Legal Mitigations for AI-Generated Harms

Explore

Expert Panel

Instruction

March 2 (Thu) 12:00-1:30 PM CSE376

Participants: Rachael Hong, Rachel McAmis, Liwei Jiang, Kaiming Cheng, Gregor Haas, Tina Yeung, Miranda Wei, Miro Enev, Yoshi Kohno, Inyoung Cheong

[ Primary goal is to get as many different things on the board as possible – optimize for breadth of use cases, stakeholders, datasets, impacts. Use “other ideas” section to record other ideas that are worth discussing later, e.g., ideas about how to build better systems ]

Overview

In this expert panel, we are envisioning the future of generative language models and their use cases. We want to understand both what good things might be in the future but, especially for our purposes, also what bad things might be in the future. When envisioning “bad things”, our goal is to consider situations in which there are no technical or policy mechanisms in place that might prevent, mitigate, or minimize those “bad things”. I.e., even if we do not think that a certain “bad thing” might happen, for any reason (policy or technical), let’s still envision them here.

When we say “generative language models”, we are thinking of large-scale systems that use an understanding of language to generate text (written or spoken or otherwise communicated). That text might be an answer to a question, an answer to a prompt, part of a communications exchange (a dialog), or more – this session is about envisioning those possible use cases, so do not restrict your thinking only to the types of systems we currently encounter.

To do at / before start of meeting

Create the following regions on the whiteboard / wall:

Use Cases

Stakeholders

Datasets

Impacts (areas for “good”, “bad”, “other”)

Changes and Resulting Impacts

Other ideas, defenses, project ideas, etc

Procedure:

Use cases and stakeholder brainstorming (different regions of whiteboard board / different color post-it notes) [5-10 minutes + discussion] [do not be constrained to current technologies] [stakeholders include users, non-users, companies, governments, M&V populations, other countries]

Datasets + incentive structure and impacts brainstorming (different regions of whiteboard board / different color post-it notes) [5-10 minutes + discussion] [datasets means “inputs to training systems” and impacts means “what happens as a result”; impacts can be “good”, “bad”, “both good and bad”, “unclear”]

Results

⁠

⁠

⁠

Use case

Writing/summarizing

Creative writing

Generating social media content / news

Meeting notes & agenda making

Writing help

Text summarizing

Summarizing articles or other content

Reading/writing assistant

Summarization (downstream tasks in NLP)

Dating conversations

News generation

Journalist news writing

News/media reporting

News generation

Programming

Creative Co-pilot

Coding

Code comment generation

Code generation

Code generation in new domains (hardware? Other misc. Boilerplate?

Text-to-app

Generating latex or code

Data generation (reverse prompting?)

Inputting problem statement & getting out an algorithm

Inputting app spec and outputting app code

Professional Services

Legal

Legal help

Writing waivers or other legal documents

Legal document drafting

Automated defense attorneys

Legal system

Making legal decisions based of court script

Writing legal documents

Generating privacy policy, Terms of Services

Business document drafting

Medical

Given symptoms, output diagnosis (automated Web MD)

Medical doctor understanding & analysis

Financial/others

Investment theses/RL

Advice: what should I do to get promoted?

Tell me recipe given ingredients

Professional assistant tools for doctors, CPAs, lawyers

Menus, food descriptions, nutrition labels

Generative tech used in other AI fields e.g., planning for debts

Research help

Idea generation

Help science communication by translating things into lay language

Generating research ideas

Research to gauge the “average sentiment” of a topic/person, etc.

Overleaf + ChatGPT

Finding research methodology to use

Writing background section or conclusion of research paper

Academic writing

Give difficult Q+A after practice talk

Arts & culture

Machine generated art & literature

Text-to-video

Writing songs, movies, plays, etc.

Entertainment (Books/TV)

Prediction of non-word features (large music models, large image models?)

Content recommendation → on-demand content generation

Personalized content creation/curation

Companionship

Give pets a voice

Communication substitute

Companionship systems (e.g., elderly who live alone to converse with)

Metarverse dating apps

Talk with deceased relatives

Deceased relative

People who want chatbot friend

Imaginary pets (Damagochi)

Virtual romantic partner

Imaginary significant other

Apps that mimics interacting with celebrity

Emotional support/Therapy

Relationship advice

Therapy sessions

Personalized emotional support. Therapist?

Automatic therapy

Talk-based therapy

Sentiment: based on chats or emails from person X, what does person X think about me?

Education

Education, homework help

Education e.g., language learning tools

Children in school learning to {write} (Insert topic)

Machine generated education curriculum

Virtual school tutor/teacher

Help learn new language

Interactive QA system for kids to learn about reading & thinking

Dissemination of history

Translation

Translation apps

Language translation to facilitate inter-cultural/regional communication

Practicing something in another language

Information

Search engine

Search for knowledge

Marketing

Creating marketing language

Advertising

Customer interactive, compelling ads

Recommendation systems maybe personalized ads

Virtual self

APs artificial personas

Personalized virtual representation

Personal assistant/planner

Personal assistant

Event planners

Virtual assistants / chatbots

Customer service

Government benefit claims & complaints

Patient intakes systems

Bank assistants

Tech support scripts

Accessibility

Accessibility (dyslexia, neurodivergence)

Accessibility tools

Harmful content

Spam/scams

Terrorism propaganda

Disinformation & astroturfing

Automated spam/harassment

Deep fake audio (phishing…)

Sexual materials

Porn

Interactive deep-fake/pornography

Celebrities

Past romantic partners

Stalking victim

Emotionally engaging content

Political persuasion

Cognitive science strong generating stimuli

Law enforcement

Police/immigration: Interrogation assistant

Government uses - summarize spy messages, search for secret information

Government summarize intercepted calls

Others

Firewall monitoring (regex → LLM) packet introspection

Circumvent your “textual footprint” by generic LLM language instead

Child protective services interview assistants

Improve training (if there is a speech synthesizer)

Online content moderation

Metal model of minds

App for “Ferris Bueller” (Help trick parents about locations)

Parent’ control on child’s personal LLM

Stakeholders

Education

Children

Children-parents

Children

Child interacting on web

Parents

Teachers

Teachers, tutors, coaches

Teachers-students

School administrators

Students

Students-children

Professionals

Authors/writers

Writers (journalists, marketing)

Newscasters, journalists

Instruction

Overview

To do at / before start of meeting

Want to print your doc?
This is not the way.

Try clicking the ⋯ next to your doc name or using a keyboard shortcut (

CtrlP

) instead.