TL;DR
"Intro to AI" is a six-part beginner mini-course from MIT Technology Review, written by Abby Ivory-Ganja (October 15, 2024) from the magazine's AI reporting and reissued as a 12-page subscriber PDF (2026 copyright). It teaches no engineering. Its value to a staff engineer is as a reference for what a careful newsroom thinks a non-technical adult should know about AI: a small vocabulary, a few hands-on exercises, a list of traps, and a framework for regulation.
Verdict: a sound skeleton for an AI literacy curriculum, but dated. It describes the chatbot era (GPT-4, DALL-E 2, Midjourney) and predates agents, reasoning models and tool use, and its policy section describes US rules that have since changed. Its durable lessons are about judgment, not technology: verify outputs, don't anthropomorphize, keep sensitive data out of prompts, and separate replacing a task from replacing a worker.
What the course covers
Six parts, each a short newsletter issue. The arc runs from definitions to use, conversation, risks, work and policy.
| Part | Topic | Core claim | Notable specifics |
|---|---|---|---|
| 1 | What is AI? | AI is an umbrella term whose boundary keeps moving; discount vendor claims. | 70-year history from Dartmouth; dictionary of neural networks, ML, deep learning, generative AI, LLMs, multimodal AI, alignment. Exam scores (bar, USMLE) say little because test items are in training data. |
| 2 | What you can do with AI | Learn by doing a low-stakes task, then iterate on the output. | Eight-step dinner-party walkthrough in ChatGPT (menu, correct it, wine, grocery list, oven schedule, playlist, poem). Other uses: writing, study aid, trip planning, comparison tables. Limits: no AGI, no causal reasoning, no world model outside training data, weak at maths, not a doctor. |
| 3 | How to talk about AI | Be specific about which AI you mean, and know when it is the wrong tool. | Ask a chatbot about raincoats, not about where to vote or how to rank job candidates. For kids: AI is not your friend, recommenders optimize for engagement, keep chat logs and version history in case of false AI-cheating accusations, deepfake legal risk, don't anthropomorphize. |
| 4 | AI traps | Bias and fabricated output are built in, and detection is unreliable. | Biased training data (US-centric images, sexualized women, hiring filters); RLHF as partial mitigation. Detectors (GLTR) miss and falsely accuse; humans improve at spotting machine text with practice (Ippolito); watermarks are voluntary and strippable. Recommends Elements of AI (University of Helsinki). |
| 5 | Working with AI | Replacing a task is not replacing a worker; the outcome is a choice. | Doctor plus AI beats either alone at breast-cancer detection; AI narrows skill gaps but over-reliance misses errors. Goldman Sachs: 300M jobs (about 18%) exposed. Legal case study: GPT-4 invented a citation to argue with the lawyer who tried the case; juniors who skip research lose the practice that builds oversight. |
| 6 | Does AI need tougher rules? | Two models of regulation: govern development or govern use. | EU AI Act (risk tiers, chatbot and deepfake disclosure, labels); US 2023 executive order and AI Safety Institute (standards, weak enforcement); China's per-product rules (recommenders, deepfakes, generative AI); African Union strategy expected in 2025. |
What has aged by November 2026
The text is from late 2024 and was not revised for the 2026 reissue. These points are from general knowledge, not re-checked for this summary.
The tools. Every example is a chat window or an image generator. Nothing covers agents that act (browse, run code, edit files, book things), reasoning models, tool use or long context, which is where most 2026 use and risk sits. "Agents that can manage the entire process of planning and booking your vacation are still some way off" no longer holds as a blanket statement.
The capability limits. "Can't solve college-level math" was already shaky when DeepMind's olympiad systems were cited, and reasoning models have since moved it further. The causal-reasoning and world-model caveats still stand as open research questions, but stated as flat facts they will read as wrong to learners who have watched a model work a proof.
US policy. The 2023 executive order it describes was revoked in January 2025, and the AI Safety Institute was renamed and refocused later that year. The EU AI Act has moved from "poised" to phased enforcement since August 2024.
Privacy. "Anything you put into a prompt becomes part of the training data" is too broad. It is true for some consumer defaults and false under most business and API terms. The safe habit it teaches is right; the mechanism is wrong.
Detection. GLTR (2019) is an odd example in 2026. The broader point, that detectors miss and falsely accuse, has only strengthened.
What holds up
The parts about judgment age well, because they don't depend on any model.
Correction is the skill. The dinner-party walkthrough is really a lesson in iterating: the first menu is all vegan, the first poem is too long, and the user fixes both by saying so. That is the same habit R012's learning study rewards: use AI as a partner you steer, not a vending machine.
Fit-for-purpose test. "Raincoat yes, where to vote no, ranking job candidates no" is a compact rule: low stakes and checkable is fine; civic, legal, medical or hiring decisions need a source of record and a human owner.
Task vs worker. Part 5 separates automating a task from eliminating a job, and notes that over-reliance makes people miss errors. The legal case study names the long-run cost: if juniors stop doing research, who will be able to check the AI later? That is the oversight-pipeline problem in any profession, software included.
Literacy as equity. Regina Barzilay's point, that AI understanding could become a divide between those with CS degrees and everyone else, is the strongest argument for teaching this material at all.
Two axes of regulation. "Regulate development vs regulate use" still sorts every regime, even after the specific rules changed.
Format. Six short parts, each ending in a reading list, with one hands-on exercise early. That is a reusable shape for teaching a team or a family.
Checklist for an AI literacy curriculum
Use this when adopting the course or writing an updated one for a team, a school or family.
Vocabulary goes one layer past the chatbot: model vs harness, tools, agents, context, and what "reasoning" means in practice.
The first exercise is low stakes and requires at least one correction, so learners practise steering.
Learners can state the fit-for-purpose rule and name decisions that need a source of record and a human owner.
Privacy guidance names the actual mechanisms: consumer vs business terms, training opt-outs, and what an agent with file or browser access can see.
Verification is taught as a workflow (check citations, open the source, compare a second source), not only as a warning.
There is a section on agents acting for you: permissions, irreversible actions, and when to require approval.
Learning vs delegating: AI as tutor builds skill, AI as ghost-writer does not (R012's RCT has the numbers).
Policy content carries an as-of date and is reviewed at least yearly.
Every capability limit is dated, so it can be retired when it stops being true.
Related entries and source
Indexed as R014 under 8.1 "AI literacy curricula" (theme 8, AI education).
R012 (AI crash projects and the learning RCT): the evidence base this course lacks. It shows tutor-style use builds lasting skill and ghost-writing erases it, and that AMIE matched physicians only under live oversight, which sharpens Part 2's "don't replace doctors".
R005 (Second Brain OS): another free course built as a learning path, aimed one level up, at people building agents. Together they bracket the curriculum, from user literacy to builder literacy.
R006 (six harness terms): the vocabulary Part 1's dictionary stops short of. It is the natural next lesson after "what is an LLM".
R004 (no-setup Claude, weak link): its Goal + Context prompting is the adult version of Part 2's iterate-on-the-menu exercise.
Source: Abby Ivory-Ganja, "Intro to AI: A mini-course to simplify complex concepts and make AI accessible for everyone", MIT Technology Review, October 15, 2024; subscriber-exclusive 12-page PDF (© 2026). Based on reporting by Will Douglas Heaven, Melissa Heikkilä, David Rotman, Rhiannon Williams, Karen Hao, Tate Ryan-Mosley and Zeyi Yang. The PDF prohibits reposting, so this summary paraphrases and quotes only short phrases.