Risk Aversion Is a Feature, Not a Bug:What That Means for AI in Research Administration

Bert Baumgaertner

There is a tempting story about research administrators and AI that goes like this: RAs worry about adopting a new technology, making them slow to adopt it, and their caution holds the institution back. This story is misguided, and misleading because the diagnosis points in the wrong direction.

Research administration is risk-averse on purpose. The work navigates federal regulations, sponsor-specific terms, institutional policies, and audit trails that survive personnel changes by years. RAs have been trained well, and the training has stuck: when in doubt, slow down; when the rule is unclear, ask; when the answer is unverifiable, do not submit. That posture is not a deficit to be retrained. It is an asset the institution already has, and it should be respected as such.

So the question is not whether RAs need to learn to take more risk with AI. They do not. The question is what needs to be brought to their table by AI vendors, by IT, by supervisors, by sponsors, etc. so that the trained caution can attach to AI tools the way it has long attached to budget templates and award letters.

Connecting to the Four Pillars

The four pillars of AI4RA — security, accuracy, reproducibility, flexibility — were not invented to constrain AI adoption. They were established because RA work already demands all four, and any tool that wants a place in that work has to satisfy them. Each pillar is a way of naming a question the trained RA is going to ask, whether or not anyone tells them to.

• Security: Can sensitive information enter this tool without crossing a boundary it should not cross? If the answer is murky, the trained answer is to wait.
• Accuracy: Are the outputs verifiable against authoritative sources? If verification takes longer than just doing the work, the tool is not actually saving anything.
• Reproducibility: If the same task is performed next month, will the tool produce a comparable result? Audits ask exactly this.
• Flexibility: Does the tool fit alongside existing systems and practices, or does it require everyone else to bend?

The pillars belong to the RA mindset already. The work of meeting them is not the RA’s alone. Security has to be designed in. Accuracy has to be supported by audit-ready outputs and verifiable citations. Reproducibility has to be engineered into the established methodology. Flexibility has to be negotiated with vendors. None of those are individual RA responsibilities. They are institutional and infrastructural ones.

What Is on the RA’s Plate, and What Is Not

There is a useful distinction to be drawn between deliberation and resources.

Deliberation is on the RA’s plate, because nothing else can substitute for it. It is the practice of asking the right questions before adopting a tool, of pressure-testing one’s own intuitions, of recognizing when a confidently presented output is in fact unverifiable.

Resources, both technological and structural, must be granted to RAs. RAs cannot supply themselves with secure environments, vetted tools, training data documentation, or institutional protocols. Those are what AI vendors, IT, supervisors, and institutional leadership have to provide. And the absence of those resources is not a failure of the RA. It is a failure of everyone upstream of the RA.

This distinction between deliberation and resources matters for two related reasons. First, AI does not clearly sit on one side or the other. The engines driving AI systems are large language models (LLMs), which by their very nature “talk” in ways that approximate the way a human might deliberate (sometimes poorly, sometimes better and faster than humans). But with a little technical understanding of how LLMs work, the mystery of how the engines work dissolves: they are incredibly sophisticated auto-completers. As such, no matter how close LLMs come to simulating deliberation, they are qualitatively different from a human agent: a human RA is held accountable, not an LLM.

The second and related reason why this distinction matters is that it changes the conversation. The right question is not how do we get RAs to adopt AI? (As if AI were just like any other tool.) The right question is: what do the rest of us owe RAs so that their caution can be productively applied to AI tools? One way to do this is to run ethics-driven workshops that enable RAs to have a voice through guided conversations. And if you’re wondering whether that means you should seek out someone like a philosopher, the answer is yes. Philosophers are particularly good at providing methodologies and guiding questions to enable portable ways of thinking through hard cases without presuming answers.

RAs Are Not Alone

I’ve attended many other panels in recent months, and there are common threads in what I’ve heard. Earlier this spring at Innovate Idaho, an industry panel was asked what AI vendors and adopters owe to the people inside organizations who will actually use the tools. The panelists converged on two answers: trust, and a safe place to try things out.

Trust, as the panel framed it, is what gets built when the people deploying a tool are visibly accountable for its failures, not just its successes. A safe place to try things out is the operational form of that trust: a sandbox, a pilot, a staging environment in which an RA can run the tool against realistic but non-binding cases, see where it falls down, and develop the calibrated confidence that responsible use requires.

Neither of these is on the RA to provide. They are infrastructural. And note what the panel was not saying: it was not saying encourage RAs to take more risks. It was saying the opposite. Provide the conditions under which trained caution can be exercised without it being mistaken for foot-dragging.

Another panel I attended surfaced a list of ethical concerns that have been circulating between RA organizations: environmental impact, access inequity, training bias, and the illegal sourcing of copyrighted training data. To that list, others on the day added the workload tax — the way AI tools, even when they speed up some tasks, often add silent overhead in verification, prompt-crafting, and judgment calls about disclosure.

These are not abstract concerns. They are concrete concerns that RAs are being asked to deliberate. And they cut in different directions. A tool that is fast may be trained on data the institution would not have approved of. A tool that improves accuracy on average may distribute that improvement unequally across PIs. A tool that saves keystrokes may shift hours into verification. To be fair, no tool will resolve all of these. But the right response is not to pretend the trade-offs are not there.

What deliberation looks like, in practice, is naming the trade-off, asking who pays the cost and who gets the benefit, and not being satisfied with a confident hand-wave. Here is where an ethics-driven workshop shines: it is designed around thinking tools that help bring these trade-offs out into the open. For example, it teaches how analogies are frequently used as arguments, and how to evaluate those arguments. And it presents questions that invite people to think about the analogies they use when they think about and discuss topics like vulnerability, harm, benefit, and reversibility.

You probably noticed that the goal of how AI4RA does ethics is deliberately NOT designed to simply provide answers. Rather, its goal is to support the way we converse and deliberate about the tools and their use. And those goals are further supported by providing resources that communicate information about the infrastructure in more depth.

Trust Is Local Before It Is Global

One pattern that is worth naming directly: trust in AI tools, for many of the RAs we have talked with, is relational before it is technical. If a colleague has used a tool and vouched for it, the opacity of the underlying system matters less. If the team has tried a tool together, surfaced its failure modes, and developed shared norms about when to use it and when not to, that team has something the institution did not have to provide: working knowledge.

This is local trust, and it is real, and it is not a substitute for the global trust that institutional protocols, vendor accountability, and transparent procurement are supposed to provide. But it is the foundation that the global structures rest on. An institution can write the best AI policy in the country and still fail if no team trusts it; an institution can have minimal policy and still operate well if its teams trust each other and the tools they share. Both layers are important. They do different work.

The implication for AI4RA is direct. While the pillars are global, the deliberation is local. Both must hold for AI to serve RAs.

What AI4RA Is Trying to Build

A short principle worth carrying around as a mental model for how RAs think, especially in moments where the pressure to adopt is running ahead of the conditions for adopting well:

If in doubt about AI, I’ll do without AI

This is not a counsel of refusal. It is a reversibility principle, and reversibility is one of the four guiding questions in our ethics workshops for a reason. Doing without is reversible: you can adopt later, when the conditions are right, when the training is in place, when the tool is verified, when the protocol is clear. Adopting under doubt is harder to reverse: outputs propagate, precedents harden, audit trails get muddied. The asymmetry is what makes the principle reasonable rather than reactionary.

The principle is also a check against the false binary that often appears in AI conversations — either we adopt this tool or we fall behind. Real choices almost always have more than two options. Adopt now, adopt later when X is true, adopt for these tasks but not those, pilot in a sandbox first, defer until there is an institutional protocol — these are all live options. Making them explicit is itself part of the deliberation.

The picture this points to is neither RAs need to be braver about AI nor AI is too dangerous for RAs to use. It is, instead, that the work of getting AI right in research administration is distributed work, something that is less flashy than a website or product. Nevertheless, it’s worth taking the time to think through how that distribution should go, especially when it comes to a tool that complicates the deliberation vs resources distinction. This is the less visible work of the AI4RA project, but it is no less important. RAs have been trained well. The job is not to undo that training, but to better resource that training towards the things
worth deliberating over.