I Read the Requirements on 69 'AI Engineer' Jobs. The Only Five That Ask for No AI Are at OpenAI and Cognition.

TL;DR: I pulled every live posting titled "AI Engineer", "Applied AI Engineer" or "AI Software Engineer" off 113 public Greenhouse and Ashby boards on August 15, 2026, and read the requirements section of all 69 of them. The running joke is wrong: 64 of 69 require a real AI skill. 28 want model work (training, fine-tuning, evals), 36 want application work on foundation-model APIs, and only 5 list no AI-specific requirement at all. The punchline is who those five belong to. Three are Cognition's Applied AI Engineer reqs, the company that makes Devin. Two are OpenAI's. Both put AI in the bonus block instead. This is a startup-and-tech sample from two ATS platforms, not the American labor market, so read it as a census of 25 named companies, not a national rate.

Indeed Hiring Lab reported on July 8, 2026 that 71% of the increase in US software development postings between May 2025 and May 2026 came from senior roles, and 37% came from roles with AI in the title. So "AI Engineer" is where the growth is. The standing practitioner complaint is that the title is decoration: a backend req with an AI paragraph stapled to the intro. I wanted the actual split, so I went and read them.

How did I build the sample?

I walked a fixed list of 230 board slugs, 126 on Greenhouse and 104 on Ashby, hitting the public JSON that both platforms serve without an account: boards-api.greenhouse.io/v1/boards/{slug}/jobs?content=true and api.ashbyhq.com/posting-api/job-board/{slug}. 113 boards resolved, 53 Greenhouse and 60 Ashby. The rest 404'd, which is what a wrong slug guess looks like, and I counted them as misses rather than swapping in replacements.

Slug collisions are real, so I verified identity before counting any row. Greenhouse returns a company_name on every job; Ashby does not, so I fetched the rendered board page for each contributing Ashby slug and read its title tag. All 25 contributing boards checked out as the company I meant: Anthropic, Databricks, OpenAI, Cognition, Stripe, GitLab, Grafana Labs, Snorkel AI, Scale AI, Arize AI, MongoDB, Klaviyo, Samsara, Gusto, Elastic, Amplitude, AssemblyAI, Alloy, Cohere, Deepgram, Gamma, n8n, Ramp, Sardine and WRITER.

That yielded 69 postings whose title contains "AI Engineer" or "AI Software Engineer". I excluded manager, director and architect titles. Then I read the requirements section of each one, and only the requirements section, because the intro paragraph is where every company sounds like an AI company.

One honesty note on the denominator: those 69 postings collapse to 51 distinct requisitions. Databricks posts the same forward-deployed req to six countries; Grafana posts one to five. I report both numbers throughout because the 69 is the number of things a job seeker sees and the 51 is the number of things a hiring manager wrote.

What did the requirements actually say?

Bucket	Postings	Distinct reqs
(a) Requires training, fine-tuning or evaluating models	28	20
(b) Requires building on foundation-model APIs (RAG, agents, evals)	36	28
(c) No AI-specific required skill at all	5	3

So the cynical read loses. 93% of these postings ask for something AI-specific in the requirements, and the split inside that is the more useful finding: application work outnumbers model work, 36 to 28. "AI Engineer" in August 2026 more often means someone who wires LLMs into a product than someone who trains one.

The bucket (a) reqs are unmistakable when you see them. Snorkel AI's Applied AI Engineer asks for "expertise across the Applied AI stack, spanning classical ML libraries (e.g., scikit-learn), deep learning frameworks (e.g., PyTorch), foundation-model ecosystems (e.g., Hugging Face Transformers), vector/embedding tooling (e.g., FAISS)... synthetic dataset curation, evaluation workflows." That is a real, checkable skill list. Nobody is padding there.

Who wrote an "AI Engineer" req with no AI in it?

Here is the part I did not expect. All five are at companies whose product is AI.

Cognition, which describes itself in the same posting as "an applied AI lab building end-to-end software agents" and "the makers of Devin, the first AI software engineer", lists these requirements for Applied AI Engineer: a STEM degree or equivalent, "3+ years as a software engineer, technical consultant, deployment strategist, forward deployed engineer, solutions engineer or similar roles with strong coding proficiency (Python, JavaScript/TypeScript, or similar)", communication skills, a track record of driving adoption, commercial instincts, and fast learning. No AI anywhere. AI appears one section later, under the heading of things that might make you excel: "Have deployed or integrated LLM or agent-based systems in production settings."

OpenAI's Applied AI Engineer, Plugins does the same thing. Required: 4-6 years of software engineering, production APIs, backend services, developer platforms, product sense, debugging distributed systems. Optional: "Familiarity with AI products, LLM APIs, tool calling, MCP, ChatGPT or Codex surfaces." Its sibling req, Messenger Integrations, files "Experience with AI product launches" under bonus too.

Read charitably, this is confidence rather than sloppiness. OpenAI and Cognition can assume anyone who applies has touched their models, and what they cannot assume is that a strong LLM tinkerer can hold a partner integration together across auth, privacy and reliability. So they gate on the scarce skill. The category error is ours: we read "AI Engineer" as a statement about the technology when it is increasingly a statement about the deployment context. That is the same collapse in meaning that happens when every company becomes an AI company.

How many name an actual framework?

21 of 69 name at least one specific tool or framework inside the requirements, and 30 of 69 name one anywhere in the posting. The most common by far is Hugging Face, at 7. Then Cursor at 4, MCP and GitHub Copilot at 2 each, LangChain at 2, PyTorch at 2, and single mentions of LangGraph and FAISS.

That distribution is the tell. Two-thirds of these reqs describe AI work without naming a single artifact you could put on a résumé and have verified. "Experience building AI-powered products" is not a skill, it is a vibe with a salary band attached. If you are trying to figure out whether a posting is real, the framework line is the fastest filter I found.

Seniority tracks Indeed's finding closely: 23 of 69 postings are senior, staff, principal or lead titled, a third of the sample, and the pattern in the reqs is that the junior-shaped ones are the ones asking for narrow model skills while the senior ones ask for judgment. Which is worth holding next to the fact that almost nobody has cleanly measured whether AI makes engineers faster.

What does this not prove?

Plenty. Greenhouse and Ashby skew hard toward funded startups and tech companies, and my 230-slug list was assembled by hand, which biases it further toward companies I could name. Nothing here supports a sentence of the form "X% of AI jobs in America." It supports "of 69 live AI Engineer reqs at these 25 companies on this date, five did not require AI."

It is also a single snapshot. I cannot tell you whether the 5-in-69 rate is rising or falling, because I ran it once. And a requirements section is a document, not a hiring decision: a company can write "must have fine-tuned a model" and hire the person who has not. All I measured is what they committed to in writing, which is more than the other end of the pipeline manages, where the claim that AI rejects 75% of resumes traces to no primary study at all.

The related question, whether AI experience is becoming a requirement across ordinary engineering jobs rather than AI-titled ones, needs a different corpus and a time axis. I ran that one on three Augusts of Hacker News hiring threads, and the answer there is much less flattering to the hype.

For now: the title mostly means something. Check the framework line, and be suspicious in the direction nobody expects, which is that the most AI-native employers wrote the least AI-specific reqs.