Group Alpha - AI Risk Decoder

Why It Matters

It demystifies AI service contracts, giving users actionable insight into privacy and safety risks, which can shape adoption choices and pressure providers toward clearer, more responsible policies.

Key Takeaways

• Legal AI documents are lengthy, posing user comprehension risks.
• Built a JSON‑based database to store AI service policies.
• CLI tool ensures consistent, duplicate‑free entries for policy data.
• Vectorized risk encoding matches user concerns via cosine similarity.
• React front‑end displays personalized risk cards with source citations.

Summary

Group Alpha introduced the AI Risk Decoder, a platform that translates sprawling terms‑of‑service and privacy policies of popular AI services into concise, user‑friendly risk summaries. Recognizing that most users lack the time or expertise to parse legal jargon, the team built a website that surfaces the most relevant hazards based on individual preferences.

The project began by scraping policy documents and storing them in a semi‑structured JSON schema, enabling flexible representation across diverse services. A custom command‑line interface handles viewing, adding, editing, and deleting entries while preventing duplicates and preserving database integrity. After manual risk extraction, the team measured inter‑annotator agreement and refined prompting strategies by comparing human‑annotated risks with those generated by large language models. Risks are encoded as semantic vectors; at runtime, a user’s input is similarly vectorized and matched via cosine similarity to surface the most pertinent concerns.

To contextualize abstract risks, the team integrated real‑world incident data from sources like the OECD.AI database, categorizing examples into six thematic groups. Each AI service appears as a flip‑card on the React front‑end, with risks displayed on the front and recent news articles on the back, complete with source links for verification. Users can search, filter, and dive deeper into detailed explanations and multiple supporting articles.

By delivering personalized, evidence‑backed risk assessments, the AI Risk Decoder empowers consumers and enterprises to make informed decisions about AI adoption, potentially driving greater transparency and regulatory compliance across the industry.

Original Description

Client - Edyta Bogucka, Nokia Bell Labs

Amir Fazel, Dhruv Patel, Finley Stirk, Igor Michalec, Mara Powney, Sritej Tummuru

AI systems are increasingly deployed in sensitive domains such as hiring, healthcare, and policing — but the risks they pose are often hidden deep inside compliance documents, risk registers, or opaque terms of service. Legal Design seeks to make these risks visible, understandable, and actionable for non-experts. Your task is to create an AI-powered assistant that can extract, interpret, and redesign how AI risk information is communicated. The system should take legal or policy text (e.g. risk assessments, safety standards, liability clauses) and transform it into accessible outputs: personalised summaries, visual diagrams of risk flows, or “at-a-glance” dashboards for different stakeholders (citizens, regulators, companies). The challenge lies in combining natural language processing, risk modelling, and user-centred design to prototype a tool that makes AI accountability tangible.