How to learn artificial intelligence: what the job data actually asks for

If you want to learn artificial intelligence in 2026, the first thing worth knowing is that two different skills go by that name, and employers pay for both. One is building models and the systems around them, which runs on Python, some linear algebra and a great deal of debugging. The other is using models well inside a job you already have, which now shows up in postings for recruiters, paralegals and marketers. The free material for the first is genuinely good and finishable in a few months. The second has almost no curriculum at all, and the best evidence we could find says people badly overrate how good they already are at it.

What follows is a path assembled from documents we opened in August 2026: Lightcast’s job posting analysis for the Stanford AI Index, two Indeed Hiring Lab studies published in July 2026, the syllabuses and prices of five courses, three cloud certification pages, and a randomised trial on developer productivity. Every figure links to the page it came from. Where the sources disagree with each other, we say so rather than picking the tidier number.

What the job postings actually ask for

Start with how common these jobs are, because the number is smaller than the coverage suggests. Lightcast supplies the labour market data for the Stanford AI Index, and its summary of the 2026 edition, published on 13 April 2026, reports that 2.5% of all US job postings now mention AI skills. That’s up 55% from the year before, 72% from 2022, and nearly 300% over the past decade. So the growth is real and the base is still thin. Roughly 97 postings in every 100 ask for none of it.

Which skills are growing tells you more than the total does. Python was the most in-demand specialised skill in that analysis, appearing in 258,674 postings, up 391% on the 2013 to 2015 baseline and nearly 30% on 2024. Skills tied to agentic AI went from 0.06% of postings in 2024 to 0.23% in 2025, which Lightcast puts at nearly 90,000 US postings. That’s a category that barely existed two years before, and it’s the same shift we worked through in how to build an AI agent. Lightcast reads it as movement from experimentation toward execution, with deployment skills like Amazon Web Services, scalability and workflow management among the fastest growers.

One caution before you lean on any of these figures. The same Lightcast post gives the US share as 2.5% in one section and 2.6% in its country ranking, where Singapore leads at 4.769%, then Hong Kong at 3.5%, Luxembourg at 3.4% and Spain at 3.3%. A tenth of a point either way doesn’t change the argument, but it’s a reminder that posting counts are estimates built on keyword matching, not a census.

The other place the demand shows up is in job titles, and that’s where the shape changes. Indeed Hiring Lab’s Pawel Adrjan counted distinct AI-touched titles on 8 July 2026 and found 822 in the US by the first quarter of 2026, about 8.3% of all titles, up from 264 in 2022. Nearly two thirds of those titles sit in non-tech occupations. Europe runs the same pattern at a lower level, with one exception.

Market	AI-touched job titles, Q1 2026	Share of all titles	Share outside tech
United States	822	8.3%	63%
Germany	288	4.2%	59%
United Kingdom	160	2.7%	54%
France	138	3.3%	54%
Netherlands	84	2.2%	58%
Spain	81	2.3%	36%

AI in job titles is now more prevalent outside tech than in tech in five of the six markets we examined.

Pawel Adrjan, Indeed Hiring Lab, 8 July 2026

That finding is not new, and it’s worth checking against an older, larger count. Lightcast’s Beyond the Buzz report, released on 23 July 2025 off more than 1.3 billion job postings, put 51% of AI-skill postings outside IT and computer science, with generative AI roles in non-tech industries up 800% since 2022. It also priced the skill: postings that ask for AI skills advertise 28% higher salaries, which the company describes as nearly $18,000 more a year. Cole Napper, its VP of research and insights, framed the consequence for employers rather than for learners: firms treating AI as a niche technical skill end up bidding for talent against firms that have spread AI literacy across the whole workforce.

The entry-level door is the one that’s closing

A wage premium is only useful if you can get through the door to collect it, and that’s where the picture turns. Indeed Hiring Lab’s Guillermo Gallacher tracked software development postings in a companion piece on 8 July 2026. US software development postings have grown by almost 15% since Claude Code launched in late February 2025, while overall job postings fell by 7% over the same stretch. Software is outperforming the wider market by a long way.

The catch is in the composition of that growth. Gallacher reports that 71% of the increase in software development postings between May 2025 and May 2026 came from senior roles, and 37% from jobs that mention AI in the title. Meanwhile the same postings sit about 27.5% below their pre-pandemic level. A recovering market that recovers mostly at the top is a hard market for anyone who’s learning, which is the pattern we found in the hiring data when entry-level roles fell while total employment held.

Gallacher also notes a reversal worth holding onto, because it cuts against the simple story. Between May 2022 and May 2026, the more exposed to AI an occupation was, the more its postings declined. In the 2025 to 2026 window that relationship flipped, and more exposed occupations rebounded more on average. One year isn’t a trend, and Indeed’s published tracker counts the share of postings carrying AI keywords rather than hires made, so it’s a signal that the destruction reading was too neat rather than proof of anything.

Official projections are steadier and slower, which is their value. The US Bureau of Labor Statistics puts data scientist employment at 245,900 jobs in 2024, growing 34% to 2034, an increase of 82,500 and about 23,400 openings a year once transfers and retirements are counted. Median pay was $112,590 in May 2024, and the typical entry-level requirement is a bachelor’s degree. Those numbers cover one occupational code, not the whole field, and a ten-year projection made in 2024 can’t have priced in much of what happened since.

A curriculum you can actually finish, and what each part costs

The good news is that the teaching material is not the bottleneck, and hasn’t been for years. Four of the five best-known courses are free, and the fifth costs less per month than a takeaway. What varies is the entry price in prerequisites, which is the thing course marketing tends to bury.

Course	Cost	Size	Stated prerequisite
fast.ai, Practical Deep Learning for Coders	Free	9 lessons of about 90 minutes, plus 25+ in part 2	A year of coding, preferably Python, plus high school maths
DeepLearning.AI, Machine Learning Specialization	$25/mo billed annually, $30/mo monthly, or via Coursera	3 courses, 151 video lessons, 94 hours 58 minutes	Basic coding and high school algebra
Hugging Face, LLM Course	Free, no ads	12 chapters in four parts	Good knowledge of Python, better after an intro deep learning course
Karpathy, Neural Networks: Zero to Hero	Free, MIT licence	8 video lectures with code	None stated on the repository
Stanford CS231n, Deep Learning for Computer Vision	Materials public, enrolment is Stanford only	10 weeks, assignments and a final project	Python proficiency, college calculus, linear algebra, basic probability and statistics

Read down the prerequisite column and the split is obvious. fast.ai asks only that “you know how to code (a year of experience is enough), preferably in Python, and that you have at least followed a high school math course”. Jeremy Howard, who leads fastai development and was Kaggle’s top-ranked competitor, teaches it across computer vision, natural language processing, tabular analysis and collaborative filtering, with the from-scratch training algorithms held back for part 2. Andrew Ng’s specialisation sets a similar floor at basic coding and high school algebra, and it’s the most measurable of the five, at 151 video lessons and 42 graded assignments across roughly 95 hours.

CS231n asks for something else entirely. Its prerequisites are Python proficiency, college calculus and linear algebra, and a statistics course. That’s a real gap, and it’s the gap between using the tools and reading the papers behind them. Karpathy’s eight lectures are the cheapest bridge across it we found, because they build the machinery in order: backpropagation from scratch in lecture one, then a language model, then a transformer in “Let’s build GPT: from scratch, in code, spelled out”, then “Let’s build the GPT Tokenizer”. That last one covers the layer that explains a surprising share of model behaviour, which we unpicked in why a model can’t count the letters in strawberry.

Nobody is short of courses. What people run out of is the prerequisite the syllabus mentions in one line and assumes you already have.

Rundowns AI

Two of these courses also date faster than the others. The Hugging Face course is tied to specific libraries, Transformers, Datasets, Tokenizers and Accelerate, so its chapters move when the libraries do. Its own stated prerequisite is “a good knowledge of Python”, and it says outright that it’s better taken after an introductory deep learning course. So sequencing matters more here than picking a winner.

Certifications cost less than the study time, and they expire

Certifications are the part of learning AI that looks most like a credential and behaves least like one. The exam fees are small next to the hours, and the recommended experience on the vendor pages assumes you already did the job the certificate is meant to prove. The clearest evidence that these are product credentials rather than skill credentials is what happens to them over time.

Credential	Fee	Format	Recommended experience	Status in August 2026
AWS Certified Machine Learning Engineer, Associate	$150	65 questions, 130 minutes	At least 1 year using Amazon SageMaker and other ML engineering AWS services	MLA-C01 last sits in English on 28 September 2026; MLA-C02 registration opens 1 September 2026
Google Cloud Professional Machine Learning Engineer	$200	50 to 60 questions, two hours	3+ years of industry experience, including 1+ year on Google Cloud	Active, no formal prerequisites
Microsoft Certified: Azure AI Engineer Associate	Not listed on the page	100 minutes	Python or C#, REST APIs and SDKs, responsible AI principles	Retired, along with its renewal assessment

Google’s page is the blunt one. It lists no prerequisites at all, then recommends “3+ years of industry experience including 1 or more years designing and managing solutions using Google Cloud”. AWS is milder, at one year of SageMaker, but it’s mid-swap: the current MLA-C01 exam stops being offered in English on 28 September 2026, and registration for its replacement opens on 1 September. Microsoft’s Azure AI Engineer Associate has gone further and simply retired, renewal assessment included.

That’s the whole argument against treating a certificate as the goal. One of these three credentials has retired outright, and another swaps its exam code in September 2026. The underlying maths did neither. Which doesn’t make certifications worthless, because a hiring filter that scans for a vendor name is a real filter and $150 is a cheap way past it. It does mean the certificate is a receipt for a product version, and receipts age.

Using AI is its own skill, and people misjudge how good they are at it

The second skill, the one that shows up in those non-tech job titles, has the weakest evidence base and the loudest marketing. The most useful study we found on it is a randomised controlled trial from METR, published on arXiv in July 2025. Sixteen experienced open-source developers completed 246 tasks in projects they’d worked on for an average of five years. Each task was randomly assigned to allow or forbid the AI tools at the early 2025 frontier, mostly Cursor Pro with Claude 3.5 and 3.7 Sonnet.

Before starting tasks, developers forecast that allowing AI will reduce completion time by 24%. After completing the study, developers estimate that allowing AI reduced completion time by 20%. Surprisingly, we find that allowing AI actually increases completion time by 19%.

Becker, Rush, Barnes and Rein, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity

Read the middle sentence twice, because it’s the one that bears on learning. The developers still believed they’d been sped up by 20% after living through the slowdown. Economists forecast 39% shorter and machine learning experts 38% shorter, so the misjudgement wasn’t confined to the participants. The sample is 16 people on mature repositories they knew well, which is a narrow slice, and the tools have moved on since early 2025. Even so, it’s the only randomised trial on the question we could find, and it points somewhere uncomfortable: self-assessment of AI fluency is unreliable in exactly the direction that flatters the user.

Which is why the vendor documentation is more honest than the course market on this point. Anthropic’s prompt engineering overview doesn’t open with techniques at all. It assumes you arrive with “a clear definition of the success criteria for your use case”, a way to test against them empirically, and a first draft prompt, and it sends you to build evaluations if you don’t have those. It also says plainly that “not every success criteria or failing eval is best solved by prompt engineering”, and points at model selection instead for latency and cost. The measurable part comes first and the wording comes second, which inverts how most prompt courses are sold.

So the honest version of this skill is narrow and testable. It’s knowing what your task’s success looks like, running the same task with and without the tool, and keeping the result. We covered the technique layer separately in 12 prompting techniques that hold up, and the tool layer in what earns a place in a developer’s stack. Neither is worth much without the measurement, which is the part the METR trial shows people skip.

What the education data says about where people are learning

None of this is happening inside universities, or at least not mostly. The Stanford HAI 2026 AI Index education chapter records master’s degrees in AI software fields rising 17% from 2023 to 2024, and AI PhDs in the US and Canada up 22% from 2022 to 2024. At the same time computer science enrolment at US four-year universities fell 11% between 2024 and 2025. So specialisation is growing while the general pipeline shrinks, and those two facts belong in the same sentence more often than they get put there.

The chapter’s finding on where skills come from is the one that matters for anyone reading this without a degree. It reports that people are acquiring AI skills outside formal education, through certificates, online learning and work itself, and that AI literacy has grown faster than engineering-focused skills in most countries. The exceptions it names are the UAE, Chile and South Africa, where engineering skills have grown more steeply since 2022. It also finds AI skills being advertised on resumes, which is a claim about what people say they can do rather than a test of whether they can.

One more number from that chapter puts the demand in perspective. Four out of five US high school and college students now use AI for schoolwork, while only half of middle and high schools have an AI policy and just 6% of teachers say those policies are clear. The usage arrived before the instruction did. That gap is the reason “learn AI” resolves to a marketing page so often: the demand is enormous, the syllabus is contested, and nobody agrees on what the finish line looks like.

What would change this plan

Three things would, and they’re all measurable. The first is the entry-level composition in Indeed’s data. If the share of software growth coming from senior roles drops well below 71%, the case for a long unpaid study runway weakens, because junior work would be absorbing people again. The second is the agentic AI skill share. It went from 0.06% to 0.23% of postings in a year, and a category growing that fast either becomes a standing requirement or gets folded back into ordinary software work within two more.

The third is the replication of the METR result on current tools. Sixteen developers and 246 tasks is a small trial by any standard. A larger study on 2026 models that found a genuine speedup would move the “using AI” skill from unproven to trainable, and it would justify the course market that’s already selling it. Until one lands, the defensible position is that model fluency has to be measured per task rather than assumed.

What stays put through all three is the boring half: Python, which appears in more AI postings than any other specialised skill, the linear algebra and calculus that CS231n asks for, and the habit of testing a claim before repeating it. None of that expired when Azure’s AI engineer certification did. That’s the reason it’s worth the year.