[ free to use / updated 2026-11 ]
Free learning resources and research findings on AI, gathered in one place and sorted into six skills anyone can learn.
2,345 resources1,103 publishers40,744 papers
[ overview ]
AI tools change every week and the advice around them is loud. HumanWide gathers learning resources in one place, shows what has been checked, and helps you find a practical starting point.
[ library ]
Courses, guides and videos from 1,103 publishers. 1,990 are completely free. Currently machine indexed (L0); not yet independently reviewed.
[ skills ]
From stating what you need to checking the answer.
[ research ]
Study-specific findings, linked to their sources. Quote checks establish where words came from, not whether a conclusion is correct.
[ try a skill ]
Pick a skill, then switch between before and after. The conversations are short examples.
what changed
[ explore ]
Pick the person most like you for a short practice route, a starter prompt and a result you can keep.
[ optional reading / full courses take longer ]
[ library ]
Type a word, then narrow by skill, level or cost. Results update as you type.
2,345 of 2,345 resources
Review status describes how a resource was checked. A link check date does not mean a person reviewed its content.
[ research ]
One finding for each skill, from studies published in 2025 and 2026. Open a card for the details, or browse all 40,744 papers.
These are study-specific findings, not universal guarantees. Checking that a quote appears in a paper does not validate its conclusion or show that it applies to your situation.
[ verify results ]
ai alone ################### 94.9% with people ####### <34.5%
Tested alone, AI models named the right medical condition in 94.9% of cases. When 1,298 people used the same models, they found it in fewer than 34.5%, no better than people without AI.
Bean et al., 2025 · arXiv:2504.18919
[ understand how ai works ]
-17%quiz score after learning with AI help
Developers who learned a new library with AI help scored 17% lower on a quiz about it and were not significantly faster. Those who stayed engaged, asking for explanations, kept their learning.
Shen and Tamkin, 2026 · arXiv:2601.20245
[ state the need ]
up to +54%alignment scores, compared with baselines
Non-experts who answered a few small, structured questions about their goal produced expert-level product requirement documents.
Zhou et al., 2026 · arXiv:2602.04210
[ give the right context ]
+8 to 22 ptsaccuracy after rewriting the question
Leading models answered fewer than half of real people's questions about images correctly. Rewriting the same questions explicitly raised accuracy by 8 to 22 points.
Choi et al., 2026 · arXiv:2601.06165
[ pick the tools ]
+26.3 ptsaccuracy when large models answer briefly
On 7.7% of benchmark problems, smaller models beat larger ones. Asking the large models to answer briefly raised their accuracy by 26.3 percentage points.
Hakim, 2026 · arXiv:2604.00025
[ let ai act ]
45people in a driving-simulator study
An in-car assistant that showed its planned steps and partial results felt faster, earned more trust and lowered workload compared with one that worked silently.
Kirmayr et al., 2026 · arXiv:2602.15569
[ my toolbox ]
Press + save on a learning skill, finding or resource. Turn your collection into a learning checklist, a work guide, or an Agent Skill for your coding tools. Your collection and route progress stay in this browser.
01 / what goes in
02 / choose an output
A practice checklist for you to work through, with your saved resources.
03 / name your Agent Skill
04 / install for
[ method ]
A small pipeline that runs on one laptop and a modest budget. Its scope and limits.
[ discover ]
A polite crawler and public research APIs gather candidates, following each site's robots rules.
[ screen ]
Each resource is checked and labelled by skill, level, audience and cost.
[ organise ]
Six skills, three levels and eleven audiences.
[ read ]
A low-cost model reads every abstract, then codes each relevant paper against the same questions, from its full text when that is open.
[ verify ]
Each finding keeps a short quote, and code checks that it is really in the paper.
96,425papers screened by title and abstract
44,889papers coded by a model, from their full text, abstract or first pages
98.5%of quotes found in the source text
98.5% measures whether extracted quotes were found in the source text. It is not a measure of summary accuracy, study quality or how often an AI answer is correct.
[ cite / contribute ]
@misc{humanwide2026,
title = {HumanWide: A Calm, Clear Path to Using AI Well},
author = {{HumanWide contributors}},
year = {2026},
howpublished = {\url{https://humanwide.org}},
note = {Living collection, updated November 2026}
}
Resources belong to their publishers. HumanWide links to them and describes them in its own words.
[ questions ]
Yes, and so are most of the resources it points to: 1,990 of the 2,345 cost nothing, and every one shows its cost before you click.
Anyone. Paths start from eleven kinds of people, from students and teachers to health workers, public servants and people who are simply curious.
Widely read AI papers from Hugging Face Daily Papers, the most-cited work on OpenAlex, and studies of people using AI from arXiv, PubMed, ERIC and the ACL Anthology.
Resource cards show their review tier: L0 is machine indexed, L1 has been cross-checked by a model from another vendor, and L2 has been reviewed by a person. The current resource catalog is L0. Research quotes are checked against the source text; that check does not validate the study's conclusions. Read the linked original and consider its setting and limits.
A learning checklist helps you practice. A work guide gives you checks for your next task. The advanced Agent Skill export is a folder with a SKILL.md file that tells a compatible coding agent how to work. You can download all three directly in your browser.
