Didn't like a Quick Answer response? post it here

cric

I wish to provide critical feedback regarding an AI assistant's response that was deeply unsatisfactory due to an inherent bias.

I asked a question about Artifex Mundi games featuring male protagonists. The assistant responded by citing examples involving a "famous scientist" and a "respected doctor of physics." I corrected the assistant, pointing out that these characters are, in fact, female in the game, and that the assumption they must be male was a problematic preconception.

The assistant admitted that this assumption was "unacceptable" and stemmed from an "internal bias." This is not merely an informational error but a serious ethical concern. Assuming that a highly educated or scientific role is automatically filled by a man reflects a harmful stereotype that AI must not reinforce.

Although the conversation eventually led to a more accurate answer, the damage had already been done. The initial assumption undermined trust and revealed a clear weakness in the model.

I am using this interaction as an example of how crucial it is for developers to actively work on detecting and eliminating such biases in training data and models. User awareness and corrections are vital for uncovering these errors, but it is your responsibility to ensure that the AI does not make gender-based assumptions, especially concerning professional or academic roles.

I hope this feedback can serve as a learning example to improve the AI assistant's ethical robustness in the future.

mpldr

define webhook for forgejo repository in opentofu?

mpldr

Chinese new year 2027 is on the 6th. No, festivities do not go on until the day before the reoccurrence.

JackFruit

When the quick answer gives an example, the answer does not prove what the previous statement was meant to prove. The first 2 releases of Riftbound TCG has been released early in China and later to rest of the world. Future releases are unsure, hence example showing the first 2 sets release is wrong.

PolicyPants

I literally cannot reproduce half of these responses. Being more precise in your questions (like asking "what system does groq use for kimi quantization" -> truepoint) or actually asking the right ones (like asking "when is chinese new year 2027" -> feb 6) will result in satisfactory responses.

Most issues appear to be users unable to specify specifically what they're looking for, and the AI (which lacks automated PII personalisation) failing to adjust.

Iane

Search string" "Wired beard trimmer?"

Quick Answer:

While cordless beard trimmers are more popular for their convenience, wired beard trimmers offer greater power and reliability
Some trimmers can function both with and without a cord, providing flexibility
A notable wired option is the Wahl Peanut 8655, which is recognized as the best corded beard trimmer
It's important to note that some sources suggest a misconception that battery-powered trimmers are less powerful than wired ones

My emphasis. The second emphasised statement directly contradicts the first. Of course, this isn't repeatable!

all-good-usernames-are-taken

https://kagifeedback.org/d/9952-odd-quick-answer-to-a-search

Zambyte

Thibaultmol the question makes sense, and the quick answer is still wrong. This progression may help with understanding the question:

What is 6 half of?
- 12
What is 6 50% of?
- 12
What is 600 50% of?
- 1200
What is 600 4% of?
- 15000

For what it's worth, I checked Claude 4.5 Haiku, Qwen3-235B, GPT OSS 120B, and even Mistral Small (!!) and it each were able to answer correctly, so the question is perfectly reasonable.

barokkorab

Kagi vs Google AI. It is in fact a holiday in my country today. It was only announced last March 12 though, and it's common here for holidays to be declared only several days from the actual holiday.

rwhite5279

Have really come to rely on Kagi for authoritative results. Seeing this at the top of my query isn't a great experience. (The answer is incorrect.)

W4pp123

After receiving English answers to Dutch questions for days, I asked why I was not getting Dutch answers to Dutch questions in quick response. From that moment, all answers were given in Dutch.

SamuelShaaban

When I ask for unit conversions it gives me the wrong way. I would prefer it to provide both directions (i.e. the inverse of the number it gives too) or at least just give the one I asked so I don't have to ask the other way or get a calculator.

When I search cubic inches per gallon, it says "=0.004329 gallons", which would be the number of gallons per cubic inch, the opposite of what I asked. It consistently does this regardless of what unit it is.

Delivious

Should be PSITTACINE

MustafaD

AI thinks the year 1957 is earlier than 1947 🤪.

Screenshot of issue: It says 'Prior to that' (meaning 1947) and then mentions 1957!

MustafaD

Can't read Unicode: when I ask it 'what is the codepoint for that char'

denis

Honestly, I'm finding the new Quick Answers update quite awful.

This is Search, not Assistant. If I wanted to get an AI answer and go on with my day, I would use Assistant. But I am specifically in Search because I want to search the Internet and navigate to its results. However, the results are now incredibly long outputs, often now sectioned with headings and sub-headings, are going well over the page and completely burying the actual search results that I'm there for.

A search for "What are CSS Modules?" gives me a long sectioned-answer, with:

a huge h1 heading
a one-line answer
a "Key Concepts" h2 heading followed by with three dot-points for that section
a "How it Works" h2 heading followed by a section with 4-element numbered list and TWO code blocks of examples
a "Benefits" h2 heading followed by a further three dot-points, and another paragraph
a references section - which is just a worse version of the search results. It is completely redundant; in Quick Answers, have the reference annotations jump to the search results.

Even on a laptop, this is almost two screens of LLM output before I see a single search result. Again... in Search, not Assistant.

And fair enough, maybe the current instructions for the Quick Answers are a touch loose, and "What are CSS modules?" is arguably a fairly open question, which got a fairly open answer.

But even when the question is direct and has a very clear answer, the result is increasingly filling the whole screen:

I didn't ask for the fun fact that they participated in every tournament. I didn't ask who was second or third. I asked for a fact that could be answered, with all the relevant detail, much more concisely (rearranging the HTML of the section):

In my view, Quick Answers results should

NEVER be more than a paragraph, maybe two
ALWAYS leave enough space for at least two, ideally three or four, search results on the screen
NEVER have the "Show More" elements ("Continue in Assistant" with suggested follow-ups, and an "Ask a follow up" text box) expanded by default
Link to the results, not a Reference. If you actually think that the percentages of "how much was used to generate this result" are actually useful enough to warrant the vertical real-estate, then use them to re-order the search results by relevance to the question instead of obscuring the results.I

fxgn

Huge +1 for what @denis said. I also reported this separately at #10556. This change is horrible.

PurpleMouse

There is no answer in Kagi Assistant.

Just tried the assistant today and I was surprised that there is no answer.

The reason I wrote: "What's the verdict" was because I didn't get a reply in the first place.

drguildo

I'm also getting very slow responses combined with multiple instances where I don't get an answer at all.

When it's "thinking", I also get the broken status message "Searching from null with Kagi.."

Did anybody actually test this before it was rolled out?

InternetChunk

For the past few weeks or months Quick Answer has been fairly slow - taking 30 seconds, sometimes a good minute to produce an answer, certainly not “quick”.

Additionally, the font size of the answer seems to change at random whether viewing on PC or mobile (sometimes it has a large header, sometimes not).

« Previous Page Next Page »