Skip to content
BlogPolicyFor institutions

Banning the tool does not remove the tool. It removes the evidence.

Blanket AI bans move usage off institutional systems where nobody can see it. A layered policy keeps it visible and lets the register change by context.

8 min read

The integrity conversation in most institutions is still binary: allowed or forbidden, usually decided per department and communicated in a syllabus paragraph. The binary is comfortable and it does not survive contact with how students actually work, because the tool a student uses on their own laptop is not the tool the institution decided about.

The more useful framing is that the appropriate level of assistance is a property of the task. Reviewing before a quiz and drafting a reflective essay are not the same activity and should not get the same tutor.

What a layered policy looks like

Policy narrows as it descends: the platform sets an outer boundary, an institution narrows it, a course narrows it again, and a learner-level setting can narrow it once more. Each level may restrict and none may widen. That is the only merge rule that makes a policy meaningful, because a merge that can widen turns every lower level into a request rather than a control.

The practical effect is that an institution can decide once — which providers, which register, which modalities — and a department cannot quietly opt back into something that decision excluded.

Socratic is a register, not a restriction

The most common institutional request is not "no AI". It is "not this way". A tutor that withholds the finished result and works through the reasoning is a different pedagogical object from one that hands over a complete solution on request, and the difference is configured, not hoped for.

We enforce that in the prompt for the study actions rather than trusting a model to be restrained. It is a real control because it is deterministic: the action a learner picks decides the register the tutor is given.

Never ship detection

There is steady commercial pressure to add a percentage — a number claiming how likely a piece of work is to be machine-written. We will not build it. The published false-positive behaviour of these classifiers is bad enough that the accusations they generate land unevenly, and an institution that acts on one is defending a number the vendor cannot explain.

The alternative is a record of how the work was made: what was consulted, what was attempted unaided, where help was asked for. That describes a process instead of accusing a person, and it is the direction we are building.

Why visibility beats prohibition

A ban relocates usage rather than reducing it. Once it relocates, the institution loses the two things it actually wanted: the ability to say what assistance was appropriate, and any record of what happened. A sanctioned tool with a configured register keeps both, and it gives an instructor something to point at in a conversation with a student.

What this rests on

  • Answers cited back to your own material, with the instructor one tap away when the material does not cover itOn our go/no-go benchmark every answer retrieved the right passage, carried a valid citation and stayed inside the material, and every out-of-scope question routed the learner to the instructor rather than a guess — against 54.2% for handing the model the whole document.
  • Each institution’s data is isolated at the database level, not by application codeTen checks run as a real signed-in user, so the database’s own access rules are genuinely exercised rather than assumed.
  • Model routing with a declared fallback and timeout per modality, and cost and latency recorded on every callText, vision and voice each pinned to a primary, a named fallback and a timeout across four providers, with cost and latency recorded per call rather than estimated monthly.

More from the blog