Introduction to Computational Linguistics and Natural-language Processing (CS 187) – Fall 2023

Module Topic: Uncertainty, Moral Responsibility, and the Precautionary Principle
Module Author: Camila Hernandez Flowerman

Course Level: Upper-level undergraduate
AY: 2023-2024

Course Description: “Natural-language-processing applications are ubiquitous: Alexa can set a reminder, or play a particular song, or provide your local weather if you ask; Google Translate can make documents readable across languages; ChatGPT can be prompted to generate convincingly fluent text, which is often even correct. How do such systems work? This course provides an introduction to the field of computational linguistics, the study of human language using the tools and techniques of computer science, with applications to a variety of natural-language-processing problems such as these. You will work with ideas from linguistics, statistical modeling, machine learning, and neural networks, with emphasis on their application, limitations, and implications. The course is lab- and project-based, primarily in small teams, and culminates in the building and testing of a question-answering system.” (CS 187 Harvard Course Catalog)

Semesters Taught: Fall 2020, Fall 2021, Fall 2023

    Tags

Module Overview

In this module, participants consider decision-making about acting under conditions of uncertainty. The module takes as its starting point the decision by OpenAI not to release GPT-2 to the public, and asks whether that decision (and their subsequent decision to release the GPT) was morally justified. After discussing why utilitarian principles or cost benefit analyses may not resolve this question, participants are introduced to the precautionary principle as a way of thinking about decision-making under conditions of uncertainty. They discuss several variations of the principle and how these variations might apply to other scenarios involving decision-making about actions with potentially large consequences. The precautionary principle is explained in three key pieces: the damage condition, the knowledge condition, and the potential remedy. These pieces of the principle can be adjusted or changed to make the principle more or less permissive with respect to what actions it allows and/or what remedies are prescribed.

    Connection to Course Technical Material

Although the module is fairly broad and does not tie in directly to the technical content covered in the course, the module is straightforwardly connected to the course material more generally, since the course is on natural language processing, and the primary case study used for the module is GPT-2.

Goals

Module Goals

1. Identify the ethically relevant similarities and distinctions between OpenAI’s development of GPT-2 and the development of other historically significant new technologies.
2. Critically evaluate OpenAI’s decision to withhold GPT-2 from the public (and its subsequent reversal) in light of the precautionary principle.
3. Participate in productive discussions with peers about what kind of moral responsibilities different stakeholders have when developing new, high-consequence technologies under conditions of uncertainty.

    Key Philosophical Questions

1. How should we make decisions about potentially high-consequence technologies under conditions of uncertainty?
2. Was OpenAI morally justified in their decision-making with respect to GPT-2?
3. Is the precautionary principle sufficiently action-guiding under conditions of uncertainty?

Materials

    Key Philosophical Concepts

• Uncertainty
• Moral Responsibility
• Precautionary Principle
• Utilitarianism

    Assigned Readings

• OpenAI blog post on GPT-2 decision

Implementation

Class Agenda

1. Introduce OpenAI GPT-2 case.
2. Warmup with think/pair/share activity.
3. Motivation for and introduction to the precautionary principle.
4. Case studies in small lab groups.
5. Larger class discussion

    Sample Class Activity

After students learn about the precautionary principle, they are given a lab packet to work through in small groups. The lab packet contains:

1. Short blurbs with descriptions of 2-3 examples of the development of other new technologies under uncertainty. Examples included in this module were the CERN large hadron collider, the development of recombinant DNA and subsequent Asilomar conference, and heritable genome editing.
2. A worksheet which has two variations of the precautionary principle. One variation is a more permissive principle, and one is a stricture principle. These are achieved by changing the damage condition, knowledge condition, and remedy within the principles.

Students discuss how each principle would apply to each example of a new technology being developed, and whether the principle gives what they think is the right or wrong answer in each case.

    Module Assignment

After the module, students were asked to write a short 500 word essay on the following prompt:

“Propose a version of the precautionary principle — that is, a setting of the three parameters — that you believe gives the “right answers” for the three case studies that we examined in class, and argue for its appropriateness; or provide an argument that no such version of the precautionary principle exists.”

Lessons Learned

Overall the module went really well. Students were clearly very engaged and wanted to continue discussing even when time had run out. One problem was simply running out of time, however. Students only had about 10 minutes to do their small group work on their lab packets, and many weren’t able to finish the worksheet in time. Further, this left only about 5 minutes for large class discussion, which would have tied the module together. In future iterations of the module, time could be saved by omitting utilitarianism as a motivation for the precautionary principle and simply diving right into the precautionary principle.