Home/AI Literacy for Everyone
AI Literacy for Everyone
What AI is, how models learn and speak, why they err, and AI's place in work and society — for everyone.
What AI is, how models learn and speak, why they err, and what AI means for work and society — no math, no coding. Elements of AI, UNESCO frameworks, Stanford HAI and more: 53 resources (50 free).
AI Fluency: Framework and Foundations — Claude Academy
A free 14-lesson course built with professors Rick Dakan and Joseph Feller around the 4D framework: Delegation, Description, Discernment and Diligence. You decide when AI fits a task at all, how to disclose its role to each audience, and how to stay responsible for AI-assisted work. It closes with a worksheet that assembles a signed personal AI-use policy — roles per task, disclosure commitments, revision triggers and a stress test. No paid subscription needed to complete it.
Ethics of AI — University of Helsinki MOOC
A free open online course from the University of Helsinki that teaches ethical thinking about AI through real cases rather than slogans. You work through what AI ethics means, what values and norms govern development, and how to start reasoning about AI from an ethical point of view — no coding required. Built with the Cities of Helsinki, Amsterdam and London and Finland's Ministry of Finance; worth two ECTS credits and completable on a slow connection.
Machine Learning for All (University of London, Coursera)
The University of London's Machine Learning for All teaches the AI taxonomy without programming: what machine learning is, how it differs from hand-written rules, and where supervised, unsupervised and reinforcement learning sit. Case studies and a drag-and-drop model builder make the categories concrete instead of abstract. Built for complete beginners and non-programmers. Free to audit on Coursera, with an optional paid certificate.
AI competency framework for students
UNESCO's AI Competency Framework for Students: 12 competencies across four dimensions (human-centred mindset, ethics of AI, AI techniques and applications, AI system design), each at three progression levels. It gives education systems a researched structure for what students should understand about AI at each stage, including responsible use and disclosure. The global reference for how children and students should learn to live with AI. Free from UNESCO, the UN agency for education.
AI competency framework for teachers
UNESCO's AI Competency Framework for Teachers: 15 competencies across five dimensions, from human-centred mindset to AI pedagogy and professional learning, with progression levels for training programs. It defines what teachers must master to use AI responsibly in teaching while protecting students' rights and their own professional roles. The teacher-side companion to the student framework, used to design national training programs. Free from UNESCO.
AI Literacy Lessons for Grades K-12
Common Sense Education's collection of grab-and-go AI literacy lessons for grades K-12, each 20 minutes or less, covering how AI works, chatbot design, bias, AI friendship and responsible use. Lessons include videos, discussion guides and activities built for real classrooms. The most widely used free AI literacy lesson set in US schools, research-backed and classroom-tested. Free for every teacher, school and district.
Build a Large Language Model (From Scratch) — Sebastian Raschka
Sebastian Raschka's Manning book that builds a working GPT-style language model step by step in Python: tokenizing text, creating training batches that predict the next token, implementing attention, pretraining, then fine-tuning for instruction following. By the end the reader has implemented the exact mechanism behind products like ChatGPT, turning next-token prediction from a slogan into lived experience. Aimed at developers with basic Python. Paid book with a free companion repository.
AI Literacy Toolkit: AI Ethics & Policy — NEIU Libraries
A university library guide dedicated to creating your own AI-use policy across conflicting contexts. It walks through analyzing institutional policies, defining personal core principles, building documentation and disclosure standards, and resolving the hard case where courses and workplaces disagree. Includes worked sample policies and a policy-development challenge with scenarios. Free, self-paced and written for students and professionals alike.
AI Literacy — Questions & Answers (European Commission)
The European Commission's official Q&A on the AI Act's AI-literacy obligation: what providers and deployers must do, why no fixed knowledge level is mandated, and the compliance steps organisations should consider. It lays out the timeline — obligation applicable since 2 February 2025, national enforcement from August 2026 — and links the living repository of literacy practices. Written for organisations but the clearest public answer to "why is AI literacy becoming law?". Free.
Cloudflare Learning Center — AI Inference vs. Training
Cloudflare's Learning Center lesson on inference versus training explains the build-once-answer-many split in plain language: what a training run is, what happens at inference time, and why the two have different cost, latency and energy profiles. Short, cross-linked and written for non-specialists. Maintained by an infrastructure company at the center of AI delivery. Free to read with no account.
Guidance for generative AI in education and research (UNESCO)
UNESCO's first global guidance on generative AI in education, addressing the questions every teacher and institution is facing. It covers the cheating-versus-learning debate, disclosure and age-appropriate use, data-privacy mandates for tool providers, and how to validate AI systems for pedagogical fit. You get concrete policy frameworks for schools plus creative uses in curriculum, teaching and research. Free, aimed at ministries, institutions, teachers and researchers.
How to generate text: decoding methods with Transformers (Hugging Face)
Hugging Face's classic long-form guide to what happens after a model scores the next token: greedy search, beam search, top-k, top-p (nucleus) sampling, and temperature, each demonstrated with runnable code and side-by-side generated text. It shows concretely why the same prompt produces different answers and where randomness helps versus hurts. Written for practitioners but readable by anyone wanting the mechanism behind 'creativity' sliders. Free, with companion notebooks.
Pretraining vs. Finetuning vs. Instruction Finetuning (Sebastian Raschka FAQ)
A focused FAQ entry from Sebastian Raschka, author of Build a Large Language Model (From Scratch), separating three terms people blur: unsupervised pretraining, fine-tuning on labeled data, and instruction finetuning that teaches a model to follow requests. Each stage gets a definition, a purpose, and when you would use it, with links to deeper write-ups. The cleanest quick reference on the specialization ladder available. Free to read.
The Illustrated GPT-2 (Jay Alammar)
Jay Alammar's step-by-step visual tour inside a trained transformer language model: the token and position embeddings, the stacked attention blocks, and the weight matrices that compress patterns from training into billions of numbers. Each stage is illustrated with labeled diagrams tracing a sentence all the way to a predicted next token. The reference visualization of what 'the model' physically is. Free to read, with a companion talk.
The Illustrated Transformer (Jay Alammar)
Jay Alammar's Illustrated Transformer is the world's most-cited plain-English explanation of the 2017 architecture behind the ChatGPT era. Step-by-step diagrams show how attention turns input sequences into output — the why-now of modern AI made visual. Written for developers and curious non-specialists; no code is needed to follow it. Free, and used as teaching material in university courses and company onboarding alike.
The Illustrated Word2vec (Jay Alammar)
Jay Alammar's illustrated introduction to word embeddings: how meaning gets plotted as lists of numbers, how relatedness becomes distance between points, and how vectors support search, recommendations, and retrieval. The visual walkthrough of the famous king-minus-man-plus-woman analogy makes the geometry intuitive without any math. Written for a general technical audience and used as course material worldwide. Free to read.
How can I tell how many tokens a string will have? (OpenAI Help Center)
OpenAI's official help article explaining tokens as the unit that models read and bill by: how to count the tokens in a string before sending it, and why embeddings and API usage are priced per token. Short and practical, it turns abstract pricing pages into a concrete check you can run on your own pasted prompts. Written for anyone estimating the cost of a long document or conversation. Free documentation from the maker of GPT.
ICO Guidance on AI and data protection
The UK data-protection regulator's detailed guidance on what actually happens to personal data inside AI systems and what the law requires. You learn the lawfulness, fairness, transparency and accountability principles as they apply across the AI lifecycle, including how bias enters and how to mitigate it. Aimed at compliance officers and developers alike, with plain-language chapters, a glossary and an accompanying risk toolkit. Free, updated as UK AI regulation evolves.
The AI Act Explorer (EU AI Act full text)
The navigable full text of the EU AI Act, the regulation that turned AI literacy into a legal duty. You can read Article 4's AI-literacy obligation and the risk-tier structure that explains why deployment rules differ for chatbots, hiring tools and medical systems. The plainest way to see where regulation is heading globally and what "AI literacy as law" actually says. Free to use, no account needed.
UNESCO Recommendation on the Ethics of Artificial Intelligence
The first global standard on AI ethics, adopted by all 193 UNESCO member states in 2021. It states the core values and principles in plain terms — human dignity, transparency, fairness, sustainability, privacy, human oversight and accountability — plus concrete policy actions to implement them. This is the reference text behind most "AI ethics in plain words" lists, including the human-determination principle. Free and open access for any learner or policy reader.
Energy and AI (International Energy Agency)
The most comprehensive global analysis of AI's electricity demand and its climate implications, built on new modelling and consultation across governments and industry. It supplies the honest numbers the environment conversation needs — data-centre demand more than doubling to roughly 945 TWh by 2030, which sources will serve it, and where AI can instead save energy. A World Energy Outlook special report published under CC BY 4.0. Free to read online with downloadable PDF.
The AI Index Report 2025 — Stanford HAI
Stanford HAI's seventh annual report collating rigorously vetted data on where AI actually stands — capabilities, investment, research, policy and public opinion. It is the calm, evidence-based backbone for "where agents are heading and what the next years plausibly bring", cited by governments and media worldwide. Free PDF with raw data and charts released alongside it, so every claim can be checked.
Weapons of Math Destruction — Cathy O'Neil
The landmark book on how big-data scoring models quietly reinforce inequality in hiring, lending, policing and education. You learn to trace opaque feedback loops and spot when a model's authority is borrowed from math rather than evidence. Written for general readers by a mathematician and algorithmic auditor. National Book Award longlist and Euler Book Prize winner, 288 pages from Crown.
Crash Course Artificial Intelligence (series)
Crash Course's Artificial Intelligence series maps where AI already lives in everyday life: search, recommendations, spam filters, voice input, translation and camera processing. Each episode breaks one familiar application down to its actual mechanism, demystifying the AI in your own apps. Aimed at general audiences and produced with PBS Digital Studios. The full series is free on YouTube.
[1hr Talk] Intro to Large Language Models (Andrej Karpathy)
Andrej Karpathy's one-hour survey talk on how modern large language models are built and trained: pretraining on internet text, supervised instruction tuning, and reinforcement learning from human feedback, all in clear technical language. It covers what parameters and tokens are, why models hallucize, and how fine-tuning specializes a base model — the full two-stage picture. Delivered by a founding member of OpenAI's technical staff and viewed millions of times. Free on YouTube.
Generative AI in a Nutshell (Henrik Kniberg)
Henrik Kniberg's animated eighteen-minute video maps the entire generative AI landscape — transformers, LLMs, training versus prompting, tools and agents — into one coherent mental model. It is explicitly survival-oriented: what changed in 2022 and how to thrive alongside it. Over a million views and widely used in corporate AI upskilling sessions. Free on YouTube with a companion illustrated PDF.
Anatomy of an AI System — Crawford & Joler
A large-scale visual map and 21-part essay tracing one Amazon Echo from mineral extraction through daily use to the e-waste dump. It makes visible the three extractive processes behind a voice assistant — material resources, human labour and data — connecting cobalt mining, clickwork and planetary logistics in a single diagram. The study piece for the sustainability principle in daily choices. Free project by Kate Crawford and Vladan Joler (AI Now Institute and Share Lab, 2018).
Embedding Projector (TensorFlow)
Google's browser-based visualization tool that plots high-dimensional embeddings in 2D and 3D so you can literally see clusters, neighborhoods, and distances between points of meaning. Load the built-in datasets — word embeddings, image features — and explore how relatedness looks spatially, the intuition behind semantic search and RAG. Free to use with no installation, part of the open-source TensorBoard ecosystem. A hands-on way to feel what an embedding is.
ExplAIn — AI Fundamentals (interactive lesson)
ExplAIn's interactive AI Fundamentals lesson rebuilds AI literacy for the post-2022 world: what AI is, seventy years of history, narrow versus general and jagged systems, rules versus learning, and honest limits — with runnable demos like a pathfinding visualizer and a neural network playground. Reading and interacting takes about thirty minutes and ends with a comprehension check. Designed for learners who want the real mental model, not hype. Free, no account.
ExplAIn — Multimodal AI (interactive lesson)
ExplAIn's multimodal lesson is a forty-minute interactive deep dive into models that see, read and listen at once: CLIP, contrastive learning, zero-shot classification and vision-language models, plus audio and video. Interactive demos and a closing quiz make the diagnostic question — is this generative AI or not? — answerable. Aimed at learners who finished an AI primer and want the next layer. Free, no account.
LLM Visualization (Brendan Bycroft)
An interactive 3D walkthrough of a GPT-style language model, from the tokenizer through the attention blocks to the final next-token prediction. You fly through a live rendering of the actual weight tensors and watch numbers flow through the network, which makes 'billion parameters' concrete instead of abstract. Built as a free open-source teaching artifact and widely shared as the clearest visual of what sits inside a model. Ideal before any deeper course on transformers.
Teachable Machine (Google)
Google's Teachable Machine lets you train a working image, sound or pose model in your browser with your own examples — the flashcards demo made real. Watching its confidence rise, misfire and bias as you add or skew training data teaches pattern-learning and data coverage viscerally. Built for students and teachers, with no code and no account needed. Used in thousands of classrooms worldwide and completely free.
Tiktokenizer
An interactive tokenizer playground showing exactly how models chop text into tokens: each token and its ID is highlighted as you type, with live counts for OpenAI's GPT tokenizers. Paste the same sentence in English and Bangla and watch the counts diverge — the clearest demonstration of why non-English text costs more and why models stumble on character-level tasks. Free with no signup, open source. The standard hands-on complement to any explanation of tokens.
Embeddings: Interactive exercises (Google Machine Learning Crash Course)
Hands-on exercises from Google's Machine Learning Crash Course where you train and interrogate embedding representations yourself, testing how models encode meaning as coordinates and how sparse categorical data becomes dense vectors. Each exercise gives instant feedback and links back to the matching lesson on embedding fundamentals. Designed for learners with a little Python, entirely in the browser. Free, from Google's official ML curriculum.
A Review of Bangla Natural Language Processing Tasks and Resources (arXiv)
A survey paper reviewing the state of natural language processing for Bangla: which tasks and datasets exist, where resources are thin compared with English, and how the script's compound characters and inflection complicate tokenization and modeling. It documents precisely why current AI tools vary in Bangla quality — training coverage, not magic. Studied material for understanding small-language realities in AI. Open access on arXiv.
BenLLM-Eval: Large Language Models on Bengali NLP (arXiv)
A research evaluation of how today's large language models perform on Bengali language tasks, measuring where quality holds up and where it degrades relative to English across understanding, generation, and code-switching. The results give concrete evidence for uneven coverage: models are fluent but thinner on low-resource languages and mixed-script input. Read it to calibrate expectations when using AI tools in Bangla. Open access on arXiv.
Datasheets for Datasets (Gebru et al.)
Datasheets for Datasets by Timnit Gebru and colleagues is the studied material behind training-data fingerprints: how datasets are collected, what they cover, and the biases and gaps they carry into models. Its questionnaire has become the standard for documenting provenance, consent and intended use. A readable paper with no code required. Free on arXiv and republished in Communications of the ACM.
Generative AI and Jobs: A global analysis (ILO)
The International Labour Organization's global study scoring every ISCO-08 occupation's tasks for GPT exposure, using microdata from 59 countries. It shows why the honest frame is task-not-job: most exposed work is augmented rather than automated, with effects varying sharply by country income group and hitting women's clerical employment hardest. This is the evidence base for discussing AI and jobs without panic or hype. Free working paper with downloadable PDF.
Green AI (Schwartz, Dodge, Smith, Etzioni)
Green AI is the reference paper on the compute, energy and financial cost of training modern models. It quantifies why training runs are expensive and environmentally significant, and argues for reporting efficiency alongside accuracy. A short, heavily cited read that grounds the environment side of AI literacy. Free on arXiv and published in Communications of the ACM.
RLHF: Reinforcement Learning from Human Feedback (Chip Huyen)
Chip Huyen's widely referenced deep dive into how a raw pretrained model becomes an assistant that follows instructions: supervised fine-tuning on demonstrations, training a reward model on human preference comparisons, and optimizing the policy with reinforcement learning. She walks through the math gently, the practical failure modes, and what 'trained to be helpful' does and does not guarantee. Written by a bestselling ML systems author. Free to read.
The Artificial Intelligence Revolution, Part 1 (Wait But Why)
Wait But Why's The Artificial Intelligence Revolution Part 1 is the classic long-form explainer of narrow AI, AGI and ASI — what each term means and why the timeline debate exists at all. Tim Urban builds the vocabulary from zero with diagrams and a careful account of expert disagreement. Written for intelligent non-specialists. Free, read by millions, and still the most-shared AGI primer on the web.
The Artificial Intelligence Revolution, Part 2 (Wait But Why)
Part 2 of Wait But Why's AI Revolution series works through what superintelligence would actually mean: capability jumps, control problems, and the honest uncertainty around takeoff scenarios. It separates credible concerns from science fiction, exactly the myth-clearing the topic needs. A free companion to Part 1 with the same accessible depth. Frequently assigned on future-of-AI reading lists.
Tracing Knowledge Cutoffs in Large Language Models (arXiv)
A research paper testing how large language models actually behave around their training cutoff, showing accuracy decay on time-sensitive facts and how unevenly that decay falls across topics and domains. It backs the classroom claim that rare facts and niche subjects are thinner in the training data than common ones. Read it as studied evidence for why models confidently fail on recent or obscure questions. Open access on arXiv.
What Is ChatGPT Doing … and Why Does It Work? (Stephen Wolfram)
Stephen Wolfram's celebrated long-form essay explaining, from first principles, what a language model actually does when it writes: probabilities over the next token, trained into billions of weights. It builds the lossy-compression and sampling intuition in plain words with worked diagrams, then connects the mechanism to why answers vary and why fluent text can be confidently wrong. Written for curious non-programmers yet rigorous enough that engineers cite it. Free to read on his site.
Why AI Is Harder Than We Think (Melanie Mitchell)
Melanie Mitchell's paper Why AI Is Harder Than We Think dissects the four fallacies behind confident AI claims, using real failures and careful reasoning. It is the best short study of why benchmark progress misleads and why general intelligence stays hard. Written for a broad audience — a few pages of focused reading. Free on arXiv and widely assigned in AI literacy and ethics courses.
Why language models hallucinate (OpenAI)
OpenAI's research paper and explainer arguing that hallucination is a natural consequence of how models are trained and evaluated: statistical guessing rewarded during training, with confident wrong answers scoring the same as abstentions on many benchmarks. It shows why fluency and confidence are features of generation, not signs of knowledge. The most authoritative plain-language treatment of why models make things up. Open access.
AI Snake Oil (Narayanan & Kapoor)
AI Snake Oil, by Princeton's Arvind Narayanan and Sayash Kapoor, is the leading myth-versus-truth analysis of what AI can and cannot actually do. Their essays dissect overhyped claims across prediction, generative AI and AGI talk, giving readers evidence-based criteria for judging AI products. Written for a general audience and grounded in their Princeton research and book. Free to read, with a companion newsletter and talks.
Hallucination (artificial intelligence) — Wikipedia
The encyclopedia entry cataloging AI hallucination: fabricated facts, citations, and features produced with full fluency, with sourced examples, proposed causes, and mitigation approaches. It covers the taxonomy a user needs to recognize — wrong facts, wrong math, invented sources — and the research literature behind each. Studied material with dense citations for going deeper. Free and editable.
Knowledge cutoff — Wikipedia
The encyclopedia entry defining the knowledge cutoff: the date after which a model's training data ends, leaving it blind to later events, and why different models carry different cutoffs. It collects sourced material on temporal boundaries in training corpora, the effect on question answering, and what users can do about it. A compact studied reference with citations to primary research. Free and continuously edited.
Stanford HAI — Artificial Intelligence Glossary
Stanford HAI's AI glossary defines the vocabulary that makes AI news legible: model, training, prompt, token, agent, hallucination, alignment, benchmark and more. Each entry is a short vetted definition from Stanford's Human-Centered AI institute, written for non-specialists. The right reference to keep open through the AI literacy pillar. Free and continuously updated.
Gender Shades — MIT Media Lab
The intersectional audit of commercial facial-analysis and gender-classification systems by Joy Buolamwini and Timnit Gebru. You study the measured error rates — up to 34.7% on darker-skinned women versus 0.8% on lighter-skinned men — and the skewed benchmark datasets that produced them. It is the classroom case for why "the data said so" can hide real discrimination. Free research project with full results and the Pilot Parliaments Benchmark.
MIT AI Risk Repository
A living database of over 1,700 AI risks extracted from 74 published frameworks, organized by a causal taxonomy and a domain taxonomy with 24 subdomains such as "false or misleading information" and "discrimination & toxicity". You can filter it to see the full landscape of documented harms in high-stakes areas — health, finance, misinformation — with every risk linked to its source paper and quotes. Free to browse, copy and use; maintained by the MIT AI Risk Initiative.
IBM SkillsBuild — Artificial Intelligence Fundamentals (badge)
IBM's free Artificial Intelligence Fundamentals credential is a structured course plus exam covering AI categories and machine learning approaches — supervised, unsupervised, deep learning and reinforcement learning — alongside responsible AI. Short modules end in a verified digital badge you can display on profiles. Designed for students and career starters with no prerequisites. Created by IBM Training and free through IBM SkillsBuild.
