Sue The Map

They sued the AI. We mapped it. Spin the globe, pick your language (lawyer or human), ask questions out loud. 307 cases. 34 states. One file.

Comment

Inspiration

AI is being sued everywhere. Facial recognition in Illinois. Generative AI copyright claims in California. Hiring algorithm bias in New York. The GW Law DAIL database captures all of it, but it lives in a spreadsheet that takes a trained researcher to navigate.

We wanted to ask: what if anyone could walk up to this data and just ask it a question?

That became Sue The Map.

What We Built

Sue The Map is a single-page interactive visualization of 307 AI lawsuits across 34 US states, powered by the DAIL dataset from GW Law's Ethical Technology Initiative. A rotatable 3D globe color-codes every state by case density: transparent where there are no cases, scaling through yellow and orange to deep red for the highest-volume jurisdictions. You filter by year, sector, or status, and the globe redraws instantly. Click any state and every case filed there appears on the right, with status, sector, media coverage, and the full story behind it.

A Globe, Not a Map

We chose a 3D globe over a flat US map deliberately. AI litigation is not a domestic issue. It is spreading across continents, and this tool is built to grow with it. When international DAIL data becomes available, the globe is already waiting with no redesign needed.

Two Modes: Because the Same Data Serves Different People

Law Mode is built for practitioners. Filterable case tables, sector breakdowns, annual filing trends, precise legal classifications. Everything is labeled in proper legal terminology so nothing gets lost in translation.

Public Mode strips the jargon entirely. Sector names become plain descriptions. Case summaries are rewritten in everyday language. And the left panel's "Coverage Gap" tab becomes "Untold Stories": a curated list of active AI lawsuits with zero press coverage, framed explicitly for journalists who need to know what nobody is covering yet.

The same row in the database. Two completely different reads. The translation layer is the product.

Ask It Anything

The AI query panel is grounded entirely in the DAIL dataset. When you ask a question, the assistant receives a structured snapshot of all 307 cases: per-state counts, active/inactive breakdowns, media coverage flags, sector distributions, and yearly trends. It is strictly instructed never to fabricate a statistic. It tone-shifts automatically between modes: citation-heavy and formal in Law Mode, plain and conversational in Public Mode.

Multi-turn conversation is supported, so you can ask follow-ups without repeating context. If a state name appears in a response, the globe highlights it automatically.

Voice First, Because Access Matters

The voice feature was built with disability in mind, not as an afterthought, but as a primary interaction path.

A lawyer who is blind should be able to use this tool as fully as anyone else. A researcher with a motor disability should not have to type to explore 307 cases. A journalist listening while commuting should be able to ask questions hands-free.

Click the microphone once and it stays on. You speak, the AI answers, text-to-speech reads the response back to you, and the mic reopens automatically so the full conversation flows without a single click. The mic pauses while the AI is speaking so the two never talk over each other. Suggested query chips and full keyboard navigation ensure the tool works entirely without a mouse as well.

The Data Pipeline

Raw case data lives in an Excel sheet maintained by GW Law. A Python processing script cleans and resolves jurisdiction names, including district court abbreviations like N.D. Cal. and C.D. Cal. that a naive string match would silently drop. It counts cases per state per year, flags media coverage, and outputs a compact JSON file. A second script bakes that JSON directly into a single self-contained HTML file. No server required, no build step for the end user. Open the file and it runs.

Challenges

Data quality. Court jurisdiction names use a dozen different formats for the same state. We were silently dropping 14 cases because district court shorthands did not match against full state names. We built a regex-based court abbreviation resolver to catch them all.

Globe performance. Calling polygonCapColor() inside a mouse-move handler triggered full 3D mesh rebuilds on every pixel of movement, making the globe stutter visibly on interaction. Removing the hover handler and switching to tooltip-only rendering solved it completely.

Color design. The color scale went through six iterations. We needed zero-case states to visually recede without disappearing from the globe geometry, and high-case states to read as urgent without looking like an error state. The final scheme scales from fully transparent through pale yellow and orange to deep red, with opacity tied linearly to case count.

Dual-mode AI tone. The same system prompt cannot serve a federal litigator and a first-time reader equally well. We built a mode-aware prompt builder that injects the full dataset context in two different framings, one for legal audiences and one for the public, and switches between them live.

What We Learned

We learned that framing data is as hard as collecting it. The same lawsuit reads completely differently depending on who is in the room. Building both modes forced us to think carefully about what information is actually essential versus what is just familiar to people who already know the field.

We also learned that accessibility is not a checklist. Building a continuous-listening voice mode forced us to think through the entire interaction loop: what happens when the mic times out, what happens when the AI responds, what happens on an error, what happens when a blind user cannot see the globe update. Those edge cases are where most voice features quietly fail.

Future Scope

Retrieval-Augmented Generation with a Vector Database. Right now the entire dataset is injected into the system prompt as structured text. This scales fine for 307 cases but will hit token limits as DAIL grows. The next step is a vector database where each case is a semantically embedded document. The AI retrieves only the cases most relevant to the question before answering, dramatically improving accuracy and unlocking datasets an order of magnitude larger.

Global Dataset. The globe is already built for it. As DAIL expands to cover EU AI Act litigation, UK cases, and emerging markets, Sue The Map adds jurisdictions without any redesign.

Predictive Case Modeling. With enough historical data, machine learning can identify which states are statistically likely to see case surges based on legislative activity, industry density, and prior litigation patterns. A prediction layer on the globe would let policymakers and legal teams get ahead of emerging litigation hotspots rather than just documenting them after the fact.

Real-Time Data Sync. The current pipeline runs manually. Connecting directly to the DAIL API or a live legal filing feed would keep the globe current without human intervention.

Offline PWA and Screen Reader Support. Packaging as a Progressive Web App with full ARIA labeling would make the tool installable and fully navigable by screen readers, completing the accessibility picture for blind and low-vision users.

Built With

Share this project:

Updates