Part 2: Using Generative AI and Ethics Matrices to Select Sustainability Indicators

Using AI as a Structured Analytical Assistant

Part 1 established that sustainability indicators are deliberately constructed representations of system conditions, impacts, risks, and changes. Part 2 provides a method for using generative AI to widen the search for plausible indicators and then using Ethics Matrices A, B, and C to evaluate, compare, and prioritize them.

The AI systems are not being asked to make the final decision. Their role is to generate possibilities, locate publicly available sources, propose operational definitions, and expose alternative ways of framing the project. Your role is to verify the information, compare the models, identify omissions and embedded assumptions, and select the final indicator portfolio.

The Two-Round Workflow

The assignment uses two rounds. The first round is an open search for plausible indicators. The second round uses the ethics matrices to analyze and prioritize a common candidate pool. Separating the rounds helps make the comparison more meaningful.

Step 1: Prepare a Common Project Brief

Begin with a concise description of the final project. Use the same brief with every AI model so that differences in the outputs are not caused by different background information.

The brief should identify:

• the project topic and the decision, technology, policy, program, or intervention being evaluated;
• the geographic boundary and relevant time horizon;
• the project stage or implementation context;
• the principal stakeholders;
• known environmental, social, economic, technical, and governance concerns;
• known constraints, uncertainties, and data limitations;
• the intended use of the final indicator portfolio.

A brief of approximately 150 to 250 words should be sufficient. Avoid asking the AI to invent the project context. State what is known and identify what remains uncertain.

What You May Upload and What You Must Not Upload

You may upload the course-provided Ethics Matrix A, Ethics Matrix B, and Ethics Matrix C files to the Penn State approved AI platform. You may also paste or upload your own non-sensitive project brief, public reports, public datasets, and publicly accessible links.

When in doubt, do not upload the material. Describe the relevant issue in general terms or rely on a public source instead.

Step 2: Open Independent Model Conversations

Use Penn State AI Studio and select two or three different AI models. Where possible, choose models from different model families. Begin a new conversation for each model and keep the conversations independent.

Access: Penn State AI Studio

Use the same project brief and initial prompt with every model. You may ask follow-up questions, but retain a record of consequential follow-up prompts so that the process can be compared and documented.

Step 3: Use Public Web Sources to Ground the Search

The models may search the public web when that capability is available. You may also conduct your own web searches and paste links into the conversation. Useful sources may include government agencies, international organizations, professional standards, peer-reviewed research, public datasets, industry reports, community plans, and established indicator frameworks.

Ask the model to distinguish between an indicator it generated independently and an indicator adapted from a published framework. Require links for factual claims and proposed data sources. A link is a starting point for verification, not proof that the claim is correct.

• Open the cited or linked source and confirm that it exists.
• Check that the source actually supports the proposed definition or method.
• Prefer primary or authoritative sources when available.
• Record the source used to define or justify each final indicator.
• Reject fabricated citations, broken links, or unsupported claims.

Round 1: Generate Plausible Candidate Indicators

Ask each model to propose approximately 10 to 12 candidate indicators. The purpose of this round is breadth. The models should search for plausible measures and operationalize them, but they should not make the final selection.

Review each output for generic language, duplicated indicators, implausible measurement methods, missing stakeholders, and unsupported claims. Do not ask one model to summarize the other models. Preserve the independence of the initial outputs.

Step 4: Create a Common Candidate Pool

Combine the strongest and most distinct candidates from the model outputs into a common pool of approximately 12 to 15 indicators. Merge obvious duplicates, but preserve meaningful differences in definition, boundary, stakeholder focus, or measurement method.

The common pool is important because every model will evaluate the same candidates in Round 2. This makes the ethical comparison clearer. Otherwise, differences in model rankings could simply reflect differences in the lists they originally generated.

Round 2: Upload and Apply Ethics Matrices A, B, and C

Upload the three course-provided matrix files to each model conversation. Instruct the model to read the matrices as analytical frameworks. Then provide the same consolidated candidate pool and the same Round 2 prompt to every model.

Ethics Matrix A: Integrity and Evidence

• Is the indicator clearly and consistently defined?
• Can the proposed evidence or data be verified?
• Is the measurement or assessment method valid and transparent?
• Are assumptions, uncertainty, and limitations disclosed?
• Could the indicator be used or communicated in a misleading way?

Ethics Matrix B: Broader Impacts and Justice

• Which benefits, burdens, risks, and public consequences does the indicator reveal?
• Which stakeholders are represented, and which remain invisible?
• Does the indicator address distributive, procedural, or intergenerational justice?
• Does it make policy, transformation, risk, or precaution visible?
• What broader impact would remain outside the indicator?

Ethics Matrix C: Embedded Choices in Indicator Design

• What boundary, category, proxy, or definition is built into the indicator?
• Which assumptions are required to calculate or interpret it?
• What is excluded or treated as fixed?
• Does the indicator privilege one stakeholder, timescale, or outcome?
• How could a reasonable alternative definition change the result?

The matrices should guide qualitative analysis rather than produce an automatic numerical score. A model may rank or group indicators, but every priority should be accompanied by a reason tied to the matrices and the project context.

Step 5: Compare the Model Analyses

Compare both the discovery outputs and the matrix-based analyses. Agreement is useful, but it is not proof that an indicator is valid. Disagreement may reveal ambiguity in the project, the indicator definition, the ethics matrices, or the models themselves.

Step 6: Select and Refine the Final Portfolio

Select six to eight final indicators. Revise the names, definitions, boundaries, methods, and limitations as needed. The final portfolio should be your synthesis, not a copied list from one model.

The final portfolio should:

• fit the specific project and decision context;
• contain operational indicators rather than broad themes;
• collectively engage Ethics Matrices A, B, and C;
• represent multiple stakeholder interests and system functions;
• address relevant short- and long-term effects;
• make significant tradeoffs, uncertainties, and burdens visible;
• avoid unnecessary duplication;
• use plausible evidence or data sources;
• state important limitations and risks of misuse.

Step 7: Verify Before Submitting

AI output is not evidence. Before using a proposed indicator, verify the factual claims and the source material needed to define or apply it.

• Confirm the definition, unit, formula, or assessment method.
• Confirm that cited frameworks, standards, and datasets exist.
• Check whether the proposed data are actually public and available at the stated scale.
• Distinguish a published threshold from a model-generated suggestion.
• Check whether the indicator measures the intended condition or only a proxy.
• Revise claims that exceed what the evidence can support.

Step 8: Document the AI Process

Retain the following materials for Assignment 7:

• the common project brief;
• the Round 1 and Round 2 base prompts;
• the model names used;
• the initial candidate outputs;
• the consolidated candidate pool;
• the matrix-based analyses;
• consequential follow-up prompts;
• the comparison table;
• the indicators rejected, combined, or revised and the reasons for those decisions;
• the final verified portfolio.

Use conversation export or another clear record available in the approved platform. The purpose is to make the analytical process visible, not to reward the longest transcript.

Responsible-Use Guardrails

Connection to Assignment 7

Assignment 7: AI-Assisted Ethical Selection and Comparison of Sustainability Indicators

The assignment asks you to use two or three AI models, compare their candidate indicators and matrix-based analyses, and produce a justified portfolio of six to eight sustainability indicators for your final project. The complete submission requirements and evaluation criteria are provided in Canvas.

Main Point

Generative AI can widen the search for plausible indicators and make the initial research process more manageable. The ethics matrices provide the criteria needed to examine evidence, broader impacts, justice, boundaries, proxies, assumptions, and omissions. Comparison across models makes variability visible. Verification and final selection remain human responsibilities.