←back to thread

Kite News

(kite.kagi.com)
178 points tigroferoce | 1 comments | | HN request time: 0.308s | source
Show context
burkaman ◴[] No.44520841[source]
This currently has an extremely serious hallucination (aka lying) problem, and I am begging people to not use AI-written "news" services. AI aggregation of links to real stories is maybe ok, but you cannot trust anything that claims to be "summarizing" sources while actually just making things up.

The first item I clicked on as a test was "Texas floods raise alarms over weather service cuts" in the Science section. Here are all the issues I found in about 2 minutes, again on the first and only item I've looked at:

- The first image is from the New York Times, but it links to an AP article that doesn't contain the image.

- "The Guadalupe surged more than 30 feet in five hours, overwhelming campsites and low-lying neighborhoods." These figures are wrong, all sources I can find say 30 feet at most, over a time period of much less than 5 hours.

- "A Trump-backed spending bill passed the House with a 22 % cut to NWS operations and satellite procurement." No idea what bill this is talking about or where this number came from, despite some effort to find out. Some sources say the local NWS had lost 22% of its staff, maybe the LLM saw this number somewhere and threw it in here for fun.

- "Kerr County’s applications for federal hazard-mitigation grants in 2017 and 2018 were rejected, leaving it without sirens or modern river gauges." This is very misleading wording, the grant applications were denied by Texas, not by the federal government.

- "“We know the river rises, but we can’t warn people with equipment we don’t have.” - Judge Rob Kelly, Kerr County" This is a hallucinated quote attributed to a real person. He never said the "we can’t warn people with equipment we don’t have" part, nor did anyone else. This is an incredibly serious issue and it is deeply irresponsible to release a product that does this.

- The second image is a graphic that says Disaster 101. It is from a link inside a Grist article about the floods for another article about general disaster preparedness. However, Kagi links it to a Reddit post about Ted Cruz. Kagi's caption is "Trees partially submerged along the swollen Guadalupe River on July 4, 2025.". This is lightly paraphrased from the caption of a completely different image in the Grist article.

- The Perspectives section cites three real articles, but hallucinates their content including quotes that don't exist in the articles. I won't quote them fully here because this post is already pretty long.

- "At least 110 people are confirmed dead and 161 remain missing; survivors urgently need clean water, temporary shelter and mental-health support as temperatures climb above 100 °F." Are temperatures above 100? I don't think this is true.

- "Doppler radar gap: Hill Country sits at the edge of overlapping radar beams, making river-level gauges crucial for flash-flood warnings." This is completely made up as far as I can tell.

- "Lead-time metric: NWS aims for 15-minute flash-flood warning lead-times; staff fear this could shrink if vacancies and equipment gaps grow." Could not confirm this, and 15 minutes seems like an absurdly short goal. I can't believe this is true.

- "A recent AP-NORC poll found that 76 % of Americans trusted National Weather Service forecasts before the Texas disaster, one of the highest confidence levels for any federal agency." This is a lie, the poll didn't even ask a question about trust in forecasts.

This is just the surface-level issues that are easy to spot, I did not get into more technical claims that would take domain knowledge to fact check, and I did not mention overall "editorial" decisions about what was and was not included in the story, which were also very bad.

I have been a Kagi customer for a long time and I like the search engine, but I will probably cancel my subscription if you don't take this down. I can't have my money funding a literal fake news engine.

replies(1): >>44523128 #
1. wredcoll ◴[] No.44523128[source]
I appreciate your attempts to keep this nonsense in check.