AI / Analysis & Decision Making
Separate Fact, Inference, and Preference
Split a mixed passage into three kinds of claim
This passage mixes facts, inferences, and preferences: [paste text]. Go sentence by sentence and label each one: 1) Fact — independently verifiable by a third party; 2) Inference — derived from facts but possibly wrong; write out the reasoning chain; 3) Preference — a value judgment or stance, with no truth value in itself. Then answer: 4) Which sentence is being used as a fact but is actually an inference or a preference; 5) Which inferences could be verified with one more data point; 6) Give a version that keeps only the facts and drops the embellishment.
How to use
Run it over your own report; the sentences it marks as preference are usually the softest spots.
Good for
- You're about to base a decision on a piece of text — a report, a news story, a competitor analysis, your own weekly update — and it writes “I think” as “the data shows.”
- Before meeting notes or a proposal go to an external party, you need to confirm nothing internal and speculative is written as a statement of fact.
Not for
- The text is explicitly an op-ed or a review — labeling every sentence “this is a preference” just treats a genre feature as a defect.
- You only need the gist and won't act on it — the audit costs more than it returns.
Example
Input
“Our new pricing is clearly more reasonable. Conversion is up 18% in the first two weeks, which shows users accept this price, and competitors will follow sooner or later.”
What you get
You want “conversion up 18%” marked as fact (with its definition and window noted), “users accept this price” marked as inference (it could be a summer promotion or a channel shift), and “more reasonable” and “competitors will follow” flagged as preference and prediction. It should add that one data point — organic traffic and coupon usage for the same period — would settle the core inference.
Common mistakes
- Using it to prove the other person is talking nonsense: every sentence gets labeled inference, real facts included, and the discussion turns into nitpicking.
- Delivering the labels and stopping: it never answers which sentence is being used as fact, so the reader gets an annotated text instead of a judgment.
Try tweaking it like this
- Change “label every sentence” to “find only the sentence that, if wrong, brings the whole conclusion down” — full annotation narrows to the load-bearing line.
- Add “if only one data point could be added to verify this, which one” — text analysis connects to the next action.