Part of my job includes reviewing and processing appeals from students who have experienced a change in their financial situation and have requested that their financial aid be recalculated to reflect their current circumstances. The FAFSA uses a family’s prior-prior year tax return to calculate their financial aid eligibility for college. Because it’s looking back so far, it’s common for families to have experienced a change in their financial situation since those tax returns were filed. Let’s consider a common case I see in community colleges: a student was laid off from a decently-paying job, then got a minimum wage part-time job while they return to school for retraining. The student is no longer earning the same income shown on the FAFSA, and their financial aid eligibility should be updated to reflect their current, lower earnings, rather than the higher salary they no longer have. My job as a financial aid administrator isn’t waving a magic wand to make the student eligible for more aid, it’s assessing the student’s change in income, reviewing their documentation, and updating the information on the FAFSA to reflect their current financial situation. This determination is about as high-stakes as it gets, and accordingly, it must be well-documented and stand up to the scrutiny of a federal auditor.
What “auditable by design” actually means in practice
One of the hardest parts of my job is how regulated the world of federal financial aid is. Appeals like this, though, fall into a special category called professional judgment. The federal regulations are intentionally vague while also being specific about my responsibility. I can’t make an adjustment with an end goal in mind (like qualifying a student for the maximum financial aid) or make blanket assumptions about students with similar circumstances. I have to examine a student’s unique circumstances with a consistent approach and articulate my rationale for why those circumstances justify my changes. Much of this nuanced approach is conveyed by my institution’s policies and procedures, but that’s more of a philosophy document, not an operator’s manual.
Much of this nuanced approach is conveyed by my institution’s policies and procedures, but that’s more of a philosophy document, not an operator’s manual.
When my department gets audited at the end of the year and this student’s file is selected for review, I must be able to clearly explain to that auditor exactly why I performed those calculations and adjustments to update the student’s FAFSA. That’s why about six months into my time working on these calculations that I started creating an audit trail document to go along with each appeal I processed, where I would briefly summarize my rationale, the documentation I collected, and the calculations I performed to get to the final data I updated on the student’s FAFSA. While I’m not required to do this, it’s a helpful tool both in case I have to defend one of my decisions months after I’ve made it, and to help me reflect deeply on the analysis I’m performing.
Where AI tools currently fail this standard
When I started more deeply exploring AI tools, something that felt slightly unnerving was that it was impossible to explain how an AI tool returned a specific result, and that the same result could almost never be replicated. Even with a rough understanding of the mathematical foundations of AI tools, we don’t know what data or reasoning is happening behind a model, we only see the output. The kind of audit trail that I use in my professional judgement decisions does not exist for AI tools. That lack of auditability and inconsistency is probably why surveys of AI adoption in higher education have ranked Financial Aid professionals as some of the biggest skeptics of AI with adoption rates significantly lower than other higher education offices (1) (2). When our work is highly regulated and accountability matters, we can’t afford to hand off our decision-making to a mysterious black box.
The special problem of consequential decisions about vulnerable people
There’s another angle worth highlighting that isn’t always captured in discussions of AI adoption in higher education: my work in financial aid directly impacts students. Because I work at a rural community college, these aren’t just any college students, they’re students who are overwhelmingly the first in their family to attend college, who possess fewer resources and have a much greater need for support. If I trust AI for the wrong decision, it doesn’t just impact me, it impacts real students who can’t afford a mistake. A wrong decision can mean a student decides not to enroll, or loses their housing, or drops out. Where some in higher education see flashy predictive analytics or productivity gains from automated decision-making, I see the potential for disproportionate impact on vulnerable populations. AI tools are trained by data that reflects human bias, and research is showing us how that bias can show up in ways that can disproportionately harm the same populations we’re trying to serve. If we can’t see how AI tools make their decisions, we can’t confirm whether they’re applying standards uniformly or discriminating and perpetuating bias.
What auditable by design would actually look like
While the concept of professional judgment is unique to the world of financial aid, the underlying framework I use to approach these cases is directly applicable to working with AI in any high-stakes environment. There are five essential principles of this framework:
01Articulable design logic
Can you explain, in plain language, what the tool is doing and why? Do you lead with this explanation, rather than trying to discover it after something goes wrong?
I lead with the evidence and reasoning principles I’m applying. That reasoning is the thing being evaluated, not just the conclusion.
02Contemporaneous reasoning trails
Does the system capture not just what it decided, but what it considered when making that decision, so that a human can review this reason?
My audit document captures the evidence considered and calculations performed, so that I (or an auditor) can review it.
03Consistency checking
Is there a mechanism to detect when similar cases are being treated differently, and to require an explanation when they are?
Following my policies and procedures and using my audit document template helps ensure that I treat cases similarly. If I stray from this framework, I always document why.
04Reconstructability
Could someone unfamiliar with the system review a specific decision and evaluate whether it was reasonable?
Since my work can always be evaluated by an auditor, I need to explain my work clearly enough for an auditor to clearly follow my decision and judgment.
05Assigned accountability
When the tool is wrong, who is responsible? Is that person empowered to override it, and is the override itself documented?
At the end of the day, I am accountable for all my decisions. Being “audit-ready” means exercising ownership of all decisions.
Each of these rows highlights a principle and the yes/no question about the AI system, answered by the practice that already exists in financial aid. These principles won’t all be applicable to all AI use cases. But they’re a helpful way to think about shifting the conversation from “does AI work” to “can we evaluate whether the AI worked, and for whom.”
The FSA Handbook GPT as a worked example
I built a custom GPT for my staff that allows them to look up regulatory guidance from the Federal Student Aid (FSA) Handbook and other official FSA guidance. When building this tool, I knew that my instructions had to deliberately tell the AI where to find the evidence for its responses and to always provide a linked citation to that evidence. The instructions clearly state that answers must never be fabricated, citations must be verifiable and accurate, and cases of ambiguity must always be flagged by the AI for additional review. I included this because it’s just as important that the AI provides a clear response as it is that it admits when it can’t find an answer. Part of building trust with these tools includes seeing the both a confident, accurate, referenced response and a confident admission that it can’t find a reference that addresses the question. As an added accountability measure, my staff know that before trusting the AI’s response, they should click through to the linked handbook section or regulatory memo to verify that the citation actually supports the claim used by the AI. Because ultimately, my staff (and I) are accountable for ensuring that they’re working from accurate guidance.
In closing
Across higher education institutions, people are asking “what can AI do?” While that’s a useful starting point, higher stakes areas, like Financial Aid, need to start with the more fundamental question of “what would it take to trust this action?” This framework, that AI should be auditable by design, helps us consider this harder question.