My approach to ethical AI use starts with consequences, not moral certainty. I am not claiming that I have access to objective moral facts. I am making a practical claim: AI-supported work should be judged by what it is likely to do once it enters a real workflow.
In my research, RAG, and institutional support work, ethical AI use means reviewing outputs before they shape decisions, guidance, documentation, or public-facing information. The main issue is not whether AI was used. The issue is whether the AI-supported output changes the reliability of the work, hides uncertainty, weakens accountability, or gives someone more confidence than the evidence supports.
A problem with AI ethics is that the criteria can become vague if they are not tied to a clear epistemological position. If I am not grounding the framework in universal moral certainty, then I should not write as if I am. A consequentialist approach fits my work better because it asks what the output is likely to cause in practice.
The guiding question is:
What predictable harms, distortions, or accountability problems could this AI-supported output introduce, and what review steps reduce those risks before the work is used?
Transformation and Intellectual Honesty
A further ethical issue is that AI does not simply preserve an input. It transforms it. Any idea, note, source, student question, policy excerpt, or institutional record that passes through an AI system is changed by the process of summarizing, classifying, rewording, prioritizing, and connecting it to other patterns. Even when the output is useful, it is not identical to the original input.
That matters because small generalizations and inferences can accumulate. A sentence can become slightly broader than the source. A student’s question can be turned into a cleaner category than the student actually expressed. A policy detail can be compressed in a way that removes an exception. A messy human problem can be made to look more settled than it is.
For that reason, ethical AI use requires honesty about transformation. I should not treat an AI-supported output as if it were simply the user’s original thought, the source’s original meaning, or the institution’s verified position. Once AI has shaped the material, the result is a derived product. It may still be valuable, but it carries the responsibility of revision, attribution, source-checking, and human judgment.
This is especially important in research, advising, and institutional support work because the harm often comes from subtle distortion rather than obvious fabrication. The ethical obligation is to preserve the boundary between source material, user input, AI-generated synthesis, and final human-approved output.
Core Ethics Criteria
Before I use, share, or operationalize an AI-supported output, I review it against five criteria.
Accuracy: Does the output represent the source material correctly, or has it overstated, compressed, or distorted what the evidence says?
Evidence boundaries: Does the output separate verified source information from inference, interpretation, assumption, or AI-generated synthesis?
Uncertainty: Does the output identify incomplete, conflicting, outdated, or ambiguous information, or does it make uncertainty look settled?
Accountability: Is there a human reviewer responsible for checking the output before it affects students, staff, research records, advising processes, or institutional decisions?
User autonomy and trust: Does the output help the user understand the basis for the answer, or does it create misplaced confidence by hiding sources, limitations, or decision logic?
RAG-Specific Review Checks
For RAG outputs, I use a more concrete review process before the output reaches a student, colleague, or operational workflow.
I check that the answer is actually grounded in retrieved source material, not just plausible language.
I confirm that the cited source supports the specific claim being made.
I check whether the source is current, authoritative, and appropriate for the question.
I mark any part of the answer that depends on inference rather than direct source support.
I escalate, revise, or withhold the output when the issue involves policy ambiguity, student records, advising consequences, or institutional authority beyond what the source evidence can support.
Source That Shaped My Thinking
Floridi and Cowls helped me name one of the main ethical problems I see in AI-supported work: explicability. They define explicability as both intelligibility and accountability. In other words, ethical AI use requires some way to understand how the system shaped the output and some way to identify who is responsible for the result.
That matters for my work because AI does not simply repeat an input back unchanged. It summarizes, classifies, generalizes, and infers. Those changes may be useful, but they still need to be visible. If an AI-supported answer makes a source, a student question, or an institutional position sound cleaner or more settled than it really is, I need to be able to identify that transformation and take responsibility for correcting it before the output is used.
For me, explicability means that AI-supported work should not hide behind fluent language. The output should make clear what it is based on, where interpretation begins, what remains uncertain, and who is responsible for reviewing it before it guides a decision.
One-Sentence SOP
Because I evaluate AI ethics by foreseeable consequences rather than claims of objective moral certainty, every AI-supported output must be reviewed for accuracy, evidence boundaries, uncertainty, accountability, and user autonomy before it is shared, operationalized, or allowed to influence decisions.