An Ai2 study, as described in a summary based on Hugging Face, tested how reliably large language models can judge the novelty of scientific research ideas. Changing just a few words in the prompt shifted the model’s accuracy at picking the more novel of two ideas by more than 50 percentage points.
In other words, the same idea can score high or low depending on wording alone.
What was reported
The task was a comparison: given two ideas, which one is more novel. The headline result is a gap of more than 50 percentage points in accuracy, caused by a small change in phrasing rather than a different model or different data.
There is a limit to what we can say. The text we had was cut off midway, so we did not see the study’s method, the models tested or the other results. The claim also reaches us through a secondary summary, not directly from Ai2. We treat it as a clear warning sign, not a complete picture.
Why it matters beyond the headline
The study dealt with scientific ideas, but the structure is familiar to any business: a model receives items and is asked to judge, rank or choose. Scoring leads, sorting customer requests and rating content are the same kind of task. The study does not show they fail to the same degree, but it justifies testing them.
The trouble is that this kind of bias is silent. There is no error message, and the answer reads as confident. In a live system, someone tidying a prompt to fix an unrelated problem could move decisions without anyone noticing.
For an organisation or a bank, this is a control and compliance problem. It is hard to claim a decision is consistent and auditable if a light rewording moves it by dozens of points. Whoever builds a system that makes or recommends decisions has to demonstrate stability, not just show good answers on a few examples.
What to do this week
First, map where a language model influences a decision in your business: who gets the first reply, which lead counts as hot, which request is urgent. Those are the places to test.
Second, build a small set of cases a person has already decided. Run each one through several reasonable wordings of the same prompt and compare the results. If they move, that decision is not ready for full automation.
Third, treat the prompt like code: keep versions, and rerun the test set on every change. Fourth, set up human verification, periodic review of a sample, and safety mechanisms before handing critical decisions to an automated system.
A system run this way stays stable even as it changes. We describe that approach in managed automation for business.
Sources
Frequently asked questions
Can I use an LLM to score sales leads?
Yes, as a first-pass sort rather than a final decision. First check that the result doesn't change when you phrase the prompt in different ways, and keep a person in the loop for the leads that matter most.
Does a newer model fix sensitivity to prompt wording?
You can't assume so. The summary we saw doesn't say which models were tested, so there is no basis to call any model immune. The safe route is to run the wording test on the model you actually use.