Home

AI Learning Hub › AI Tutorials › Introduction to AI for Beginners

Foundational skills · Ten-part course

Introduction to AI for Beginners

Start here. Ten parts take you from what generative AI actually is to a working strategy for using it: how it works, what it can and cannot make, how to prompt it, how to prepare what you give it, which tools to avoid, and where the field is heading. Nothing assumes you have used AI before.

Start with part one AI explained using analogies Back to the tutorials →
Format
Ten parts to read, with videos, hands-on activities, and a set of prompts to try
Length
An afternoon end to end; each part stands alone and takes fifteen minutes or so
Written for
The whole SLS community — students, faculty, and staff. No technical background assumed
What you need
A browser, and a free account on one AI tool if you want to do the activities

Written in autumn 2025. The concepts hold; the model names, prices, and features in parts three and ten move every few weeks, so read those as examples of a shape rather than a current inventory.

Two rules before you start. Never put personal, confidential, or proprietary information into an AI tool. And if you are a student, check your instructor’s policy before using AI on coursework — permission is theirs to give, and the Honor Code applies either way.

Part one

What is generative AI?

By the end of this part you should be able to

  • Define generative AI in a sentence, without jargon
  • Say what separates it from the AI that was already in your spam filter
  • Name what it is good at, and where it needs you
  • Explain why the co-pilot — or intern — framing is the useful one

What it actually is

Generative AI is software trained on millions of examples that can now produce new content — text, images, code, music, video — resembling what it learned but not copied from it.

The word doing the work is generative. Ordinary software follows explicit rules. Older AI recognized and sorted things. Generative AI makes something that did not exist before, which is why it is both remarkably creative and reliably wrong some of the time. Those two facts have the same cause.

Traditional AI and generative AI are not the same thing

Traditional AI

Recognizes, analyzes, classifies

It sorts what already exists.

  • Spam filters, sorting mail
  • Recommendations on Netflix or Amazon
  • Image and face recognition
  • Voice assistants answering commands

Generative AI

Creates something new

It produces content from learned patterns.

  • ChatGPT, Claude, Gemini — new text
  • Sora, Veo, and the image tools — pictures and video
  • GitHub Copilot — code
  • Suno, ElevenLabs — music and voice

How it works, in three steps

1. Training

The model is exposed to enormous amounts of material — billions of words, millions of images — and learns the patterns in it: which words tend to follow which, what makes a face a face, which chords sound resolved.

2. Your prompt

You ask for something. The model works out what you are asking for and pulls up the patterns relevant to it.

3. Generation

It produces new content by predicting, over and over, what should come next. It is not retrieving anything it saw. It is assembling something that fits the shape of what it saw.

The comparison that makes this click is a chef who has eaten at ten thousand restaurants and memorized every pattern of flavor, technique, and plating. Hand them your ingredients and they do not look up a recipe — they build something new inside the rules they have internalized. There is a longer version of that idea, and eighteen more like it, in AI explained using analogies.

What it can do

Text

Drafting, answering, summarizing, translating, and creative writing.

Images

Artwork from a description, photo editing, logos, mockups, style variations.

Code

Writing it, debugging it, explaining it, documenting it, translating between languages.

Audio and video

Music, voiceover, sound effects, editing, and generated video.

What that looks like in practice

It is a tool, and you are the editor

AI is not thinking or understanding. It recognizes patterns and applies them statistically to produce output that looks intelligent. It feels like magic because the mechanism is invisible — and like a magic trick, it stops feeling magical the moment you see behind the curtain.

It will be confidently wrong. AI produces incorrect information, reproduces bias, and misreads intent. It will invent facts that sound true — the term for that is hallucination. This is not a bug being fixed next quarter; it follows from how the thing works. You are the quality control.

Treat it like a brilliant, fast, overconfident intern. It will hand you a first draft in seconds, and it may have invented half of it. You are the editor-in-chief.

Which makes your job concrete: set clear expectations, give real instructions, read the output critically, check the claims, add the context and expertise it does not have, and make the decision yourself.

Your turn: learn with AI

20 minutes

Both major tools now have a mode built for learning rather than answering. Try the same question in each and compare what you get.

  1. ChatGPT’s Study Mode. Sign in at ChatGPT (opens in a new tab), open the dropdown in the message box, and select Study and Learn. There is an explanation of the mode (opens in a new tab) if you want the detail.
  2. Gemini’s Guided Learning. In Gemini (opens in a new tab), click the plus symbol or Tools, then Guided Learning. Google has written it up here (opens in a new tab).
  3. Paste one of the prompts below into both, and read the two answers side by side.

Prompts to try

Explain the concept of 'AI hallucinations' like I'm a 5th grader. Provide a real-world example of how this could be a problem in a legal context.

Create a simple timeline of the 5 most important milestones in the history of artificial intelligence, starting from the 1950s.

What are the key ethical considerations when using generative AI to create art or music? Present the arguments from two different perspectives.

Describe three ways generative AI is currently being used in the legal profession. For each, list one major benefit and one potential risk.

Act as a tutor and teach me the difference between a Large Language Model (LLM) and a neural network. Use an analogy to help me understand.

Then reflect. Did the two take different approaches? Was one clearer? Which did you prefer, and can you say why? Noticing the difference is the skill.

A habit worth starting now: keep an AI journal

Log your AI interactions from the beginning. A spreadsheet with five columns is enough: date, tool, prompt, a summary of the output, and your notes. It pays off three ways:

  • Provenance and citation. You can always say where something came from, which academic integrity and professional practice both eventually require.
  • Skill. You start to see which prompts and which tools actually work for which tasks, instead of guessing each time.
  • Reflection. You end up with a record of how these tools are changing your research, writing, and thinking — which is worth knowing.

Applied AI, in one example

The video below was made entirely with NotebookLM (opens in a new tab), a Google tool that grounds Gemini in only the sources you give it and turns them into overviews, mind maps, quizzes, and flashcards. It is one of the most useful things you can add to your toolkit, and you will see more examples like this as the course goes on. The rule from the top of the page still applies: nothing sensitive, confidential, or proprietary goes into it.

The library is part of this. Drop in at the Curiosity Corner to ask anything AI in person, read The AI Upload for what changed this week, or email library@law.stanford.edu for a one-to-one session.

Key takeaways

  • Generative AI creates new content from patterns learned in enormous datasets
  • Traditional AI classifies what exists; generative AI produces what does not
  • It is not magic, not sentient, and not reliable — it is sophisticated pattern matching
  • Your role is expert guide, editor, and quality control

Part two

A brief history of AI

By the end of this part you should be able to

  • Say roughly when the field started, and how old the idea is
  • Name the milestones that mattered, and one that mattered most
  • Explain what an “AI winter” was, and why progress has not been a straight line
  • Describe what the Transformer changed, and what may replace it

Artificial beings with minds are an ancient idea; the scientific pursuit of them is about seventy-five years old. The route from there to here runs through enormous ambition, two long stalls, and a handful of breakthroughs that arrived faster than anyone expected.

Why bother with the history? Because it calibrates you. It shows what these systems were built to do, it makes the current moment legible rather than magical, and it is the cure for assuming the last three years predict the next three. Progress here has never been inevitable or linear — it has needed sustained funding, a conceptual breakthrough, and hardware that could keep up, all at once.

Before computers

The modern timeline

1950 · The Turing test

Alan Turing publishes “Computing Machinery and Intelligence,” proposing a test for whether a machine’s behavior is distinguishable from a human’s, and asking the question the field has never quite escaped: can machines think?

1956 · The field gets a name

The Dartmouth Conference, organized by John McCarthy and others, coins the term artificial intelligence and establishes it as an academic discipline.

1970s–1980s · The first AI winter

Early optimism collides with slow computers and hard problems. Critical reports land, funding dries up, and the field spends a decade out of favor.

1997 · Deep Blue beats Kasparov

IBM’s Deep Blue defeats the world chess champion — the first widely felt demonstration that a machine could outplay the best human at something we considered pure thinking.

2012 · Deep learning arrives

AlexNet wins the ImageNet competition by a wide margin using deep learning, proving that with enough data and enough compute, neural networks work.

2017 · The Transformer

“Attention Is All You Need” introduces the Transformer architecture, which becomes the foundation of GPT, BERT, and effectively every generative model you have used. Read the paper (opens in a new tab).

2022 · It reaches everyone

DALL·E 2 puts image generation in front of the public, and ChatGPT launches in November and reaches a hundred million users in two months — the fastest consumer adoption on record at the time (opens in a new tab).

The one to understand: attention

If you only remember one moment, remember 2017. Before it, models read text in sequence, one word after another, like a person reading aloud — which made it hard to hold the beginning of a long document in view, and slow to train.

The Transformer let a model take in the whole input at once and, for any given word, weigh how much every other word matters to it. That is the attention mechanism, and it bought two things: a much better grip on context regardless of distance, and training that was far more efficient. The efficiency is what made scale possible — billions and then trillions of parameters — and scale is what produced the models we now argue about.

What comes after the Transformer

The Transformer is not the end of the story, and its main weakness is cost: attention gets quadratically more expensive as text gets longer, so analyzing a book or a large case file is slow and expensive by construction. Two directions are worth knowing the names of.

State space models — the efficient reader

Imagine a speed-reader who scans each line and keeps a running summary in their head, updating it as they go, instead of re-reading the page to understand one sentence. SSMs process sequentially and stay fast over very long inputs. Mamba (opens in a new tab) is the example to look up.

Hybrids — both engines

Like a hybrid car using the electric motor around town and the engine on the highway, these combine attention’s contextual power with the efficiency of older recurrent designs. RetNet (opens in a new tab) from Microsoft is one line of that research.

The takeaway is not which one wins. It is that the field is in a permanent race for cheaper and faster, and the next jump in what AI can do may come from architecture rather than from size.

See it for yourself

Your turn: be the historian

45 minutes

This one teaches two things at once: what the raw data of AI progress actually looks like, and how to get a research tool to work from sources you chose rather than the whole internet.

  1. Pick a dataset. Go to Epoch AI’s data page (opens in a new tab), choose something that interests you, and download it.
  2. Prepare the files. Unzip it, open each .csv in a spreadsheet program, and export each as a PDF. This step is the unglamorous heart of data work.
  3. Load it. In NotebookLM (opens in a new tab), create a notebook and upload your PDFs as the sources.
  4. Interrogate it. Ask questions of the data, then use the studio tools to generate an audio overview, a video overview, and a mind map.
  5. Record what happened. In your AI journal: what it got right, what it flattened, and what you would do differently.

The video below is that exercise done with Epoch’s AI Companies data, start to finish.

And cite it. Epoch’s data is free to use and redistribute under a Creative Commons Attribution license (opens in a new tab) provided you credit the source and authors — which is the last step of any research, not an optional courtesy.

Sources for this part
  • Turing, A. M. (1950). “Computing Machinery and Intelligence.” Mind, 59(236), 433–460.
  • McCarthy, J., Minsky, M. L., Rochester, N., & Shannon, C. E. (1955). “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence.”
  • Lighthill, J. (1973). “Artificial Intelligence: A General Survey.” Science Research Council.
  • Campbell, M., Hoane, A. J., & Hsu, F. H. (2002). “Deep Blue.” Artificial Intelligence, 134(1–2), 57–83.
  • Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). “ImageNet classification with deep convolutional neural networks.” NeurIPS 25.
  • Vaswani, A., et al. (2017). “Attention is all you need.” NeurIPS 30.
  • Ramesh, A., et al. (2022). “Hierarchical text-conditional image generation with CLIP latents.” arXiv.
  • Shelley, M. (1818). Frankenstein; Čapek, K. (1920). R.U.R.

Key takeaways

  • The idea is ancient; the field began in the 1950s
  • Progress has come in bursts separated by AI winters
  • The 2017 Transformer paper is the hinge everything modern turns on
  • More efficient architectures are already being built to succeed it

Part three

How it works

By the end of this part you should be able to

  • Say what a large language model is, and what the “large” refers to
  • Explain next-token prediction to somebody else
  • Describe the two phases of training, and what each one produces
  • Define tokens and context windows, and notice when you have run out of the latter
  • Say why the same design that makes it fluent makes it hallucinate

The engine: large language models

A large language model is something like an enormously detailed map of human language, built by observing an immense amount of writing. Ask it a question and it behaves like a navigator across that map, predicting the most probable path of words that answers you. It is a map, not a traveler: it has never been anywhere. (The longer version of that analogy is on the analogies page.)

“Large” means three things at once:

One label, many engines

“LLM” is a category, not a product. Companies build and tune models for different goals, the way carmakers build engines for speed, towing, or economy. As of late 2025 the families you will meet are the GPT models from OpenAI, Claude from Anthropic, Gemini from Google, Llama from Meta, and Mistral’s models — and the specific version numbers will have moved by the time you read this.

What they are known for differs, and it is worth choosing deliberately:

A model that is excellent at creative writing is not therefore good at reading a contract. The analogy for this — a chef who has never tasted food but has documented every dinner party ever held — is on the analogies page.

How they learn

Phase one: pre-training

The model reads an enormous amount of text and learns to predict the next word, or token, in a sequence. Given “The sky is,” it learns that “blue” is highly probable. Repeat across billions of examples and it has absorbed grammar, facts, styles of reasoning — and whatever biases were in the material.

Source: Devlin, J., et al. (2019). “BERT.” NAACL-HLT.

Phase two: fine-tuning

Then it is shaped for behavior rather than knowledge:

  • Supervised fine-tuning — training on high-quality examples of what a good answer looks like.
  • Reinforcement learning from human feedback — human raters score outputs, teaching the model which responses people prefer.
  • Instruction tuning — teaching it to follow instructions and hold a conversation at all.

Source: Ouyang, L., et al. (2022). “Training language models to follow instructions with human feedback.” NeurIPS.

The core mechanic: predicting the next token

The whole of generation is this, repeated: given everything so far, what is the most likely next token? Take the input “The capital of France is.” The model has seen that phrase in countless documents, computes that “Paris” is overwhelmingly the most probable continuation, emits it, and then asks the same question again with “Paris” now part of the input. It does not know that Paris is the capital. It calculates that the token is likely.

It is the world’s most sophisticated autocomplete — and that is not a dismissal, it is the mechanism.

One wrinkle: models do not always take the most probable token, because that produces flat, predictable text. A setting called temperature controls how much randomness is allowed. Low temperature is more predictable and more factual; high temperature is more varied and more creative, and drifts further from the safest answer.

Source: Holtzman, A., et al. (2020). “The curious case of neural text degeneration.” ICLR.

Tokens

A token is a chunk of text: a word, part of a word, or a punctuation mark. Models break everything you write into tokens before processing it, which is why they are also the unit of pricing and of limits.

The video below shows tokenization visually, and the song analogy covers the idea in prose.

Context windows, and the limit you cannot see

Every model has a context window: the maximum number of tokens it can hold at once, counting your prompt, everything you have attached, and its own replies. The window is where the conversation lives, and when material falls out of it, it is gone. The moving train window is the analogy for this.

Windows have grown fast:

Nothing tells you when you run out. Chat interfaces do not warn you, and there is no counter. You find out indirectly: the model starts contradicting itself, forgets a detail you established twenty messages ago, or answers as though it has never seen the document. If a long conversation suddenly gets worse, suspect the window before you suspect the model.

Two techniques for long work.

  • Periodic summaries. Every so often, ask: “Summarize the key decisions and important points from our conversation so far.” That pulls the essentials back to the front of the window.
  • Chain-summarize into a new chat. When a conversation is done but the project is not, ask for a comprehensive summary, open a fresh chat, and paste it as the first message. You get a clean window with the history intact.

Your turn: watch text become tokens

10 minutes

There is no universal standard for tokenization — each company trains its own. The same sentence might be ten tokens for one model, twelve for another, and nine for a third, which is a concrete example of these engines differing under the hood.

Paste the same sentence into each of these and compare. Which model is most efficient with your writing?

Try a sentence with unusual words, a citation, and a long compound term. The differences are most visible where language is least ordinary.

Under the hood, briefly

Models are built on neural networks: layers of artificial neurons, each receiving numbers, transforming them, and passing them on. Training adjusts the connections until the network recognizes patterns. If it helps, picture a spreadsheet with billions of cells, each holding a number that influences how text flows through the system.

Modern LLMs specifically use the Transformer, whose attention mechanism lets the model weigh how much every word matters to every other word. In “The animal didn’t cross the street because it was too tired,” attention is what lets the model tie “it” to the animal rather than the street.

Does it think?

No. There are no beliefs, no consciousness, no understanding. Even when a model appears to reason — and newer ones will show you a chain of steps — it is producing a sequence of tokens that resembles reasoning it has seen. Pattern matching at extraordinary scale is not a mind.

When it writes “I understand” or “I think,” those are tokens it learned belong in conversation. Nothing is being understood or thought. The analogy for this — someone who has studied a million photographs of a city and never walked its streets — is on the analogies page.

What follows from that, practically:

Source: Bender, E. M., & Koller, A. (2020). “Climbing towards NLU.” ACL.

Why it makes things up

A hallucination is not a malfunction. Asked for a fact it does not have, the model still does the only thing it does: produce the most probable continuation. What comes out has the shape of a fact, or a citation, without the substance of one. The remembered birthday party is the analogy.

Rates vary enormously by task and model. Top models hallucinate rarely on constrained work like summarizing a document you supplied; loosely grounded questions are far worse, and studies through 2024 found chatbots getting scientific references wrong in a large share of cases. One widely-reported example had a model recommending glue in pizza sauce, a “fact” it had absorbed from a sarcastic forum post. The harder problem is that newer models are more likely to produce a confident, detailed wrong answer than to say they do not know.

Researchers do not expect hallucination to be eliminated, because it is a consequence of the architecture rather than a defect in it. The goal is to make it rarer and easier to catch — which is your job as much as theirs.

The other three limits worth naming:

What to do about it. Verify anything that matters, and especially every citation, legal reference, and number. Use AI as the strong start of a piece of work rather than the end of it. And when you are unsure, ask the same question again in a fresh conversation — inconsistent answers are a reliable tell.

Source: Jones, N. (2025). “AI hallucinations can’t be stopped — but these techniques can limit their damage.” Nature, 637(8047), 778–780.

Free training straight from the model builders

The companies that build these tools publish good free material, and it is worth reading them on their own products. OpenAI Academy (opens in a new tab) is the most developed. Two sessions there follow directly from this part — the recordings move around their catalog, so the guides are linked here and the videos are findable by title:

Key takeaways

  • An LLM is a pattern engine trained on a vast amount of text
  • Training has two phases: pre-training for patterns, fine-tuning for behavior
  • Generation is next-token prediction, one token at a time
  • The Transformer’s attention mechanism is what made this possible
  • It does not understand anything, and hallucination follows from the design

Part four

What it can create

By the end of this part you should be able to

  • Name the main categories of output and a real use for each
  • Say what AI does well and what still needs your expertise
  • Explain why factual research is the hardest of these, and what grounding fixes
  • Set expectations that survive contact with an actual task

AI is good at well-defined, pattern-shaped tasks with plenty of examples behind them. It is bad at anything needing real-world understanding, long chains of reasoning, or expertise its training did not contain. Almost every disappointment traces back to that line.

Text

The most mature use, and the one to start with. A professional drafting a client email about something complicated gets structure and clear language from the AI, then supplies the specifics, the relationship, and the accuracy.

What it does well: drafting correspondence, summarizing long documents, marketing and social copy, lesson plans and teaching material, creative writing, translation, reformatting between structures, and brainstorming outlines.

The ethics of AI-assisted writing

Students: permission first

Ask your instructor before using AI on any coursework. Even with permission, the accuracy of what you submit is yours.

Authors and creators

Be transparent about where AI helped. Many publications now require disclosure, and the norm is moving that way fast.

Legal professionals

Never enter confidential client data. Verify every AI-generated authority — it invents cases. The work product is your responsibility.

Business professionals

Good for outlines and polishing email. Not a place for sensitive business or customer data, and not a substitute for strategy.

Citing AI: Bluebook Rule 18.3

The Bluebook’s 22nd edition (2024) added Rule 18.3 for citing generative AI. It specifies what a citation to AI output must contain, but leaves genuine ambiguity about when you cite — whether the AI is a source of authority or merely a tool you used. Until that settles, save your outputs so you can produce them.

The shape of the citation

OpenAI, ChatGPT-4, "Exact prompt text" (Date) (conversational artifact on file with author).

Worth reading on this:

A word about AI detectors

AI detection tools do not work reliably, and the research is consistent about it. They produce false positives — flagging human writing as machine-written — and simple paraphrasing defeats them. Stanford provides iThenticate (opens in a new tab) for plagiarism review, which is a different problem from detecting AI authorship. Treat a detector score as evidence of nothing on its own.

Images and video

From a text description you can generate original images, edit photographs, and produce video; AI also helps across the production workflow rather than only at the point of generation.

What it does well: concept art and illustration, marketing and social visuals, variations on an existing image, background removal and extension, scripts and storyboards, short generated clips, and simple animation.

Sources: Ramesh, A., et al. (2022), “Hierarchical text-conditional image generation with CLIP latents”; Brooks, T., et al. (2024), “Video generation models as world simulators.”

Code

This has become the most transformative application, and not only for developers. A teacher who has never written a line of code can say: “I want a quiz app with multiple-choice questions, a score at the end, and confetti when someone gets everything right” — and after a few rounds of conversation, have something their class can use.

What it does well: boilerplate, explaining code you inherited, finding bugs, translating between languages, regular expressions, unit tests, documentation, and optimization suggestions.

Jargon buster: “vibe coding.” Describing what you want in plain language and letting the AI handle the syntax. You are the architect; it is the construction crew. It is genuinely how a lot of software now gets started — GitHub has reported that Copilot writes a large share of the code in files where it is enabled — and it is also how you end up with something you cannot debug. Both are true.

Tools people actually use:

If you want to learn the underlying craft: Harvard’s CS50 (opens in a new tab) is the classic introduction and now integrates AI tools, and freeCodeCamp (opens in a new tab) is a complete free curriculum with a community attached. The hub’s own Understanding APIs course and GitHub portfolio guide are the next step from here.

Facts and research — the hard one

AI produces text that looks like facts. That is not the same as producing facts, and this is where it fails most expensively. Four distinct problems stack up:

The safer pattern is to bring your own vetted facts and let AI do the shaping. If you must research with it, use tools built for the job — a dedicated deep-research mode, or an answer engine like Perplexity (opens in a new tab) — and verify every claim and citation anyway.

Grounded AI, or RAG. Retrieval-augmented generation is the technique that makes AI usable in high-stakes fields: instead of answering from everything it absorbed, the model is restricted to a specific trusted set of documents, and cites from those. Think of it as locking the AI in a library containing only the books you selected. This is what sits behind the AI assistants in Westlaw, Lexis, and Bloomberg Law (opens in a new tab), and it is the reason those answers can be checked when a general chatbot’s cannot.

Data and spreadsheets

AI is a capable data assistant and a poor data analyst. It will clean and format messy data, categorize text entries, pull information out of unstructured sources, summarize a dataset, suggest a chart type, and write the formula or script you cannot remember. Interpreting what any of it means is still yours.

Your turn: a creative break

15 minutes

Two tools, no experience needed, and a reminder that this technology is also fun.

Make a song. At Suno (opens in a new tab), try: a blues song about a lawyer who lost a brief; a synth-pop anthem for finals; a folk song about the founding of Stanford.

Make a story. In Gemini’s Storybook (opens in a new tab), try: a law book that comes to life at night; an adventure whose hero is a gavel; a mystery set in the tunnels under campus.

Share what you make in #techchat (opens in a new tab) — then come back, there are six parts left.

Fantasy versus reality

Most frustration comes from asking for one enormous thing instead of several specific ones. Two examples of the pattern:

Formatting

“Format this 300-page report perfectly”

What happens: it loses the thread, times out, and starts creatively rewriting your section titles.

What works: ten to fifteen pages at a time, with explicit formatting rules, processed in chunks you check.

Analysis

“Analyze this spreadsheet and give me insights”

What happens: attractive charts that may mean nothing, and confident conclusions that miss the context of your data.

What works: have it clean the data, suggest chart types, and surface basic patterns — then interpret them yourself.

Key takeaways

  • One engine, many outputs: text, code, images, audio, and data work
  • It is strongest on first drafts, repetitive work, pattern-finding, and simplification
  • Strategy, fact-checking, nuance, and judgment stay with you
  • Small specific tasks beat large vague ones, every time

Part five

Responsible AI use

By the end of this part you should be able to

  • Hold a realistic view of AI as a tool rather than a replacement
  • Explain what an “AI stack” is and why one tool is not a strategy
  • Walk through the five steps of the PAUSE framework
  • Apply the two tests — the front page, and the public billboard

The new power tool in the workshop

Imagine a skilled carpenter whose workshop acquires a fast, versatile new power tool. It cuts, sands, and shapes at astonishing speed. It does not replace the carpenter’s judgment about which wood to use, their knowledge of joinery, or their hands on the final pass. Using it to hammer a nail would be foolish — there is a hammer for that — and letting it design the piece would produce something generic.

AI is that power tool. It is here to augment your skills, not to replace your thinking.

Build a stack, not a habit

You would not use one tool for every job, and you should not use one AI for every task. The goal is a small, deliberate set chosen for specific strengths — which also stops you developing tunnel vision about whichever tool you happened to start with. You do not need to subscribe to everything that launches; adopt what genuinely saves you time and meets professional standards, and let other people evaluate the rest.

A working stack

Chosen for what each is good at

  • Claude for summarizing long depositions and drafting a first memo
  • Lexis or Westlaw AI for grounded, citable research
  • Perplexity for quick, sourced background questions
  • A specialized contract analysis tool for contract analysis

A misused stack

The same tools, wrongly assigned

  • A free general chatbot for all legal research — hallucinated citations waiting to happen
  • A music generator for client-related content — unprofessional and beside the point
  • An image generator for courtroom sketches — inappropriate, and not authentic

Your AI Stack on this site is a searchable directory of the tools to choose from, and it lets you save the few that fit your own work. And the honest answer to “should I use AI for this?” is often no — recognizing that is the valuable skill, not the tool list.

Further reading: deep and narrow beats scattershot

A 2025 Harvard Business Review article (opens in a new tab) argues against running many small AI pilots and in favor of pointing real resources at transforming one core part of a workflow. It is the organizational version of the same advice: a purposeful stack beats chasing every new tool.

The PAUSE framework, in short

PAUSE is the workflow the library teaches for deciding when and how to use AI. It moves you from passively accepting output to actively directing the work. The five steps:

P — Precisely define your task

Break the goal into concrete tasks. What has to exist at the end? Who is it for? What are the constraints? Task first, tool second — if you picked the tool before defining the task, you will struggle with the tool.

A — Is it a good AI fit?

Drafting, brainstorming, and outlining: good fit. Anything fact-based: poor fit, because it hallucinates. And sometimes the right tool is a calculator, a template, or a colleague.

U — Use it ethically and securely

Two tests. The front page test: would you be comfortable if your use of AI here were reported on the front page? The public billboard rule: never paste anything into a public AI that you would not put on a billboard — which rules out confidential and personally identifiable information entirely.

S — Do you have the skills to assess it?

You must be able to tell whether the output is any good. If you cannot evaluate it, you cannot responsibly use AI for it. Back to the intern: brilliant, fast, overconfident, and in need of an editor who knows the subject.

E — Does it genuinely enhance your workflow?

Two checks. Time: if verifying and fixing takes longer than doing it yourself, do it yourself. Growth: is this building your skill or eroding it? Use AI as a tutor and a sparring partner, not only as a shortcut.

Just “do it”

“Write a Python script that scrapes a website.”

“Teach me how”

“Act as a tutor. Explain the steps to write a Python web scraper using BeautifulSoup. Give me commented code for each step, and tell me what could break.”

This is the short version. The full framework — with worked examples, checklists, and a one-page decision flow — is the PAUSE Rule on this site. Read it before your next real AI task.

Your turn: research with answer engines

30 minutes

An answer engine is a different class of tool from a chatbot: rather than a list of links, it synthesizes an answer from several sources and cites them inline. That makes it checkable, which makes it useful — and makes checking it the whole exercise.

Use both: Perplexity (opens in a new tab) and Google’s AI Mode (opens in a new tab).

Pick one question

What are the best practical tips for using AI ethically in a professional setting?

How can I identify and mitigate bias when using generative AI tools?

What are the most effective ways to use AI in an environmentally responsible way?

Then verify everything. Open every source each tool cites. Does the source actually support the claim made? Is the source itself credible? Note the verified facts in your AI journal — and note anything the tool asserted that its own citation did not support. That happens more than you would expect, and catching it is the skill this exercise builds.

Key takeaways

  • PAUSE: precise task, AI fit, use ethically, skills to assess, enhance workflow
  • Define the task before choosing the tool
  • You are the human in the loop, and the decision is yours
  • Treat it as a brilliant but fallible intern that needs an editor

Part six

Mastering the prompt

Imagine hiring an extraordinarily capable assistant: they read thousands of pages in seconds, write in any register you need, and spot patterns instantly. The catch is that they are completely literal. Ask them to “help with the case” and they will wait for instructions. Give them specific, structured direction and the work comes back transformed.

That is the whole of it. The difference between people who get remarkable results from AI and people who find it useless is almost never the model — it is how they communicate with it. Prompt engineering is not arcane knowledge; it is clear thinking, written down.

By the end of this part you should be able to

  • Build a prompt with the C.R.A.F.T. formula
  • Explain context engineering, and why it matters more than wording
  • Recognize the five pitfalls in your own prompts
  • Manage what an AI can see, deliberately

From vague to specific

Telling a five-year-old to “clean your room” is a gamble; they are capable, but the instruction is not. Toys go under the bed. “Clothes in the hamper, books on the shelf, toys in the bin” produces a clean room. Prompting is the same trade: enormous capability, no ability to infer what you meant. The full analogy is on the analogies page.

C.R.A.F.T.

A good prompt is a creative brief. Build it from five elements:

C — Context

Background it cannot guess. Who is involved, why this matters, what has already happened.

R — Role

Who it should be. “Act as an experienced appellate clerk.” “You are a patient tutor.”

A — Action

The task, as a verb. Summarize. Draft. Compare. Prioritize. Translate. One of them, not five.

F — Format

The shape of the output. A bulleted list. A professional email. A table with three named columns.

T — Tone

The register. Formal and academic. Friendly and brief. Firm, but not aggressive.

The library’s prompting coaches

Two custom GPTs the library built for exactly this, and which the exercise below uses:

The guardrails: what it cannot do

Most disappointment with AI is a gap between expectation and mechanism. People expect the model to infer intent, read between lines, and know unstated context; when the output is mediocre they blame the technology, though they prompted for a capability it does not have. Six limits, and what each one means for how you write the prompt:

Knowledge cutoff

It does not know events, cases, or regulations after its training data ends. So: state time ranges explicitly, and verify anything recent yourself.

No live data

Without search, it cannot look anything up, check a database, or read a paywalled source. So: do not ask what a court ruled yesterday — give it yesterday’s ruling.

Hallucination

It will produce fabricated citations and statutes with total confidence. So: request sources, and say “only cite cases you can quote directly.”

Context limits

It sees a bounded amount of text and forgets the rest. So: break documents into sections and start new conversations for new topics.

No real understanding

No common sense, no intuition, no experience of the world. So: spell out what would be obvious to a colleague.

Inconsistency

The same prompt can give different results on different days; it is not deterministic software. So: use structured formats and test a prompt more than once.

How to study the limits yourself, instead of taking anyone’s word
  • Run controlled experiments. Ask the same question several ways. Ask about a case you know does not exist. Watch where it succeeds and where it fabricates.
  • Read the documentation. Companies publish model cards describing capabilities, limits, and known failure modes.
  • Follow the community. Researchers and practitioners find new limitations constantly, in public.
  • Stress-test your own use case. Before relying on AI for something that matters, run it on examples where you already know the right answer.

And keep a failure log. Every time you get an incorrect or problematic response, note what you asked and what went wrong. The patterns that emerge will shape your prompting more than any guide.

Five pitfalls, from real legal prompts

Learning what makes a bad prompt is as useful as learning what makes a good one.

Pitfall 1 — the vague request: “Research employment law for me.”
  • No jurisdiction. Federal? Which state?
  • No issue. Discrimination? Wrongful termination?
  • No context about the actual problem
  • No output format

Result: a generic overview that costs you ten minutes and tells you nothing you can act on.

Pitfall 2 — assumed omniscience: “What was the holding in the Johnson case about the non-compete?”
  • There are thousands of Johnson cases
  • No citation, no jurisdiction
  • No timeframe

Result: a confident, plausible, fabricated answer — or a real holding from the wrong case, which is worse.

Pitfall 3 — the kitchen sink: “Draft a motion to dismiss, research the case law, prepare a brief, outline oral argument, and write a client memo.”
  • Five tasks in one prompt dilutes all five
  • Each needs different context and a different format
  • You cannot give useful feedback on any of it

Result: shallow work on everything. Break it into six or seven focused prompts.

Pitfall 4 — implicit expertise: “Review this contract and tell me if it’s good.”
  • Good for whom — the buyer or the seller?
  • No concerns identified: liability, payment terms, term length?
  • No jurisdiction, so no applicable standard

Better: “Review this employment contract from the employer’s perspective in California, focusing on the enforceability of the IP assignment and non-solicitation clauses.”

Pitfall 5 — the unconstrained creative request: “Write a demand letter for my client.”
  • No facts and no claim
  • No tone — aggressive or conciliatory?
  • No desired outcome — settlement or litigation?

Result: a template full of placeholders that needs rewriting from scratch. It needs facts, strategy, and constraints to produce anything usable.

Context engineering

A prompt is your instruction. Context is everything the model can see: the prompt, the files you attached, and the conversation so far. Managing that is more consequential than phrasing the prompt perfectly.

Think of packing a carry-on. You cannot bring the closet. You choose the clothes that match the itinerary, roll them tight, and buy toiletries when you land — the right information, structured clearly, supplied when it is needed.

1. Load only what you need

Do not upload a hundred-page report when the conclusion on page 98 is what matters. Give it the relevant pages.

2. Put important information first

Models attend most to the beginning and end of a long prompt. State the goal and the critical constraints up front.

3. Use structured templates

Instead of “summarize this,” ask it to fill fields: a one-sentence summary, key points, open questions. Consistent input, consistent output.

4. One goal per conversation

Analyzing a report and drafting an email about it are two chats. Start the second with a short summary from the first.

5. Always request sources

Ask for links, page numbers, or direct quotes for every factual claim, so that verifying is possible at all.

For precision: prompting in JSON

You can structure a prompt as a JSON object instead of prose. Rather than one crowded drawer, you get labeled folders. Models were trained on enormous amounts of code, so they read this structure natively and answer it more consistently.

a prompt as an object
{
  "research_topic": "renewable energy trends 2024-2025",
  "focus_areas": ["solar", "wind", "battery storage"],
  "output_structure": {
    "key_trends": "5 bullet points with data",
    "sources": "minimum 3 credible sources with URLs"
  }
}

Your turn: fix a bad prompt

20 minutes
  1. Choose one of the five bad prompts above — “Research employment law for me” is a good start.
  2. Open Prompt Architect (opens in a new tab).
  3. Paste it in and ask the coach to help you rebuild it. It will interrogate you for context, role, format, and tone — which is the part you skipped.
  4. Run the before and after versions in the same tool and compare what comes back.

Twenty minutes of this teaches more than an hour of reading about prompting, because the difference is visible in the output rather than described.

Prompting guides from the model builders, and two for legal work

Key takeaways

  • Prompting is clear communication, not magic words — use C.R.A.F.T.
  • What you put in the context window matters more than how you phrase it
  • Know the limits, and prompt for what the tool can actually do
  • Be specific about jurisdiction and constraints, and do one task at a time
  • Give it only what it needs, most important first, and start fresh for a new task

Part seven

Anatomy of an AI interface

By the end of this part you should be able to

  • Name the core components of a modern AI interface
  • Tell chat, canvas, and agent modes apart, and know when to use which
  • Say what web search, deep research, and artifacts are for
  • Explain how custom GPTs and Gems extend a tool
  • Find the two settings that matter most for privacy

Beyond the chat box

The chat box is the front door, but these platforms have become workbenches. Knowing which tool is which is the difference between using AI for questions and using it for work.

Mode

Chat

Turn by turn. Best for quick questions, brainstorming, and refining an idea by argument. Like texting a well-read colleague.

Mode

Canvas or project

A persistent workspace for a document or a codebase, where you edit the output directly and prompt for changes in place. Like a shared document whose co-author is the AI.

The features worth knowing by name

Building your own assistants

Custom GPTs, in ChatGPT

A version of ChatGPT with your instructions and your reference files, and optionally web browsing or data analysis switched on. Good for a task you do the same way every week.

Custom Gems, in Gemini

The same idea in Google’s tool: give a Gem a role and standing instructions — “you are a legal research assistant” — and stop re-explaining yourself.

The library publishes a set of ready-made skill files that do this without any building: the AI Skills, with installation instructions for ChatGPT and Claude. If you would rather build one from scratch, the Writing Partner agent guide walks through a real example end to end.

Somewhere safe to experiment

Before you take a new capability to real work, try it in a controlled environment. For the Stanford community that is the Stanford AI Playground: university-hosted, free, and with several vendors’ models behind one interface, so your experiments do not go to a consumer account. The hub has a full guide to it.

For advanced users, the vendors’ own playgrounds offer earlier access and more control, usually with an API key or a payment method attached: Google AI Studio (opens in a new tab), the OpenAI Playground (opens in a new tab), Hugging Face (opens in a new tab), and Microsoft Copilot Studio (opens in a new tab).

Reading the fine print

An AI provider’s terms of service and privacy policy are your contract with them. They determine how your data is used, who owns what you produce, and what rights you keep. Skipping them is signing something unread — and the answers differ meaningfully between providers and between the free and paid tiers of the same provider.

Use AI to read them. Download the terms and the privacy policy, upload both into NotebookLM (opens in a new tab) as sources, and interrogate them in plain language. Because the answers are grounded in the documents you supplied and cite their sections, this is one of the few research tasks where AI is straightforwardly reliable.

A prompt for the terms of service

Act as a data privacy and compliance analyst. Based only on the provided source documents (the Terms of Service and Privacy Policy), answer the following. For each answer, cite the specific section or paragraph.

1. Data usage for training: is my content used to train the models by default? If so, how do I opt out?
2. Data retention: how long is my data stored?
3. Confidentiality and access: who can access my content, and under what circumstances?
4. Content ownership: who owns the output I create?
5. User data rights: how do I delete my data and my account?

Platform tour: ChatGPT

The layout generalizes, so learning one interface well makes the others legible.

The one setting to find today. In Settings → Data controls, the option named something like “improve the model for everyone” governs whether your conversations are used for training. Turning it off is the difference between your work being private and your work being training data. Every provider has an equivalent, in a slightly different place.

Platform tour: Claude

Platform tour: Gemini

If you would rather have fewer of those features switched on than more, Digital Wellness collects the guidance for reducing AI features in the tools you already use, including a step-by-step guide to the Google ones.

Working with your own sources

The video below is a walkthrough of NotebookLM, which is built for exactly this: upload your material, ask questions of it, and generate summaries grounded in what you provided. The interface is specific but the skills transfer — uploading sources, asking questions of them, and checking the answer against the source is the same everywhere.

Your turn: an interface field study

25 minutes

Be a digital anthropologist. Go to Perplexity (opens in a new tab), ask it a few real questions, and open every menu. Then answer:

  • What does the Focus control do — academic, video, and the rest? How is that like or unlike a custom GPT?
  • How does it present sources compared with the tools you have used? More prominent, or less?
  • What is here that you have not seen elsewhere, and what is it for?

The point is transferable: the first thing to do with any new AI tool is find out what it is willing to show you about where its answers came from.

Optional: AI skills bingo — five in a row

Complete any five in a line. The goal is range, not mastery.

  • Ask for five facts about ancient Rome, then check two of them
  • Draft a professional email to a colleague
  • Generate an image of a cyberpunk library
  • Have a complex topic explained to a five-year-old
  • Build a three-day itinerary for a city you do not know
  • Write a short poem about coffee
  • Summarize a long news article in three bullets
  • Ask for a fifteen-minute workout plan
  • Run a deep research report on something you know well, and grade it
  • Get a recipe from three ingredients you actually have
  • Translate a sentence into five languages
  • Debug a broken code snippet
  • Create a custom GPT or Gem with a personality
  • Use Canvas or Artifacts to produce a document or a block of code
  • Compare two films, and see whether it has opinions or averages
  • Generate interview questions for a job you might want
  • Write a thank-you note that does not sound like a template
  • Use web search for next week’s weather, then check it against a real forecast
  • Build a quiz about a hobby
  • Write a social post for a product that does not exist
  • Plan a week of meals
  • Get book recommendations from your favorite author, and see how many are real
  • Ask for a short story in the style of a writer you know well — and judge the imitation
  • Make a table comparing three phones, then check one specification

Key takeaways

  • These are workbenches, not chatbots — the mode you choose matters
  • Chat for quick work, canvas or a project for anything long
  • Web search beats the knowledge cutoff; deep research beats a single answer
  • GPTs and Gems are specialists you build once and reuse
  • Find the training and data-sharing settings on day one

Part eight

The art of AI data

By the end of this part you should be able to

  • Say why what you provide matters more than how you ask
  • Classify Stanford data correctly before it goes anywhere near a tool
  • Select, structure, and format material so a model can use it
  • Use AI to prepare data for AI
  • Run a pre-upload check in under a minute

Garbage in, garbage out is the oldest rule in computing and the most relevant one here. The model is the engine; what you provide is the fuel. Selecting and formatting it carefully is how you tell the model what to pay attention to.

Ask for a summary of a contract and supply the contract plus fifty unrelated emails, and the termination clause you needed is buried in noise. Good material means better accuracy, faster answers, and lower cost, since most platforms charge by how much they have to read.

Stanford data governance comes first

These are requirements, not suggestions. Before any data goes into any AI tool, you are obliged to know its classification and the rules attached. Getting this wrong is a compliance problem, not a style problem.

A special case: FERPA and student data. The Family Educational Rights and Privacy Act protects student education records, and it does not bend for a convenient tool. Any AI that handles student data — assignments, grades, personal information — has to be compliant, which usually means a contract making the vendor a school official with a legitimate educational interest and real security obligations. The Department of Education’s guide (opens in a new tab) is the authority.

Be a surgeon, not a shovel

The context window is short-term memory, and overloading it is the most common mistake there is. Provide only what is essential.

Triage your documents

Rank them: critical, important, supplemental. Start with the critical ones only, and add more if the answer needs it.

The five-page rule

Uploading more than five pages? Stop and ask whether every page bears on your actual question. Extract the sections instead.

Be token-aware

More material means more tokens, slower answers, and on many platforms a higher bill. Precision is cheaper.

Four ways this goes wrong

Structure is the whole trick

Models do well with organized input. Label everything, keep dates and names in one format, and use lists to show hierarchy. Sixty seconds of formatting saves ten minutes of re-prompting.

how to hand over two documents
### DOCUMENT 1: Employment Agreement
**Parties:** Jane Doe ("Employee"), Acme Corp. ("Company")
**Effective Date:** 2024-01-15
**Key Clause (Section 4.a):** "Employee agrees not to solicit any Company
client for a period of 12 months following termination."
---
### DOCUMENT 2: Termination Email
**From:** hr@acme.corp
**To:** jane.doe@email.com
**Date:** 2025-10-20
**Subject:** Notice of Employment Termination
**Body Snippet:** "Your final day of employment will be October 31, 2025."

Your pre-upload checklist

  • Is this the minimum data needed?
  • Are my instructions clearly separated from the data?
  • Is every document labeled?
  • Are dates and numbers in one consistent format?
  • Have I removed the duplicates?

Red flags that your data needs work

  • The answers are vague or generic
  • It asks you to clarify basic facts
  • It pulls material from the wrong section
  • The response takes unusually long, or fails
Different data types, different preparation
  • Text documents. Pre-process them. Pasting clean text beats uploading a file with complex formatting or poor OCR.
  • Spreadsheets. Describe the columns and state the goal: “the attached CSV has columns Date, Amount, and Category — find all Legal Fees transactions in Q3.”
  • Transcripts. Add speaker labels (Attorney Jones:, Witness Smith:) so the model can follow who said what.
Advanced: two-stage prompting, validation, and practice data
  • Two-stage prompting. First prompt extracts the key clauses; second prompt analyzes only what was extracted. Smaller context, better answer.
  • A validation prompt. Before the real task, ask it to “review this data and flag any inconsistencies or missing information.”
  • Practice on public data. Build the skill on non-sensitive datasets from somewhere like Kaggle (opens in a new tab) rather than on anything real.

How AI uses the web, and why that is not the same as research

When a model fetches a page, it looks for patterns; it does not evaluate whether the source is any good. A .org domain guarantees nothing. Cross-check factual claims against a reputable source, and trust your own skepticism — if an AI-generated claim feels off, verify it by hand rather than asking the same tool again.

Your turn: read a robots.txt file

15 minutes

A robots.txt file is a website’s public rulebook for crawlers, AI ones included. It is the closest thing to a public statement of how a site feels about being used as training data, and reading a few is unusually clarifying.

Stanford SearchWorks, excerpt
Sitemap: https://searchworks.stanford.edu/sitemap.xml
User-agent: AdsBot-Google
Disallow: /
User-agent: SemrushBot
Crawl-delay: 60

That tells Google’s ad crawler to stay out and one commercial crawler to wait a minute between requests. It is a request, not a wall — a badly behaved crawler can ignore it. Read the whole file (opens in a new tab).

Canva, excerpt
User-agent: GPTBot
Disallow: /
User-agent: Claude-Bot
Disallow: /

A stricter position: named AI training crawlers are refused the entire site. Their full file (opens in a new tab) is long, and worth skimming for how specific the refusals get.

Now do it yourself:

  1. List five websites you rely on.
  2. Visit each one with /robots.txt on the end of the domain.
  3. Copy the contents of one of them.
  4. Paste it into an AI with this prompt: “Explain this robots.txt file to me like I’m a beginner. Based on these rules, how friendly is this website to AI crawlers?”

You are checking the model’s explanation against a document you can read yourself, which is the safest way to learn how much to trust its explanations.

When policy says you cannot add data to AI

A blanket “not allowed” usually means “not this data, or not this tool.” The productive move is to find the permitted workflow rather than to argue or to ignore it.

  1. Clarify the policy. Find the actual clause and read what it restricts.
  2. Classify the data. Regulated data, secrets, and attorney-client confidences do not go into unapproved tools, full stop.
  3. Minimize. Summarize or excerpt rather than pasting whole documents, and strip identifiers.
  4. Prefer the approved tool. An enterprise or university-hosted option, such as the Stanford AI Playground, over a consumer app.

A script for the conversation with a colleague: “Fair concern. Let’s confirm the policy, classify the data, and pick the safest permitted workflow. If it is regulated or confidential, we use an approved tool or we redact.”

An alternative: synthetic data

Artificially generated data that mimics the statistical shape of the real thing lets you work on a problem without exposing anything. Start from a clear purpose, use clean source data, and involve someone who knows the domain, or you will produce something realistic-looking and wrong.

Do: check statistical fidelity, validate against more than one metric, and treat anonymization as the point rather than a step.

Watch for: bias inherited from the source data, temporal gaps that make it quietly out of date, and privacy leaks through data that is less anonymous than it looks.

Key takeaways

  • If you are unsure whether a document is needed, leave it out — you can add it later
  • Structure is not optional: label, order, and format everything you provide
  • Use a first prompt to clean or extract, and a second to analyze
  • Before your next complex prompt, spend two minutes pulling out only the critical passages

Part nine

AI we do not recommend

By the end of this part you should be able to

  • Name specific tools that carry legal, ethical, or operational risk, and say why
  • Spot the red flags in terms of service and data handling
  • Evaluate a new tool on more than its capabilities
  • Recognize how AI is used against users, not just by them

Not all AI is equally safe to use. What follows is a risk-based assessment rather than a condemnation: these tools are capable, and the specific problems below are why we do not recommend them for professional or institutional work. The reasoning matters more than the list, because the list will change.

Assessed in autumn 2025, on the basis of active litigation, regulatory action, published research, and the terms of service as they then stood. Any of those can change; check before relying on this.

Meta AI — do not use for anything sensitive

Meta’s consumer assistant carries privacy problems serious enough to rule it out for work involving confidential information or institutional content.

  • Private chats have surfaced publicly. Its discovery feed has exposed sensitive user prompts — medical and legal among them — because the sharing controls are confusing and not private by default.
  • Group conversations are training data. Its policy allows training on user interactions including group chats, where not everyone present consented to anything.
  • Regulatory and legal exposure. Ongoing GDPR challenges in Europe and copyright litigation from authors over training data, both of which make the service itself less predictable.
Apple Intelligence — not for anything that has to be right

Promising, early, and not yet reliable enough for reference-grade work.

  • Accuracy. A documented history of generating false news headlines and pushing them as notifications.
  • Marketing and legal issues. A class action over advertising features that did not work, and an order from the National Advertising Division to drop misleading “available now” claims.
  • Availability. Key features require the newest hardware, and the rollout has been delayed in major markets over regulatory questions.
DeepSeek, the public version — use the vetted one instead

The model is genuinely interesting. The public service is at the center of intellectual property disputes, privacy investigations, and poor safety results.

  • Training provenance. OpenAI has presented evidence that DeepSeek trained on ChatGPT outputs, and researchers have found linguistic fingerprints suggesting Gemini outputs too.
  • Regulatory scrutiny. Investigated by regulators in Italy and South Korea over privacy, and flagged for national security review in the United States.
  • Safety. Audits found a very high misinformation failure rate, easy circumvention of its guardrails, and political censorship in its answers.

The safer route. A vetted DeepSeek runs inside the Stanford AI Playground, where you get its useful transparency about its own reasoning without the data going to the public service. That contrast is the clearest argument for using a managed environment at all.

Midjourney — not for institutional or commercial use
  • The terms shift all liability to you. If the tool generates an image that infringes somebody’s copyright, its terms make that the user’s problem, not the platform’s.
  • Active litigation from Disney and NBCUniversal, alleging the service functions as a vending machine for infringing images of their characters. An adverse ruling could change or end the service.
  • Operational unpredictability. Its terms allow it to change access and pricing at will, and association with a platform accused of large-scale infringement carries its own institutional risk.
Character.AI — not for personal or sensitive conversation
  • Privacy. Conversations here are often deeply personal, and the privacy policy contemplates review and use for training, with little clarity about retention.
  • Content and safety. Users can create any persona, which makes moderation genuinely hard and the environment risky for younger users.
  • Blurred lines. The design encourages treating the model as a companion, which raises real questions about emotional dependency that the product does not answer.

A resource for researching bad AI

The Midas Project (opens in a new tab) is not an assistant; it is a watchdog that tracks and documents AI-driven manipulation, disinformation, and privacy abuse. When you need a verifiable example of AI going wrong — for a paper, a policy, or an argument — it is a better starting point than searching.

How AI is used against people

The same capabilities that help you are being used to automate and personalize attacks, and knowing the shapes makes them easier to spot.

Two literacy skills

Spotting AI-written books

Generated titles are appearing in print, often generic, factually shaky, and strangely flat. Look for repetitive phrasing, unnatural word choices, a mechanical tone, and an absence of perspective. Check whether the author exists in any verifiable way, and whether the publisher has a reputation to protect. Read reviews before buying.

Context, and the screenshot problem

A model can be led. Feed it inflammatory or contradictory prompts long enough and it may produce something offensive — and a screenshot of that output, without the conversation that caused it, can be presented as the AI speaking unprompted. Before believing or sharing any AI output, ask to see the whole conversation.

And do not buy prompts

Marketplaces sell “expert” prompts. It is an unnecessary purchase for four reasons:

Your turn: assess a tool

20 minutes

The scenario. A new tool called “Synapse Weaver” is spreading through your team. The website promises next-generation insights and looks polished. Its terms say all user data is “anonymized and used to improve the service for everyone,” and that the company is not liable for any output. Signing up requires connecting your primary Google account.

Your job, as the person asked whether to approve it:

  1. What are the immediate red flags?
  2. Which of the tools above do those red flags resemble, and in what specific way?
  3. Name three questions you would put to the vendor before any pilot.
  4. Approve, deny, or pilot with restrictions — and justify it in writing, briefly.

This is the assessment you will actually be asked to make, usually with less time than this.

Key takeaways

  • Popularity and capability are not evidence that a tool is safe to use
  • Read the terms, especially on infringement liability and training on your content
  • Private by default is the standard; opt-out privacy is a warning sign
  • Active litigation and regulatory action predict instability you will feel

Part ten

The future, and your AI strategy

By the end of this part you should be able to

  • Name the directions the field is actually moving in, with examples
  • Explain why smaller models matter as much as larger ones
  • Describe what agentic tools change, and what they risk
  • Set up a way of keeping current that does not depend on hype

You have the fundamentals: what this is, how it works, what it makes, how to direct it, and where it fails. The last thing worth having is a sense of trajectory — not to predict it, but so that the next announcement lands as a variation on something you understand rather than as news.

This part is a snapshot from October 2025. Treat every named model and product as an example of a direction, not as a current release.

Where it was heading, as of late 2025

1. Small models with big minds

Samsung’s Tiny Recursive Model has around seven million parameters — thousands of times smaller than a flagship — and uses recursive reasoning to match or beat far larger models on logic tasks. The implication is that sophisticated reasoning need not require a data center, which changes where AI can run. The code is open (opens in a new tab).

2. The chatbot as an operating system

Apps inside ChatGPT (opens in a new tab) let you invoke third-party services in the middle of a conversation, in plain language. The chat stops being a tool and starts being a platform — which is a significant amount of leverage to hand to one interface.

3. Agentic browsers

The browser is being rebuilt around an assistant that acts: Perplexity’s Comet (opens in a new tab), OpenAI’s Atlas (opens in a new tab), and Gemini inside Chrome (opens in a new tab). A browser that reads and acts across sites on your behalf is useful and is also a new category of thing to be careful about.

4. Reasoning as a product decision

GPT-5 (opens in a new tab) routes a question to a fast model or a slower thinking model depending on difficulty; Anthropic ships a family that trades speed against depth; Microsoft’s Phi-4 reasoning models (opens in a new tab) match much larger models on math and code through careful data work. The pattern: paying more compute for hard questions, and not for easy ones.

5. Generated worlds, and video that obeys physics

Genie 2 (opens in a new tab) builds playable 3D environments from a single image, and video models such as Sora 2 (opens in a new tab) and Veo (opens in a new tab) produce increasingly physically plausible motion and sound. Useful for simulation and training; also the evidence problem of the next decade.

6. Science

The 2024 Nobel Prize in Chemistry (opens in a new tab) went in part to the work behind AlphaFold, which predicted the structure of nearly every known protein and closed a fifty-year problem. This is the strongest existing case that AI changes what is discoverable, not just what is fast.

7. Efficiency at scale

Mixture-of-experts designs keep an enormous model on disk but activate only a fraction of it per token — the DeepSeek-V3 report (opens in a new tab) describes 671 billion parameters with 37 billion active. Large models are getting cheaper to run as well as bigger.

8. Adoption without payment

Around 1.8 billion people used AI tools by mid-2025 (opens in a new tab) and roughly three percent paid for anything. Whatever else that gap means, it means the business model is unsettled — and unsettled business models are why free tiers change.

9. Open models

Alongside the closed systems, an ecosystem of downloadable models has grown up — Llama (opens in a new tab), Gemma (opens in a new tab), Mistral (opens in a new tab), and Switzerland’s Apertus (opens in a new tab) among them, mostly distributed through Hugging Face (opens in a new tab). Anyone can run and fine-tune these, which matters for transparency, for research, and for the cases where sending data to a vendor is not an option.

Keeping current without drowning

The field produces more announcements than anyone can follow, and most of them will not matter to your work. A small number of habits beats trying to keep up.

Read one weekly digest

The AI Upload, the library’s own, comes out on Fridays and is written for this community. One digest read properly beats five feeds skimmed.

Your strategy, in five lines

Worth repeating

Keep learning continuously, aim AI at a few high-value tasks rather than everything, choose tools deliberately, follow the data policies without exception, and stay the person who decides.

Where to go from here. Walk through the PAUSE Rule before your next real task. Download the AI Skills if you want structured help with writing and research. Browse Your AI Stack to build the set of tools that fits your work. Drop in at the Curiosity Corner with a question, or email library@law.stanford.edu for a one-to-one session. And if the useful next step is less AI rather than more, that is what Digital Wellness is for.

Key takeaways

  • The field is moving toward efficient reasoning, deeper integration, and multimodal work
  • Smaller and open models matter as much as the flagship releases
  • A personal strategy is continuous learning plus a few high-value uses
  • Institutional data policy is not optional, whatever a tool makes easy
  • Keep human oversight, and apply the PAUSE Rule before you start

That is the course

You are done

Ten parts, and you now have what you need to use these tools deliberately: what they are, what they cannot do, how to direct them, what to give them, and which to refuse. Keep experimenting, keep the journal, and keep the judgment on your side of the table.

Read the PAUSE Rule Try it in the Stanford AI Playground Back to the tutorials →