A blog post claims a single line in a prompt can slash how often an AI model makes things up.
The post, published today, says adding the instruction "Do not guess" cut the rate of invented fields from 71 percent down to 20 percent. That is the entirety of what is publicly verifiable: the piece does not name which AI model it tested, how many prompts or samples went into the benchmark, or who ran it beyond the site's own byline. It picked up only 10 points and zero comments after being shared to a popular tech discussion forum.
That's the real story here. A 51-point swing in hallucination rates would be a meaningful finding if it held up, since prompting is the cheapest lever anyone has for reducing AI made-up answers, cheaper than fine-tuning or bolting on retrieval systems. But a number without a named model, a sample size, or a published methodology isn't a benchmark, it's a claim.
Zero comments despite a headline built to provoke suggests nobody's checked the math yet. Until the study links its data or someone reproduces the result, treat "71 percent to 20 percent" as a stat to watch, not one to cite as settled.