Best Practices for Responsible AI Use
Artificial intelligence is no longer experimental at Columbia University. It is embedded in teaching, research, and daily operations across the institution. As adoption accelerates, so does the need for shared expectations around how these tools are used.
To address this, Columbia University Information Technology has released University-wide best practices for responsible AI use.The guidance applies to faculty, students, researchers, and staff, and reinforces that while AI can support productivity and discovery, accountability remains with the human using the tool.
The recommendations complement existing guidance from the Provost’s Office and the Center for Teaching and Learning including CTL’s guide on communicating with students about AI.
Why CUIT Issued This Guidance
AI tools can draft text, summarize documents, analyze data, and support decision-making. Used well, they save time and expand what is possible. Used carelessly, they introduce risks related to data exposure, bias, accuracy, and compliance.
CUIT’s guidance is designed to help the Columbia community benefit from AI while avoiding common failure points, particularly when institutional or sensitive data is involved.
At the center of the guidance is one principle:
AI should assist human judgment, not replace it.
Who the Guidance Is For
The best practices apply across three major areas of University activity.
Academics
Faculty and students may use AI in teaching and learning, but expectations vary by course. Instructors are encouraged to clearly communicate what is allowed, and students must continue to follow academic integrity and copyright standards. Detailed academic guidance remains with the Center for Teaching and Learning.
Administration
Staff using AI for operations, communications, or decision support are expected to maintain human oversight, verify outputs, and be transparent about AI use where appropriate.
Research
Researchers using AI for analysis or modeling must follow IRB protocols, data use agreements, and sponsor requirements. Sensitive or identifiable research data should only be used in CUIT-approved secure environments.
Core Principles for Responsible AI Use
Across all roles, CUIT emphasizes a small set of shared responsibilities:
- Protect University data and privacy
- Use only approved tools for sensitive or regulated information
- Verify accuracy and relevance of AI output
- Maintain human review for decisions that affect people
- Watch for bias and inequitable outcomes
- Be transparent about when AI contributes to your work
AI output is not authoritative.
Responsibility does not transfer to the tool.
Administrative Use: Where Risk Is Highest
For administrative workflows, data protection is critical. CUIT advises against entering the following into public or unapproved AI systems:
- Personally identifiable information for students, faculty, or staff
- Health or clinical data
- Financial, legal, or confidential institutional records
- Unpublished research or proprietary materials
Staff should consult Columbia’s Data Classification Standards and use only AI tools that have been reviewed and approved by CUIT.
Accuracy and accountability are equally important. AI-generated content must be reviewed before being used in University communications or decisions, and automated systems should never make final determinations about individuals without human involvement.
Research Use: Integrity and Transparency Matter
In research settings, CUIT places strong emphasis on documentation and reproducibility.
Researchers are encouraged to clearly record how AI tools are used, including models, prompts, parameters, and datasets. Results should be explainable, replicable, and consistent with disciplinary norms and sponsor expectations.
Ethical and legal considerations remain in force. This includes reviewing intellectual property rights, examining datasets for bias, and disclosing AI use in publications or proposals when required.
Academic Use: Clarity Over Assumptions
AI can support learning and productivity, but assumptions create problems.
Faculty are encouraged to clearly state expectations around AI use in their courses. Students must adhere to course policies, academic integrity standards, and copyright law.
For teaching-specific guidance and examples, CUIT directs instructors to the Center for Teaching and Learning’s AI resources.
Using AI Effectively, Not Just Responsibly
CUIT’s guidance goes beyond risk avoidance. Effective AI use requires intention.
Drawing on resources such as Ethan Mollick’s An Opinionated Guide to Using AI Right Now, CUIT encourages users to think of AI as a collaborator rather than a source of truth. Clear goals, strong context, and iterative prompting consistently produce better results than one-off queries.
To support this approach, the Emerging Technologies Collective has published practical articles on:
- Choosing the right AI model for different tasks
- Choosing an API model
- Prompt design and advanced prompt techniques
- Context engineering for more reliable outputs and advanced context engineering
- A growing Columbia Prompt Library for academic use
These resources focus on how to work with AI thoughtfully, not blindly.
Tools, Support, and Reporting
When working with University data, users should rely only on CUIT-vetted AI tools. Requests for new systems must go through CUIT’s procurement and review process.
Support is available through the CUIT Security Office, Research Computing Services, the Service Desk, and the Center for Teaching and Learning.
Users should promptly report suspected data exposure, unapproved AI tools, bias concerns, or harmful AI-generated output. Early reporting allows CUIT to respond quickly and reduce institutional risk.
What This Means in Practice
AI is becoming part of how work gets done at Columbia.
That reality is not changing.
What CUIT is asking is equally clear: use AI deliberately, protect University data, verify what you produce, and stay accountable for the outcomes.
The tool may be new.
The responsibility is not.