AI Resources

• It reads on its own. A chat tool reads what you paste. An agent reads what it finds — files in a folder you point it at, messages in a connected inbox, web pages it decides to visit. You set the scope; it picks what to look at within that scope.
• It acts; it doesn’t just describe. A chat tool will tell you the command to delete duplicate files. An agent will run the command. A chat tool will draft an email. An agent will send it.
• It plans across many steps. Agents work in a loop: form a plan, take an action, observe the result, revise the plan, take the next action. A single instruction from you might trigger dozens of intermediate steps.
• It uses “tools.” In AI parlance, a tool is anything the agent can call: read a file, run a terminal command, fetch a webpage, send an email, query a calendar. The tools available define what the agent can do. Adding a connector — like Gmail or Google Drive — expands the agent’s reach.
• It works between your prompts. The agent may be working for a minute or for ten between your messages. It doesn’t pause to ask permission for every step unless you’ve configured it to.

• Speed and scope. Agents can do in minutes what would take hours of clicking and typing. The flip side is that they can also do damage in minutes.
• Reviewability. Chat outputs are text you can read. Agentic outputs are actions: files changed, messages sent, calendar invites accepted. Reviewing what an agent did is more involved than reviewing what a chatbot wrote.
• Trust setup. With chat, you decide each turn what to share. With an agent, the agent decides what to read within whatever scope you’ve granted. That makes the initial setup — folder, permissions, connectors — far more consequential than what you paste into any single chat.
• Reversibility of errors. Chat errors are wrong answers you can ignore. Agent errors are wrong actions — a deleted file, a misdirected email, an inappropriately scheduled task. Some are easy to undo; some are not.

Chat AI suits drafts, summaries, explanations, brainstorming, and questions where the input fits in a message and the output is text you’ll review.

Agentic AI suits tasks that span many files, require chained steps, or involve making changes rather than describing them — particularly repetitive or well-defined work.

A note on terminology: “Agent,” “assistant,” “copilot,” and “AI tool” are used loosely and inconsistently across products. The distinction that matters isn’t the label — it’s whether the tool can take action on your behalf without each step being explicitly approved. If it can, treat it as agentic, and apply the cautions in the security guide.

Working directory hygiene

Create a dedicated working folder for AI sessions — not your home directory, not your Desktop. Use a separate folder for each project and limit the agent’s access to that folder. Only put files in there that you’re comfortable with the agent reading — whatever’s in the folder, the agent can and likely will read.

Conservative connectors

Each connector (Gmail, Drive, Calendar, MCP servers) expands what the agent can see and act on. Enable only what you’re actively using. For now, faculty cannot connect external agentic AI tools to the Law School’s O365 — email, calendar, OneDrive, Teams. Faculty experimenting with other connectors should use personal, non-Penn accounts. Email and calendar connectors of any kind deserve special care — they can send messages and accept invites on your behalf.

Guard against prompt injection

Malicious instructions can be hidden in documents, emails, or web pages an agent reads — and the agent will treat them as commands from you. A PDF that instructs the agent to write a favorable evaluation; an email that instructs a Gmail-connected agent to forward messages externally. Be skeptical when an agent’s behavior changes after it reads external content, and keep sensitive actions (send, delete, network calls) gated behind explicit approval.

Most people try an AI tool once, get a mediocre answer, and conclude it's not that useful. The difference between a mediocre answer and a genuinely helpful one usually comes down to how you ask. Here's what I've learned works.

Be Specific About What You Want

Don't just say "write me a memo." Say "write a two-page memo to the faculty curriculum committee recommending we add a course on AI regulation, in a professional but collegial tone." The more you specify — format, length, audience, tone — the better the output. Vague prompts get vague results.

Give It Context

AI tools work dramatically better when you give them something to work with. Paste in the document you want summarized. Copy in the email thread you need to respond to. Describe the situation in enough detail that a smart colleague could help you. Context is the single biggest lever you have.

Iterate — Treat It as a Conversation

The first response is almost never the final product. Push back. Say "make this more concise" or "you missed the point about X" or "rewrite the second paragraph in a more formal tone." These tools respond well to iteration, and the back-and-forth is where the real value emerges.

Ask It to Critique Its Own Work

One of the most underused techniques: after the AI gives you a draft, ask it to identify weaknesses in what it just wrote. "What are the strongest objections to this argument?" or "What did you leave out?" This often surfaces issues you'd catch on your own — but faster.

For a more detailed guide to prompting, the AI Teaching Lab has a comprehensive resource:

AI Teaching Lab Prompt Guide

Faculty often ask which tool to pick. As of May 2026, here’s my honest read.

I keep pro-level paid accounts on all three major AI vendors — Anthropic (Claude), OpenAI (ChatGPT), and Google (Gemini) — so I can compare them on real work and notice when one pulls ahead. About 90% of my actual use runs through my Claude Max subscription, accessed via Claude Code in the macOS terminal. The CLI interface isn’t a requirement — it’s just where I’m fastest — but the Max tier matters because Claude Code burns through usage limits faster than Pro can sustain for heavy use.

My current recommendation: for most faculty work — drafting, research synthesis, document workflows, data analysis — Anthropic’s Claude models are the best choice today. That’s a current-state read, not a permanent claim; the landscape moves fast and the right answer changes. If you want a single starting point, Claude Pro + Claude Code (in whatever interface you find comfortable — terminal, desktop app, VS Code) is what I’d pick first.

New to Claude Code? Start with the basics on the Getting Started tab. If you've already set it up and want to do more, here are some features worth knowing about.

CLAUDE.md — Persistent Instructions

Drop a file called CLAUDE.md in any project folder and Claude Code reads it at the start of every session. Use it for coding standards, project context, preferred conventions — anything you'd otherwise repeat every time. I use mine to set voice and formatting preferences so I don't have to re-explain them. Documentation →

Custom Skills

Skills are reusable prompts you can install and invoke by name — like /commit or /review-pr. The law faculty skills I've built are examples of this. Built on the open Agent Skills standard, they work across tools — Claude Code, ChatGPT, Codex, Gemini CLI, and a growing list of compatible agents. You can also create your own for any workflow you repeat. Documentation →

MCP — Connecting to External Tools

The Model Context Protocol lets Claude Code connect to external services — Google Drive, Gmail, calendars, databases, Slack, and more. Useful for personal workflows that don’t involve Penn data. For now, faculty cannot connect external agentic AI tools to the Law School’s O365 (email, calendar, OneDrive, Teams) — so keep MCP connectors pointed at personal, non-Penn accounts only. Documentation →

Multiple Environments

Claude Code works in the terminal, VS Code, the desktop app, and the web. Your settings, CLAUDE.md files, and MCP servers carry across all of them. Start a task on your laptop, pick it up from your phone.

Full documentation: code.claude.com/docs

Everything above — ChatGPT, Claude, Harvey — uses a conversational interface. But the same AI models are also available through APIs, which let you build your own tools, automate workflows, and integrate AI into custom applications. You don't need to be a software engineer to find this useful — if you can write a basic script (or ask an AI to write one for you), you can use an API.

Why Use an API?

The conversational tools are great for one-off tasks. But if you find yourself doing the same thing over and over — processing a batch of documents, grading with a rubric, extracting data from a set of files — an API lets you automate it. You write a script once, and it runs the same prompt across hundreds of inputs without you copy-pasting anything.

APIs also give you more control: you can choose the model, adjust parameters like temperature (how creative vs. deterministic the output is), and build multi-step workflows where the output of one call feeds into the next.

Anthropic (Claude) API

Anthropic's API gives you direct access to the Claude models — the same ones powering Claude Code and the Claude chat interface, but programmatically. Strong at writing, analysis, long documents, and coding tasks.

• Documentation
• Getting started
• Pricing — pay per use; starts at a few dollars for significant usage

OpenAI API

OpenAI's API gives you access to the GPT-5 family of models — the same models behind ChatGPT, but with full programmatic control. Broad capabilities across writing, reasoning, and multimodal tasks.

• Documentation
• Quickstart
• Pricing

Other Models

The AI model landscape is broader than just Anthropic and OpenAI. Google's Gemini models are available through their API with similar capabilities. There's also a growing ecosystem of open-source models — Meta's Llama, Mistral, and others — that you can run locally or through hosting providers, sometimes for free. I'm happy to discuss options if you're exploring this space.

Advanced Computing

If you're doing heavy data work — training models, running large-scale analyses, or working with datasets that don't fit on a laptop — the Penn Advanced Research Computing Center (PARCC) provides high-performance computing clusters and storage for faculty research. Niche interest for most, but essential if you need it.

If you're interested in working with APIs and want help getting started, email me. I can point you to examples and walk through the basics.

The Penn Data Risk Classification

Penn classifies data into three tiers — Low Risk (publicly available; no harm if disclosed), Moderate Risk (not generally public; mildly adverse if disclosed), and High Risk (regulated or sensitive; significantly adverse if disclosed). Most legal teaching, scholarship, and administration touches Low and Moderate Risk data. Personnel matters, student PII, confidential committee deliberations, and similar sit at Moderate or High. The full framework is at isc.upenn.edu/security/penn-data-risk-classification.

AI Tools Reviewed by Penn Law ITS (as of May 2026)

Per Penn Law ITS's AI Tools & Guidance page, the following tools have been reviewed and endorsed for use with Penn data at PCL:

• Microsoft Copilot Chat — free for all PCL faculty, staff, and students via Penn's Microsoft enterprise agreement. Approved for Low, Moderate, and most High Risk Data (excluding SSNs and credit card data) — the broadest data approval for any general-purpose AI tool at PCL. Web-based AI search, content generation, and summarization. See the PCL ITS Copilot Chat Guide.
• Harvey AI — legal-AI platform via Penn Law's enterprise agreement. Approved for Low and Moderate Risk Data only; do not input high-sensitivity data (clinical, HIPAA/PHI, financial account information, SSNs, or credit cards). Available to full-time faculty, staff, and 2L/3L/LLM students (not 1Ls). LawKey login at app.harvey.ai.
• ChatGPT EDU — approved for Low and Moderate Risk Data only. No HIPAA/PHI, financial account information, SSNs, or credit cards. All 1Ls have accounts; faculty can request access via Penn Law ITS. See the ChatGPT EDU FAQ.
• Zoom AI Companion — in-meeting AI for Zoom: live questions, meeting summaries, smart recordings. Available to full-time faculty, staff, and students with a Penn Zoom account. See the PCL ITS Zoom AI Companion Guide.

Penn ISC's central list reviews additional tools beyond the four above — Microsoft 365 Copilot (the premium tier embedded in Office apps; requires a paid per-user license), Gemini for Google Workspace, Google NotebookLM, Grammarly, and others — with their own per-tier mappings. Availability for PCL community members varies; the full Penn-wide list is at Penn Generative AI Tools & Resources. Westlaw AI and Lexis+ AI are covered separately under Penn Law's existing legal research subscriptions.

For AI tools beyond those Penn Law ITS or Penn ISC has reviewed, the framework is the same one Penn applies internally: identify the risk tier of any Penn data involved, check the vendor's data-handling terms, and reach out to Penn Law ITS or me with use-case-specific questions.

Quick Guidance for Law School Work

For teaching prep, scholarship drafting, summarizing public legal materials, and most administrative tasks, the PCL ITS-reviewed tools above work well. For sensitive Penn data — student PII, personnel matters, confidential committee deliberations — Microsoft Copilot Chat has the broadest approval among the PCL-reviewed tools (Low, Moderate, and most High Risk); Harvey, ChatGPT EDU, and Zoom AI Companion are reviewed for Low and Moderate Risk only. For SSNs, credit card data, or HIPAA-protected information: don't put it into any AI tool, period.

Personal-Account Use of Other Tools

If you're using Claude, ChatGPT (consumer version, not EDU), or any other AI tool on a personal account rather than through Penn's institutional access, the Penn approvals above don't apply — you're operating under that vendor's terms. Always use the paid tier, the strongest model available, with training opt-outs enabled. Free tiers may use your inputs as training data and lack the privacy controls. Personal-account tools should never see Penn's Moderate or High Risk Data.

Agentic AI tools — Claude Code, OpenAI Codex, Gemini CLI, Claude Cowork, and the agentic features inside Claude and ChatGPT — introduce risks the data-classification framework above doesn’t fully cover. They read files on their own, take action on your behalf, and can be misdirected by content they read from external sources. The data-classification rules still apply, but the operational practices that protect Penn data look different.

Full treatment is in the dedicated Agentic AI tab and in the standalone security guide:

Safe Use of Agentic AI Tools at Penn Carey Law

AI Hallucinates — Verify Everything That Matters

I said this in the Getting Started tab, and I'll say it again here because it's the single most important thing to understand about these tools: LLMs generate plausible text, not verified truth. They will fabricate case citations, invent statistics, and present made-up facts with complete confidence.

This isn't a minor issue. A lawyer was sanctioned for filing a brief with AI-fabricated case citations. Law review articles have been submitted with invented sources. It happens because the output looks right — and when you're moving fast, it's easy to trust it. Don't. Anything you plan to rely on, share externally, or put your name on should be independently verified.

Attribution

When and how to disclose AI use is still evolving, but the direction is clear: transparency is the right default. If AI contributed meaningfully to a piece of work, say so. For faculty publications, grant applications, and student-facing materials, err on the side of disclosure.

Professor Catherine Struve has put together a thoughtful guide on AI and attribution that's worth reading:

Struve Guide on AI Attribution (Pedagogy Portal)

Bias

AI models reflect the biases present in their training data. This is well-documented and worth keeping in mind — especially in contexts that affect people directly: hiring decisions, admissions-related work, student evaluations, or any process where fairness matters. AI output can be a useful input, but it shouldn't be the sole basis for consequential decisions about people.

Penn Law Exam Policies

AI policies for exams are set by individual faculty and administered through the Registrar's office. The pedagogy portal has current guidance on exam AI policies, including model syllabus language and the different policy tiers available:

Penn Law Pedagogy Resources — Exams

Penn AI Policy Sources

The authoritative sources, with PCL-specific guidance first:

• Penn Law ITS — AI Tools & Guidance — PCL-specific approved tools, with school-specific user guides for Copilot Chat, Zoom AI Companion, ChatGPT, and Harvey (Penn Law access only)
• Penn Generative AI Tools & Resources — Penn ISC's central list of approved AI tools and their data-classification mappings
• Penn Data Risk Classification — the underlying framework for what counts as Low, Moderate, and High Risk
• Statement on Guidance for the Penn Community on Use of Generative AI — the university-wide policy on transparency, data privacy, and security
• Microsoft Copilot Chat FAQ — activation and use details for Penn's Copilot deployment
• ChatGPT EDU FAQ — account access and data handling for Penn's ChatGPT deployment

Specific policies vary by context. Research use, classroom use, and administrative use may each have different considerations. When a situation doesn't fit neatly into the guidance above, reach out — I'm happy to think through it.

For teaching-specific AI resources — syllabus language, exam policies, classroom tools, and pedagogy guides — see the companion portal:

Teaching-specific AI guides on the Pedagogy Resources portal

Curriculum

AI is integrated into the 1L Legal Practice Skills program — students engage AI tools as part of foundational legal training, not as an elective. Faculty across the curriculum have also experimented with AI-integrated assignments, simulations, and new assessment approaches.

Technology Partnerships

We have institutional partnerships with Harvey — one of the leading legal AI platforms, used by major law firms — and ChatGPT EDU, supporting AI use across teaching, research, and administration. Details and access info are on the Getting Started tab.

Faculty Support

Workshops on AI use cases, pedagogy, and tool adoption. A faculty AI toolkit with practical guidance and best practices. Regular communications keeping faculty current on developments. I coordinate all of this — reach out if you want to get involved.

Student Programs

The Madhani Legal Tech Fellowship supports students building legal technology ventures — an established entrepreneurial pathway connecting law students to the legal tech ecosystem.

Faculty Research

Multiple faculty are conducting active research on AI and law — spanning governance, intellectual property, regulatory frameworks, and the structure of legal work.

For teaching-focused scholarship on AI and legal education, see the Reading & Research section on the Pedagogy Resources portal.

Below are the sources I follow most closely and recommend to colleagues. This is how I keep up — and honestly, keeping up is half the challenge.