A university-hosted place to use several leading AI models with your SUNet ID, at no cost, without
sending your work to a vendor. This guide covers what it is, how to drive it, which model to reach for,
what the agents do — and what it deliberately leaves out.
Faculty, staff, students, and other Stanford affiliates, with a SUNet ID
Cost
Free to eligible Stanford users
Not approved for
High Risk Data, including PHI and PII — that limit is the point of the next box
The Playground is not a vault. It is a Stanford-managed environment, which is a
different claim from approved for anything. Do not enter personal, confidential, or High Risk Data,
verify facts, citations, and calculations before you rely on them, and keep a person — you
— responsible for the final decision. Stanford’s
Responsible AI guidelines↗ (opens in a new tab)
are the governing document.
The model names in part two were current in autumn 2025. Vendors rename, retire, and add models
constantly; treat the names as examples and the model menu in the Playground as the authority.
Part one
What it is, and how to use it
One secure interface to models from several vendors, reached with the credentials you already have.
A secure environment. You sign in with your SUNet ID, and your chats and data stay inside Stanford’s environment rather than going to a vendor account.
Several models, one window. Switch between OpenAI, Anthropic, and Google models mid-conversation, or run two side by side.
Stanford-specific tools. Agents that search the Admin Guide, the Faculty Handbook, and DoResearch policy.
Free to the community. No subscription, no billing, no payment method.
Six things to know how to do
1. Log in and set up
Go to aiplayground.stanford.edu↗ (opens in a new tab)
and sign in with your SUNet ID. The first time, you may meet an Information Release
screen: choose a duration and accept. Read the data use policies while you are there — the
Playground is not approved for High Risk Data.
2. Prompt, and attach files
Type in the text field and press Enter. Attach PDFs, CSVs, Word files, or images with the paperclip
icon. Once an answer comes back you can edit the prompt and re-run it, copy the reply, or use
Fork to branch the conversation and try a different direction without losing the first one.
3. Explore and compare models
The dropdown at the top switches models mid-conversation. To compare, select one model, click the
+ button, then select a second: your next prompt goes to both and the answers
appear side by side. Part two is about which two to pick.
4. Use assistants and agents
Azure Assistants includes a data and code analyst that can work through a file you upload.
Agents connects the model to outside services — web search, ScholarAI for academic
papers, Wolfram for computation, image generation. Part three covers the ones worth knowing at SLS.
5. Customize and control
The right-hand panel is where you save prompts you reuse, adjust model parameters such as
temperature and output length, manage uploaded files, and group conversations with bookmarks.
6. Share a conversation
Share creates a view-only URL for a conversation. Only someone who can authenticate with a
SUNet ID can open it, and you can manage or delete every link you have made under
Settings → Data Controls.
How it compares to the public tools
The honest summary is that the Playground trades the newest features for a managed environment. Both
halves of that trade are real.
The Stanford AI Playground compared with public AI tools
Feature
Stanford AI Playground
External AI tools
Security and privacy
Stanford-managed environment with single sign-on; data stays within Stanford systems.
Data goes to a third-party vendor, under whatever its privacy policy says.
Model access
Several vetted models — OpenAI, Google, Anthropic — in one interface.
Usually one vendor’s models.
Cost
Free to eligible Stanford users.
Paid subscription or usage-based billing.
Newest features
Added after Stanford review for safety and compliance, so later.
Direct, early access to the newest models and experiments.
Best for
Safe experimentation, prototyping, and compliant workflows for academic work.
Niche capabilities, or the absolute latest version of a model.
Where to ask questions
Two Slack channels, for two different kinds of question.
#ai-playground↗ (opens in a new tab)
— the official university-wide channel, owned and monitored by University IT. Announcements,
technical questions, feedback.
For technical support, submit a Help request through Stanford IT; the
UIT service page↗ (opens in a new tab)
is the official home of the tool, and the Playground itself carries a feedback form.
Part two
Choosing a model
The Playground is not one tool but a menu of them. Knowing what each family is good at is most of
the skill, and none of it requires understanding how they work.
Azure OpenAI
The ChatGPT models
The versatile workhorses. For legal work they are good at drafting communications, sketching an
initial argument, and turning something complicated into plain English. The more capable ones will
analyze a document or write code for computational tasks.
“Azure” is Microsoft’s cloud. An Azure OpenAI model is OpenAI’s technology
delivered through Microsoft’s enterprise cloud, which is the part that matters to Stanford:
it is what makes the security and data-handling story different from the consumer app.
Best for: general drafting, complex reasoning, document analysis.
Reach for it when: you want a reliable writing partner, or a legal concept explained simply.
Anthropic
The Claude models
Built with an explicit emphasis on safety, and unusually good with long documents. That makes them
useful for reviewing a lengthy contract, summarizing a deposition, or working through extensive
case law: a large context window means more of the document is in view at once, and the analysis
holds together better across it.
Best for: summarizing long texts, code analysis, careful writing.
Reach for it when: you have a dense document to interrogate, or a nuanced argument to draft.
Google
The Gemini models
Known for speed and very large context windows. The “flash” versions are made for quick
work — a translation, an email — while the Pro versions take an enormous amount of text
at once, on the order of a very long book. Useful for deep research across a pile of material, and
for reading charts and images rather than only text.
Best for: multimodal analysis, complex coding, deep research.
Reach for it when: the volume is the problem, or the source is not plain text.
DeepSeek
The model to run here rather than anywhere else
DeepSeek is an open model, so its architecture is more transparent than its commercial
counterparts, and in the Playground it will show its reasoning as it works. For a profession where
the rationale matters as much as the conclusion, that is worth having.
It is also the clearest argument for using a managed environment at all: running DeepSeek outside a
vetted one carries security and data-privacy risk you do not need to take, and inside the
Playground you get the transparency without exposing your work.
Best for: coding, technical questions, math, logic.
Reach for it when: you want to watch how a model reasons through a problem, step by step.
Comparing two models on the same prompt
The Playground’s best feature, and the fastest way to learn what the differences actually are:
send one prompt to two models and read the answers next to each other.
Select your first model from the dropdown at the top of the page.
Click the plus icon (+) in the top menu. The selected model’s name appears in the prompt field.
Select your second model from the dropdown.
Type your prompt and press Enter.
Read the two answers side by side.
Part three
Agents
The specialized tools in the model menu, and what the word means here — which is not what it
means in the news.
Two different things called “agent”
In the Playground, an Agent is a specialized tool that gives a model one specific
capability — these used to be called plugins. Select ScholarAI and you have equipped the model to
search a database of academic papers. It uses the tool; you still direct the process.
Agentic AI, the phrase in every headline, means something else: a more autonomous
system that reasons, plans, and carries out a series of tasks toward a goal without being told each
step. The Playground’s agents are not that.
Think of the Playground’s agents as on-demand specialists you can call in to help a generalist.
The agents worth knowing at SLS
Stanford-specific resources
A secure, quick way to query the University’s own documents.
Faculty Handbook Search — policies, appointments, leave, and governance, answered out of the official handbook.
DoResearch Policy Guide — grants, compliance, and research administration policy, for anyone with research obligations.
Admin Guide Search — the operational questions, from procurement to event planning.
Research and scholarship
These connect the model to knowledge bases outside it.
ScholarAI — searches and analyzes peer-reviewed articles, journals, and legal publications.
Google web search — real-time search, for a recent ruling or a legislative change the model was not trained on.
Wolfram — a computational engine, for law and economics or statistical evidence.
Content generation
Imagen and DALL·E — custom visuals for a presentation or course material, or an illustration of a concept that is hard to describe.
What the Playground leaves out is a decision, not an oversight. Knowing what is missing tells you
when to open something else — and what you are accepting when you do.
What it intentionally leaves out
Consumer apps carry capabilities the Playground does not offer:
Connections to your personal accounts, such as Gmail and Drive
One-click deep research modes that read the live web and return a cited report
Canvas-style workspaces for longer writing or live code
Built-in data workbenches for spreadsheet and file analysis
Real-time voice conversation
Creating and sharing your own custom assistants
Six things the consumer apps do that this does not
Read your Google files and email
Gemini extensions can look through your Gmail or Drive with your permission — summarize a long
thread, pull facts out of a folder, draft a reply with links back to the source. You keep control and
can withdraw the access.
Produce an auto-researched, cited report
Deep research and web search modes browse many pages, compare what they find, and return an
organized report with links you can check. Ask for a comparison of three degree programs and you
get sections, trade-offs, and references.
Work in a builder rather than a chat
ChatGPT Canvas and Claude Artifacts open a panel where you and the model co-edit text or code,
preview the result, and iterate in one place.
Run code over your spreadsheet
Upload a file and have the assistant clean it, chart it, and run statistics on it. ChatGPT’s
advanced data analysis runs Python in a private sandbox to do this.
Talk to it
ChatGPT, Gemini Live, and Claude support real-time voice, which is the difference between using AI
at a desk and using it while walking.
Build a reusable assistant
For a task you repeat, ChatGPT has custom GPTs and Gemini has Gems — your own small expert,
with its own instructions and starter files.
Stay in the Playground when you want to
Try different models on Stanford-managed infrastructure
Draft, summarize, or get coding help without connecting a personal account
Keep it simple — nothing to install, no permissions to grant
Open a consumer app when you need to
Search the live web and get a source-backed report
Let the model read your own Gmail or Drive
Work in a canvas on long writing or live code
Upload data and have it analyzed and charted
Use your voice
And Perplexity, the “answer engine”
Perplexity↗ (opens in a new tab)
is a public tool that searches the live web and shows its sources beside each answer. It is built for
quick, checkable answers, and it is not a Stanford-managed environment. What it adds:
Cited, live-web answers. It fetches current pages and shows citations inline, so you can check the source rather than trust the summary.
Deeper research modes. A multi-step search that synthesizes many sources for a complicated question.
Focused searches. An academic mode that aims at scholarly sources when you need papers.
Spaces. Project areas that group related threads, hold files, and take collaborators.
Pages. Turning a finished thread into a clean, shareable write-up.
Good fit for finding and citing current information, or gathering sources you intend to verify. Use the
Playground instead when you want the managed environment, or when you are drafting and thinking rather
than searching.
Before you use any of them on real work. Walk through
the PAUSE Rule, and if a term on this page is new, it is probably in
AI Explained Using Analogies. This is an unofficial guide maintained by
the Robert Crown Law Library; University IT owns the Playground itself.
