Artificial Intelligence in Accreditation

Institutions currently use various forms of artificial intelligence (A.I.) in the ongoing business practices of running an institution of higher education. For the generation of material sent to The Commission on Colleges and Universities, institutions might gather, analyze or summarize documents and data using A.I. tools, including generative A.I.. However, using A.I. tools for the preparation of institutional materials for submission to The Commission presents several risks.

This guideline was created to ensure the confidentiality, integrity, security, and veracity of institutional reports and materials submitted, handled, and reviewed by The Commission to ensure fair and equitable outcomes of the review and accreditation process.

Use of generative A.I. in accreditation should be done by users who are familiar with the concepts and constraints of generative A.I. models in accordance with the guidelines of this document. Users should have a high-level understanding of best practices regarding model prompting, the capabilities and pitfalls of large language model (LLM) systems, and familiarity with any guidance provided by the A.I. platform provider describing the best use of the platform.

Accreditation reviews are opportunities for institutions to step back and reflect upon their success in fulfilling their missions. These campus conversations often are leveraged by institutions to foster relationships across functional areas, support transparency and accountability, and promote the professional development of those engaged in the process. Over-reliance upon generative A.I. to compile narrative regarding an institution may limit the value of institutional and peer review discussions regarding each institution’s success story and reduce the benefit of the process for prompting continuous improvement.

Reports submitted to The Commission may contain sensitive information, such as financial details or strategic initiatives, not intended for public disclosure. Submitting these documents to external A.I. platforms could risk breaches of confidentiality, as these platforms may lack adequate security measures, leading to unauthorized access and data leakage. A.I. platforms may also use any uploaded materials including institutional or student data, reports, images, audio, and video for model training, which can result in the integration of private, confidential, personally identifiable information (PII), or proprietary information into A.I. models used globally. Though rare and dependent on the A.I. platform, models may reproduce this information verbatim under certain circumstances to any user of the A.I. platform. Institutions should inquire with their Chief Information Officer whether they have access to locked or enterprise LLMs that are secure for sensitive information. Institutions should also confirm the maximum-security classification of data that they are able to submit even to locked models. Only those platforms that are private, secure from public dissemination, and where transmitted prompts, files, or data are not used for model training should be used. Further, within such platforms, institutions are encouraged to review the data privacy and model training policies to ensure that institutional data is adequately safeguarded.

Institutions may use A.I. tools to summarize, analyze, or produce narratives or reports for submission to The Commission; however, those institutions are attesting to the veracity of those reports when they are submitted. Therefore, we encourage institutions to pay close attention to the outputs of A.I. tools to make sure that they are free from errors (e.g., those of omission, commission, factual, and otherwise). Though increasingly robust, generative A.I. models often suffer from “hallucinations” - instances in which the models invent citations, people, places, or other information which is incorrect or fictional though conveyed in a convincing manner by the LLM system. Additionally, some A.I. platforms have the capacity to do extensive data analysis, producing figures, tables, suggestions for decisions or actions, and conclusions. In these cases, institutions must rigorously assess these model artifacts to ensure that they are accurate, reasonable, and that the institution wholly agrees with the generated analytical products.

Overreliance on generative A.I. models should be avoided particularly in cases where it may suppress human creativity related to innovation and new initiatives. Additionally, generative A.I. systems should not be perceived as a comprehensive, definitive solution to the complex task of generating or reviewing a The Commission report. Such dependency can result in suboptimal outcomes when the LLM output is misaligned with the expectations needed to efficiently meet deadlines or report requirements.

When generative A.I. systems are used for report creation or review, the LLM products must also be scrutinized to ensure that bias present in the model does not contribute distortions to the generated artifacts. For example, passages regarding gender or race may be differentially interpreted or written by the LLM system depending on the training of the model and alignment decisions of the model creators these statements may not accurately represent the values or beliefs of the institution.

The accreditation process relies on the expertise, judgment, and discretion of peer reviewers. The use of A.I. tools by reviewers could compromise the confidentiality of documents submitted to The Commission and could also compromise the quality and integrity of the review, potentially leading to flawed outcomes.

Evaluators are prohibited from saving, copying, uploading, downloading, or distributing these materials through any electronic means including A.I. tools, that would make them accessible to unauthorized individuals outside the designated review process. External review committees should not use A.I. agents in peer review deliberations, including sending/using A.I. assistants for meeting summaries. Similarly, no The Commission employee should use external A.I. tools in the handling of institutional materials submitted for review until a secure and locked A.I. platform is provided for The Commission internal use.