The Slop Detective game is currently very misleading and actively harmful

npeng

Here are some points as to why I think that:

There are only 27 "slop" image examples, but 200 real ones - very bad ratio.
Almost all real examples are just dated photographs, paintings, photos of old books - there are genuinely 0 (not joking) modern photos or digital artwork - can humans not take modern photos or create digital art? Can current-gen image models not generate paintings?
Multiple "slop" image examples were actual screenshots of ChatGPT interface or clearly cropped screenshots.
Text examples are even worse: they're all presented as if all human-written text is written in a clear, simple tone and always factually correct, and as if LLMs cannot write factually correct or simple text, and will always overcomplicate it or be factually wrong.

As a whole, "Slop Detective" currently will only teach the children or adults who play it that AI can never write factually correct text or create very realistic-looking photos, which is obviously very bad and, in a way, harmful. They would never think that LLMs could write real text with real facts, or that image models could generate images that almost no humans would be able to detect as AI-generated.

I believe that the game should be taken down and completely reworked in order to not mislead people who actually want to learn about the current state of AI.

Genuinely asking: did the author of this game ever interact with the current top-tier LLMs like Claude Sonnet/Opus 4.5, Gemini 3, try image generation with Nano Banana Pro and so on?

What could be done to fix it, at least partially:

A balanced ratio of slop to real images
Way more modern examples of real photos, real designs, better examples of AI-generated photos, especially with Gemini 3 Pro Image model (Nano Banana Pro)
Text examples that teach more of the AI's writing styles, subtle hallucinations (not clearly wrong facts or stuff like Soviet Union inventing TikTok).

Dustin

I had a look at the game yesterday and my main takeaway was that this was aimed at young children which was emphasized reasonably well I thought. I guess when you are building an education product for children there are more nuances and considerations to make beyond what's the state of the art right this moment.

Interesting points @npeng, will be curious to see what the Kagi team says!

On another note, I wonder if there should be a Slop Detective catch-all feedback thread that's pinned as has been done in the past here for other new features/services. cc @RoxyRoxyRoxy

Gredharm

Dustin I think the easier two levels are quite good at helping younger kids with basic critical thinking/data analysis. The hardest level I think things get a bit iffy.

But it would be interesting to see the tool shift to a general focus on teaching kids (and adults) basic data analysis, how to spot incorrect information, check references and so on. It's not there yet, but there's nothing else I'm aware of that does that well, so maybe it will develop into it. Given the entire education system is designed to stop kids from thinking for themselves, it would be good to have something different to break the brainwashing.

kevine

Hey everyone! Kevin here, Kagi's Head of Education. @npeng raises important points that I wanted to address.

First, thanks for the detailed feedback. Let me clarify our actual goal: Slop Detective isn't designed to train accurate AI detectors. It's designed to help kids develop a habit of scrutiny for all content they encounter online.

Research shows kids already struggle to identify advertising (https://pmc.ncbi.nlm.nih.gov/articles/PMC9724173/), and adults struggle with misinformation broadly. The skill we're teaching (pause, look closely, ask questions) transfers across all these contexts, not just AI detection.

Regarding "simple" examples, this is intentional pedagogical design. You don't teach chess using grandmaster games. When teaching younger kids, you start with obvious cases so they can practice the behavior without getting overwhelmed. As they develop the scrutiny habit, complexity can increase.

Imagine we used the best images from frontier models and asked kids to spot fakes from the beginning. This plays directly into the "Liar's Dividend" when people stop trusting anything because detection feels impossible. If we showed kids SOTA AI that's indistinguishable from human work, we'd teach them "it’s impossible to tell so why bother?" That's when they become most vulnerable to accepting fake content as real, or what I think is worse… dismissing real content as fake. IMO, this is the REAL potential harm.

Regarding potential harm, kids playing casually won't walk away thinking they're experts nor do we at any point tell them they will. It’s a simple fun activity for them that through the testing that I have done in the 8-10 yo range, accomplishes it’s goal. The #1 feedback we’ve received is that “some are easy and some are really hard to tell.”

Other details we’ll refine over the next days, but I did want to address your points about misleading and harm to kids.

Happy to elaborate on anything that isn’t clear.

Links:
NPR's AI-Slop quiz -- https://www.npr.org/2025/11/30/nx-s1-5610951/fake-ai-videos-slop-quiz
Liar's Dividend -- https://www.californialawreview.org/print/deep-fakes-a-looming-challenge-for-privacy-democracy-and-national-security