AI Isn't Lying About Your Institution. It's Guessing.

Picture the most well-read person you've ever met. Ask them about the French Revolution, photosynthesis, the offside rule, flawless, every time. Now ask them your university's late-submission policy. They won't have a clue. They've never set foot in your registry office.

A sensible person would simply say “I don't know.” But in the latest episode of AI Levels, ICS.AI's Head of Education Sales, Andy Logan, points out that most AI was trained to always be helpful, and to it, helpful means giving an answer every single time. So instead of admitting the gap, it reaches for the nearest plausible-sounding thing, assembled from every late-submission policy it's ever seen anywhere else. Right shape, right tone, wrong institution, delivered with total confidence.

Andy calls this “confidently generic,” and it's more dangerous than being obviously wrong, because nobody double-checks an answer that sounds right. A prospective student applies on the strength of it, gets rejected, and your admissions team inherits a conversation that begins with “well, technically no human ever told you that.”

Here's the twist: this isn't AI's fault. For the last twenty years, your IT and information-security teams have done exactly what good teams should: moved the genuinely useful material onto the intranet, onto SharePoint, behind a login. Policies, procedures, the granular detail of how your institution actually runs. They built a wall, and the wall was the point. It was good practice. Which means the fix isn't making AI cleverer in general; that still won't get it through the door. The fix is building something that's genuinely allowed through, what Andy describes as an organisational language model, a deliberate exception to a wall you built on purpose.

Getting through the door, though, is the starting gun, not the finish line. Plugging a chatbot into SharePoint is an afternoon's work for a half-decent developer. What separates a system that's technically connected from one that's actually trustworthy is everything that happens afterwards: retrieving the right document live at the moment of asking, respecting who's entitled to see what (get that wrong and you haven't built an AI, you've built a data breach with excellent customer service), and the years of testing, guardrails and refinement that teach it to say “I don't know” plainly rather than pad the gap.

This is the part we care about most at ICS.AI. Grounding in sanctioned institutional content, permission-aware retrieval, and humans kept firmly in the loop for the judgement calls are core to how the SMART: platform is built, not a checkbox added at the end.

Andy offers a single question that does most of the work, and we'd happily be asked it ourselves: what happens when the answer isn't in the source material? If the reply is some version of “it does its best,” that's not grounding. That's hallucination wearing a company lanyard. Andy unpacks all of this in Episode 3 of the AI Levels Podcast, including the one question that separates real grounding from an expensive guessing machine.

Listen to the full episode here: AI Levels Episode 3 - Grounding.

---