a codebase does AI actually write now?

Close to half, and at some companies far more. In 2026, Sundar Pichai told Google Cloud Next that 75% of new code at Google is AI-generated, up from around 50% the previous fall. That's one company with unusual tooling, but the wider signal points the same way.

GitHub and Accenture's enterprise research puts Copilot at roughly 46% of the code written in files where it's turned on, rising to 61% in Java. Stack Overflow's 2025 survey of about 49,000 developers found 50.6% now use AI tools every day, and 84% use them or plan to.

By 2026, AI co-authors close to half the code a typical developer ships, and three-quarters of new code at Google (Google, April 2026). The volume of machine-written text landing in repositories every day has no precedent, and almost none of it is prose — it's Markdown.

Here's the part the headline stats miss. When an agent writes code, it rarely stops at code. It writes a plan first, a spec alongside it, a summary after, and a review comment on the pull request. All of that arrives as .md.

Why did AI agents standardize on Markdown?

Because Markdown is plain text a machine can parse and a human can read without a renderer getting in the way: diffable, version-controllable, and cheap in tokens. The industry didn't vote on it; the tools converged on it.

The clearest tell is at the infrastructure layer. In February 2026, Cloudflare launched Markdown for Agents, which converts any HTML page to Markdown at the edge when an agent asks for it. On their sample page that dropped the payload from 16,180 tokens to 3,150 — about 80% less. GitHub, meanwhile, has been documenting spec-driven development, where a Markdown spec is the thing you edit and the agent implements from it.

Developer Hartley Brody named the shift in March 2026: "Markdown is the new source code." It sounds like a slogan until you watch a coding session. The file you're actually steering is the plan, not the diff.

What does AI actually write in Markdown?

Far more than config files. The category people fixate on, agent instruction files like CLAUDE.md and AGENTS.md, is the smallest slice. In practice, the .md an agent generates spans:

• Instruction files — CLAUDE.md, AGENTS.md, .cursorrules, copilot-instructions
• Plans and specs — plan-mode output, PRDs, design docs
• Reports and analyses — audits, research summaries, migration write-ups
• Docs and READMEs — generated or updated automatically
• Review artifacts — pull-request descriptions and code-review comments
• Conversation exports — the transcript of the session itself

Two academic passes give a sense of the config-file layer alone. A 2025 study analyzed 253 CLAUDE.md files and found 77.1% cover build-and-run instructions, 71.9% implementation details, and 64.8% architecture (arXiv 2509.14744). A follow-up widened the lens to 2,303 agent context files across 1,925 repositories. And that's just the files developers commit — it excludes the plans, reports, and summaries that never reach version control but still land in your editor.

How did AGENTS.md become a standard so fast?

It went from one vendor's convention to a Linux Foundation project in about sixteen months. OpenAI released the AGENTS.md format in August 2025 as a shared, tool-agnostic instruction file. By December 2025 it was folded into the new Agentic AI Foundation, with AWS, Anthropic, Cloudflare, Google, Microsoft, and OpenAI as platinum members.

As of December 2025, AGENTS.md appears in more than 60,000 open-source projects and is read by tools including Codex, Cursor, GitHub Copilot, Gemini CLI, and Devin (Linux Foundation, 2025). A Markdown file is now the closest thing the industry has to a portable configuration standard for AI agents.

The fragmentation is real too — CLAUDE.md, .cursorrules, and copilot-instructions still coexist — but they're all the same idea in the same format. Plain Markdown, read by a machine, edited by a human.

If AI writes it, why do you still have to read it?

Because almost nobody trusts it unread. In SonarSource's 2026 survey of roughly 7,000 engineers, 66% said they won't merge AI-written code without a manual review. Stack Overflow's 2025 data agrees: only 32% of developers trust AI output, 46% actively distrust it, and 66% are worn down by code that's "almost right, but not quite."

So the review load didn't shrink when the writing sped up — it moved. GitHub reported 60 million-plus AI-assisted code reviews by March 2026, now one in five reviews on the platform. The plan you skim before approving, the spec you red-line, the report you check against reality — that's Markdown, and reading it well is now part of the job.

The productivity gain from AI writing is only banked once a human reads the output and signs off. With 66% of engineers refusing to merge unreviewed AI code (SonarSource, 2026), the bottleneck shifted from writing Markdown to reading it — which is a rendering problem, not a generation problem.

Raw .md doesn't help here. A plan full of ## hashes, pipe-delimited tables, and fenced code blocks is exactly the thing you're squinting at when you most need to spot the flawed assumption. Rendered, the same file surfaces the gap in seconds. If most of the Markdown you read now comes from an agent, it's worth reading it properly — see our notes on reading ChatGPT and Claude Markdown output on a Mac and viewing AI Markdown output, or the Quick Look approach for previewing files straight from Finder.

MacMD Viewer: native Markdown viewer for Mac with Mermaid diagrams and QuickLook. $19.99 one-time purchase →

The takeaway

The inflated "80% of docs are AI-written" stat isn't the story, because it isn't verifiable. The verifiable story is bigger anyway: AI now writes most of the code (75% at Google), hands the surrounding work back as Markdown, and a config file just became a Linux Foundation standard. What didn't change is the human at the end of it, reading the plan before it ships. Markdown became the new source code — so reading it well became a real skill, not an afterthought. For the full landscape of tools, our roundup of the best Markdown viewers for macOS is the place to start, and the 2026 AI productivity statistics put these numbers in wider context.

Sources

Every figure above was checked against its primary source on 2026-07-08.

• Google / Sundar Pichai, Cloud Next 2026 keynote, April 2026 · blog.google
• Stack Overflow, 2025 Developer Survey (AI section, ~49,000 respondents), December 2025 · survey.stackoverflow.co
• Linux Foundation, Agentic AI Foundation announcement (AGENTS.md, 60,000+ repos), December 2025 · linuxfoundation.org
• Cloudflare, Markdown for Agents, February 2026 · blog.cloudflare.com
• GitHub, "60 million Copilot code reviews and counting," March 2026 · github.blog
• GitHub / Accenture, quantifying Copilot's impact (~46% of code) · github.blog
• "On the Use of Agentic Coding Manifests" (253 CLAUDE.md files) · arXiv 2509.14744; "Agent READMEs" (2,303 files) · arXiv 2511.12884
• Hartley Brody, "Markdown is the new source code," March 2026 · blog.hartleybrody.com

Frequently Asked Questions

How much of new code is written by AI in 2026?

75% of all new code at Google is now AI-generated, up from about 50% in fall 2025, per CEO Sundar Pichai at Google Cloud Next on April 22, 2026. Across the wider industry, GitHub Copilot writes roughly 46% of code in files where it's enabled. AI now co-authors close to half the code developers ship.

What is AGENTS.md?

AGENTS.md is an open, cross-tool Markdown file that tells AI coding agents how a project works — build commands, conventions, constraints. OpenAI released the format in August 2025; it moved under the Linux Foundation's Agentic AI Foundation in December 2025 and now appears in 60,000+ open-source projects, read by tools including Codex, Cursor, Copilot, Gemini CLI, and Devin.

Is Markdown the standard output format for AI agents?

In practice, yes. Agent instruction files use it, coding agents emit plans and specs in it, and infrastructure now assumes it: Cloudflare's Markdown for Agents (February 2026) converts HTML to Markdown at the edge for crawlers, cutting a sample page from 16,180 tokens to 3,150 — roughly 80% fewer tokens.

Do developers still review AI-generated output?

Yes, heavily. In SonarSource's 2026 survey of about 7,000 engineers, 66% said they won't merge AI-written code without a manual review. Stack Overflow's 2025 survey of ~49,000 developers found only 32% trust AI output while 46% actively distrust it, and 66% are frustrated by answers that are "almost right, but not quite."

What does "Markdown is the new source code" mean?

It's a framing from developer Hartley Brody (March 2026): as AI agents plan, specify, and document work, the Markdown files they read and write become the artifact humans actually steer — the plan, the spec, the reviewable surface — rather than a byproduct of the code.