Can GenAI replace GSRAs?

Seth Guikema

Civil & Environmental Engineering

Industrial & Operations Engineering

University of Michigan

What ChatGPT thinks I look like

How do you use GenAI in your research?

Please download and install the Slido app on all computers you use

ⓘ Start presenting to display the poll results on this slide.

What single words come to mind when asked “How does widespred reliance on GenAI for research make you feel?"

Please download and install the Slido app on all computers you use

ⓘ Start presenting to display the poll results on this slide.

What is a LLM – as told by a LLM (ChatGPT)

Prompt 1: “prepare a set of powerpoint slides that explain what a large language model is. Use graphics. There should be at least 10 slides. The explanation should be giving a college student level. The slides should look appealing.”

Result: 6 slides with one sentence per slide. No graphics. Terrible formatting. Very basic information.

Prompt 2: “You did not follow my instructions. The slides were supposed to use graphics and look appealing.”

Result: 10 slides. 3 graphics. Still one bullet point per page. Still very basic information.

​

LLMs as Described by a LLM

• An artificial intelligence model trained on vast amounts of text data to understand and generate human-like language.
• They use neural networks, specifically transformer architectures, to analyze text patterns, predict words, and generate coherent responses.
• Neural networks mimic human brain functions. Transformers, like GPT and BERT, improve how AI understands context and meaning in language.
• LLMs are trained on massive datasets, learning statistical relationships between words and phrases to generate meaningful text.
• LLMs can produce biased, incorrect, or misleading information. Ethical AI usage is crucial.
• LLMs are powerful tools in AI-driven communication, but responsible development and use are essential.

A story of exploring boundaries of what ChatGPT can do

• Colleague and I uploaded an abstract of a paper we were planning to work on – focused on individual vs. collective action for climate change adaptation
• Asked ChatGPT 4.o to do a series of task:
1. Formulate a mathematical model (game theory) representing the described situation
2. Solve for the Nash Equilibrium(s) of the model
3. Create the comparative statics graphics
4. Write a description of the model, the equilibria, and the comparative statics
5. Write a discussion of the results and what they mean for climate adaptation

Assume this experiment worked and we published this paper. Is this ethical?

Please download and install the Slido app on all computers you use

ⓘ Start presenting to display the poll results on this slide.

Model formulation

It did ok, but the model required tweaking

​

​

​

​

​

​

​

Fairly standard game theory model. Not much innovative here.

But it missed a key subscript on x!

Model Solution

• Basic structure is fine – ChatGPT4.o can do basic symbolic math reasonably well
• But that missing index i on the x’s carries through – the solution is not really wrong, just not fully complete

Comparative Statics�(how does changing a model parameter impact equilibrium output?)

Huh? As the effectiveness of collective

prevention measures increases we invest

in them more – and more in individual

measures? And investments can be negative?

Let’s try that again: “redo the graphs of x and y vs. alpha but restrict y to be non-negative”

LLMs do not think. You have to make all constraints very specific, or you may get nonsense answers.

Your turn

• Put the abstract on the next slide into your favorite LLM
• Ask it to formulate the problem, then solve the problem if it does not automatically do that

This article presents an approach that a manager can use to allocate resources needed to design a system among the members of a concurrent design team. The system being designed is assumed to be composed of a number of subsystems, each designed by a different engineer. These engineers possess private information about the performance of their subsystem as a function of the design resources that they are allocated. This article shows how the manager of such a concurrent design project can induce rational self-interested engineers to reveal truthfully their private subsystem performance functions. This is accomplished through an incentive contract that ties each engineer’s pay to the contribution of their subsystem to the performance of the overall system. The approach builds from a Vickrey–Clarke–Groves mechanism to achieve, as an ex-post Nash equilibrium, truthful reporting of private subsystem performance functions by risk-neutral agents.

Please download and install the Slido app on all computers you use

ⓘ Start presenting to display the poll results on this slide.

Describe what you think of this as an approach for conducting research. Use single word responses, but put in as many as you would like.

Please download and install the Slido app on all computers you use

ⓘ Start presenting to display the poll results on this slide.

Ask your LLM for references used to solve the problem. Did it find this one: Guikema, S. D. (2006). Incentive compatible resource allocation in concurrent design. Engineering Optimization, 38(2), 209-226.

Please download and install the Slido app on all computers you use

ⓘ Start presenting to display the poll results on this slide.

Practical Issues with current generation of LLMs

• LLMs lack originality in “thinking”
• LLMs responses lack depth
• LLMs hallucinate – sometimes – and it can be hard to tell when it’s happening
• Each GPT4.o query is reportedly (by GPT4.o) accounting for 4-7 grams of CO₂ released. A google search is reportedly 0.2-7 grams.

Ethical Issues

• What does it mean to author a paper if the problem formulation, solution, and analysis of the solution are done by a LLM?
• Should the LLM be listed a co-author? In what circumstances?
• Who is responsible if there is a mistake in the output from a LLM, and this output is used in the paper?
• What should journal policies on the use of LLMs be?
• Can a LLM replace a GSRA for mathematical analysis?

What does it mean to author a paper if the problem formulation, solution, and analysis of the solution are done by a LLM?

Please download and install the Slido app on all computers you use

ⓘ Start presenting to display the poll results on this slide.

Should the LLM be listed a co-author? In what circumstances?

Please download and install the Slido app on all computers you use

ⓘ Start presenting to display the poll results on this slide.

Who is responsible if there is a mistake in the output from a LLM, and this output is used in the paper?

Please download and install the Slido app on all computers you use

ⓘ Start presenting to display the poll results on this slide.

What should journal policies on the use of LLMs be?

Please download and install the Slido app on all computers you use

ⓘ Start presenting to display the poll results on this slide.

Can a LLM replace a GSRA for mathematical analysis?

Please download and install the Slido app on all computers you use

ⓘ Start presenting to display the poll results on this slide.

My Take

• No, LLMs cannot replace PhD students.
• LLMs can, however, supplement them by taking over some of the basic analysis tasks.
• But then how do we train PhD students in this new reality?

Open Discussion Questions