Get startedGet started for free

Responsible generative AI applications

1. Responsible generative AI applications

We've discussed two fundamental considerations for responsible AI: bias and ownership. Let's turn now to ethical considerations for usage.

2. On the eve of the election

Generative AI can be used maliciously to manipulate society. Imagine on an election night, a slew of videos appears online of a candidate doing unexpected things: resigning from the election, being offensive, or committing crimes. Even official news organizations could fail to determine which information is false. Amidst the confusion, another candidate could win.

3. Types of malicious use

Several kinds of malicious use exist. First are deepfakes; synthetic media depicting events that never happened. While sometimes benign, as with this viral image of Pope Francis in a puffer jacket, deepfakes can defame, influence people, and undermine truth. Secondly, misinformation campaigns can be created with AI. For instance, AI with the ability to post on social media could create a plethora of deepfake content to manipulate opinions or decisions. Finally, AI can enhance hacking. For example, malicious actors using generative AI can access critical infrastructure with devastating consequences. Responsibly developed but poorly moderated AI can inadvertently be used disastrously.

4. Detection and prevention

To avoid negative outcomes, we should apply some usage principles for AI. First, human-in-the-loop ensures humans review generated output. A human editor might review a generated news article before publication. Second, harm prevention involves proactively limiting malicious behavior. Rejecting prompts associated with harmful content is one way. Finally, generative AI products require regular review and updates when problems are detected or new norms arise. Next, we'll look at implementing these principles across the generative AI product lifecycle: user access controls, prompt acceptance, response moderation, application usage, and feedback.

5. Access

AI can unintentionally aid criminal groups' non-criminal activities. To mitigate this, developers can adopt practices like Know Your Customer, a process to verify user identities at sign-up. This is a first line of defense to prevent fraud and illicit use by ensuring only legitimate entities gain access.

6. Prompts and responses

Developers should monitor both prompts and responses for abuse. Calls to violence and hate speech are blocked on mainstream web applications and offending users may be banned. Generative AI developers can similarly moderate use. Even then, the complexity of generative AI still allows the creation of malicious content. For instance, users can subvert developer guidelines with jailbreaking prompts. So, developers should have a method to screen generated responses before passing them to users.

7. Applications

Despite these undertakings, a malicious actor can still apply benign responses to illegal or unethical activity. To detect such activity, developers can include invisible signatures on content, called watermarks, that identify content as generated by their AI. But, this is a cat-and-mouse game, as malicious actors find ways to remove the watermarks. Activity outside the direct control of developers like this may even require law enforcement intervention.

8. Communication and feedback

As generative AI evolves, so will best practices for responsibility. Just as car seatbelts became commonplace over time, norms for generative AI will also emerge. To stay abreast of what's right for our use cases, it's important to engage key stakeholders. One way is through clear and accessible usage guidelines. These can be a touchstone for customers, end users, and regulators to manage expectations about products. Second, regular feedback loops can assess how generative AI applications affect end stakeholders. Developers can also conduct roundtables with those stakeholders, partner with civil society to address concerns, and build feedback opportunities directly into the product.

9. Let's practice!

Practice time!