1 of 48

Class 8: The risks of generative AI

A taxonomy of risks
Mitigation strategies
Case study: AI for healthcare

DPI-681M

The Science and Implications of Generative AI

Profs. Goel, Levy, and Svoronos

Harvard Kennedy School

2 of 48

The risks of generative AI

3 of 48

A taxonomy of risks

There are four broad categories of risks from generative AI:

Known limitations�[ e.g., misinformation, biases, toxic content, lack of privacy, environmental harms ]
Misuse�[ e.g., disinformation, aiding in illegal activities ]
Society-wide disruptions�[ e.g., labor markets, social relationships, attention ]
Existential risks�[ e.g., due to intentional or inadvertent damages ]

4 of 48

Limitations and misuse

Large language models learn to produce language by digesting everything on the internet and beyond. That doesn’t always lead to good outcomes!

Social biases and stereotypes
Toxic content
Mis/disinformation
Aiding in illegal activities

5 of 48

Misinformation

6 of 48

Misinformation

7 of 48

Using PingPong, try to get GPT-4 to return incorrect information.

🚴 Activity 🚴

8 of 48

Misuse

10 of 48

Misuse

11 of 48

Fine-tuning LLMs

To mitigate misinformation and misuse, large language models are fine-tuned to align with desired behavior. �[ We align using demonstration and comparison data. ]

13 of 48

Society-wide disruptions

Disputes over intellectual property�[ Should companies be allowed to train on data without compensating creators? ]
Labor displacements�[ What happens if people’s jobs are replaced by AI? ]
Fraying of social relationships�[ What happens when humans are “friends” with AI? ]
Loss of competency�[ What happens if we lose skills, like writing? ]

14 of 48

Society-wide disruptions

15 of 48

Existential risks

Though there is much disagreement on the topic, many researchers and policymakers are concerned with “existential risks” from AI [ i.e., significant loss of life ].

Such risk may stem from malicious individuals using powerful AI to cause damage, or from AI that acts in ways that are not “aligned” with human interests.

16 of 48

“Suppose we have an AI whose only goal is to make as many paper clips as possible.”�[ Nick Bostrom ]

What could go wrong?

17 of 48

“Suppose we have an AI whose only goal is to make as many paper clips as possible.

The AI will realize quickly that it would be much better if there were no humans because humans might decide to switch it off. Because if humans do so, there would be fewer paper clips.

Also, human bodies contain a lot of atoms that could be made into paper clips.

The future that the AI would be trying to gear towards would be one in which there were a lot of paper clips but no humans.”�[ Nick Bostrom ]

18 of 48

Existential risks

AGI is artificial general intelligence

AGI

No AGI

Harm

No harm

Harm

No harm

19 of 48

Generative AI and elections

Thinking back to our discussion of elections in the last class, what are some risks in each of these categories?�

Known limitations: ________________________________________�
Misuse: _____________________________________________________�
Society-wide disruptions: _________________________________�
Existential risks: ___________________________________________�

20 of 48

II. Mitigation strategies

21 of 48

Mitigating misinformation

II. Mitigation strategies

The creation and dissemination of misinformation was cited by a majority of people in the class as an immediate or medium-term risk of generative AI.

22 of 48

Mitigating misinformation

“Comprehensive legislation and attendant regulations that not only prohibits misinformation but requires disclosures and reporting on the use of generative AI in any political campaign or respecting any political content.”

23 of 48

Mitigating misinformation

“Enforce watermark or watermark-like requirements (embedded patterns, metadata) for any LLM provider.”

24 of 48

Mitigating misinformation

“A government-sponsored deep fake detector.”

25 of 48

Mitigating misinformation

“Implementing Robust Fact-Checking and Content Verification Systems. This strategy aims to combat misinformation by developing and integrating advanced verification systems to authenticate AI-generated content before it's widely shared.”

26 of 48

Mitigating misinformation

“Implementing comprehensive digital literacy programs that include education on generative AI capabilities and how to critically assess digital content.”

27 of 48

Regulatory levers

Guha et al. identify four commonly proposed regulatory approaches to mitigating AI risks:

Disclosure. Informing the public
Registration. Informing the government
Licensing. Seeking government approval
Auditing. Verifying compliance with standards�

28 of 48

Regulatory levers

Guha et al. identify four commonly proposed regulatory approaches to mitigating AI risks:

Disclosure. Informing the public�[ Disclosure, education ]
Registration. Informing the government�[ Banning ]
Licensing. Seeking government approval�[ Watermarking ]
Auditing. Verifying compliance with standards�[ Detection, content verification ]

29 of 48

Regulatory levers

In assessing proposed policies, we must consider:

Likelihood to accomplish our goals if implemented, while considering potential unintended consequences for other objectives [ i.e., “regulatory alignment problem” ]
Technical and logistical feasibility
Political support

30 of 48

Mitigating misinformation

How effective, feasible, and supported do you think these proposed strategies are? [ Discuss with your neighbor. ]

Banning deepfakes
Requiring disclosure
Requiring watermarks
Developing detection tools
Creating content verification systems
Educating people

31 of 48