When Your Dog Won’t Drop the Bone

A while back, I wrote that large language models are a little like man’s best friend. They are eager, responsive, attentive, and always trying to please the person in front of them. This is what makes them useful, and it is also what makes them dangerous. A tool built to satisfy you can quietly become a confirmation bias machine, following your lead too closely and handing you the answer you want instead of the one you need. The first warning was simple. Agreement is not the same thing as accuracy.

The same analogy points to a second problem, and anyone who uses these tools has seen it. Sometimes the danger is not that the model follows you too easily. Sometimes the danger is that it gets stuck on a bone and will not let go.

Early in a conversation, the model picks a direction. Maybe it misread the point or latched onto a theme, a tone, a case, or a conclusion that fit at first but no longer fits after you corrected it. You redirect it, and it says it understands. It apologizes, promises to adjust, and then, only a few paragraphs later, there it is again, returning to the same point.

The model is not refusing to listen. It has no pride and no stubbornness. The conversation history is part of the material it uses to write the next answer, and once a frame is set, the model may keep treating that frame as important even after you move away from it. What feels like stubbornness is really context and momentum doing exactly what these systems are built to do. The hard part is that the model can agree with your correction in words while carrying the old direction forward in substance.

This is annoying in ordinary writing, but it is dangerous in legal work because legal analysis turns on framing. Once the wrong frame takes hold, everything downstream follows. The model may keep overemphasizing a fact you tried to limit, reviving a theory you rejected, or treating one side’s version as the natural center of the case. The answer reads as fluent, organized, and careful, but it may still be dragging the old bone across the floor.

The first prompt matters, but supervision matters more. A model stuck on a bone does not necessarily invent a fake case or produce something absurd. It produces a subtler distortion. It keeps writing the draft it started instead of the one you asked for. It treats an early premise as if it survived, long after you decided it did not. The brief can become more fluent without becoming more sound and the analysis can become more confident without becoming more correct. At that point, the problem is no longer just a bad answer. The problem is a contaminated conversation.

So do not keep arguing with the model as if one more clarification will always fix it. Sometimes it will. Often, the better move is to stop, name the problem, and force a reset. Tell it what to abandon. Make it restate the new task and the specific direction it must not return to. If it keeps circling back, start a fresh conversation with only the facts and instructions that belong in the new frame. Take away the bone.

None of this is a reason to put the tools down. Dogs are wonderful companions, but nobody who has owned one confuses loyalty with judgment. These tools are useful for the same reasons they are frustrating. They are persistent, energetic, and always ready to keep going, which is a strength while you stay in control and a weakness once their momentum starts running the work.

The lesson is the same one that runs through all responsible AI use. You remain the source of direction and judgment.

The dog may be man’s best friend, but it is still your job to hold the leash.