How do you stop an AI model from making up details in review summaries?
I created a feature that summarizes app store reviews for my side project. In the version about six out of every twenty summaries included complaints that no one had actually written. My first fix was adding "do not make things up" to the prompt. It changed almost nothing because the model had no idea of what a safe summary looked like. What worked was giving it two examples. One summary had a quote from the review behind each claim and in the other the input was too thin so the correct output was "not enough data." I also added a rule that every claim must have a quote under twelve words. After that seventeen out of twenty summaries were usable on the try. The three failures were expected so I added an example covering reviews written in mixed languages. Telling a model what to avoid gives it little to copy while a finished good answer and one correct refusal give it something to match. What is the smallest change to a prompt that gave you your jump, in accuracy?
p/airuncode
p/runevr
p/producthunt
p/google-cloud-platform
p/aperture-7
p/raycast
p/kodlo
p/ai-context-flow
p/kuc-kuc
p/lasso-11
p/offiai
p/dino-kids
p/many-displays
p/reso
p/supabase
p/jev
p/dex-by-exmergo
p/murror
p/htmldrag
p/nexupira
Replies
One step past quotes that we use: let the model point, and let code copy. We have a model listen to a finished song and say which lyrics were actually sung, since the singer sometimes drops or repeats a line. Free text answers were tried first and came back with lines respelled. Now the planned lines go in numbered, and for each sung line the model answers either a line number, which our code turns into the exact original text, or its own words only where the singing really differs. A quote can still be invented; a number can only point at something that exists. For reviews that could mean numbering them, having every claim cite review numbers, and showing the cited text from your own data rather than from the model's output.