Responsible generative AI applications

Concetti di IA generativa

Daniel Tedesco

Data Lead, Google

On the eve of the election

Collection of fake news headlines

Concetti di IA generativa

Types of malicious use

  • Deepfakes
  • Misinformation campaigns
  • AI-enhanced hacking

AI-generated image of pope in puffer jacket

1 Pablo Xavier
Concetti di IA generativa

Detection and prevention

Key usage principles

  • Human-in-the-loop
  • Harm prevention
  • Continuous monitoring

 

Points of Detection and Prevention

A flowchart of points across AI usage: access, prompts, responses, application, and communication and feedback.

Concetti di IA generativa

Access

AI can unintentionally aid criminal groups' non-criminal activities.

  • Avoid supporting malicious groups
  • Know Your Customer (KYC)
    • Verify user identity

An ominous-looking character shushing the view.

Concetti di IA generativa

Prompts and responses

Moderating prompts

  • Similar to website or chat group moderation
  • Jailbreaking prompts can still subvert developer guidelines

Moderating responses

  • Screen or filter responses before showing user
Concetti di IA generativa

Applications

Malicious actors can apply benign responses to illegal or unethical activity.

  • Invisible watermarks can help determine source of content
  • May require law enforcement intervention
Concetti di IA generativa

Communication and feedback

  • Clear usage guidelines
  • Feedback loops
    • User studies and stakeholder roundtables
    • Partner with civil society organizations
    • Feedback opportunities in product

A car driver bucking a seatbelt

Concetti di IA generativa

Let's practice!

Concetti di IA generativa

Preparing Video For Download...