Kagi Update: AI Image Filter for Search Results

Original Article ↗ Hacker News Discussion ↗

Paid search engine Kagi has introduced an AI image filter that downranks results from sites with a high proportion of AI‑generated images, aiming to help users find “real” photos and human-created artwork more easily. Many users praise Kagi’s ad-free, customizable search and see this feature as a valuable response to the growing flood of low-quality or misleading AI imagery, especially for tasks like drawing reference or visual research. Others question its effectiveness and priorities, noting current weaknesses in Kagi’s image search, the difficulty of reliably separating AI from non-AI content at scale, and differing preferences from users who actually want AI images surfaced more prominently.

Overall sentiment on Kagi as a search engine

  • Many commenters are enthusiastic users, calling it the best current alternative to ad-driven search and “like Google from 10 years ago.”
  • Key value: fewer junk/SEO results, no ads, domain blocking/downranking, and customization; some say it’s the only engine that reliably finds obscure technical posts.
  • Others see only marginal improvement over DDG/Google and don’t feel it justifies the ~$10/month subscription, especially for light or casual search use.
  • Some users cancelled due to cost or being between jobs, but miss it and consider resubscribing.
  • A subset worries about all searches being tied to one login and about future shifts in business model.

Local and maps search limitations

  • Several note weak performance for local queries (restaurants, services, maps/routing) and often fall back to Google (!g).
  • There is mention of optional location sharing, but discoverability and granularity (beyond country-level) are unclear and possibly incomplete.

Search sources and ecosystem

  • Kagi is said to aggregate results from multiple providers (e.g., Bing, Brave, Mojeek, possibly Google).
  • Some dislike any association with Brave due to its crypto/ads angle; others frame it as a mere API integration, not a deep partnership.
  • Mojeek’s role in powering some “organic” results is noted and praised by a few.

Image search baseline quality

  • Multiple users say image search is Kagi’s weakest area: poor relevance for specific queries, filters (especially minus-filters) not respected, and mediocre reverse image search for source-finding.
  • Others report substantial improvement over the last year, or find Google Images worse due to indirection and UX.
  • Feedback flow via kagifeedback.org exists but is perceived by some as slow or fragmented; separate login is a minor annoyance.

Reactions to the AI image filter

  • Strong interest from people seeking drawing/photo references and wanting to avoid “AI slop” overwhelming results.
  • The current approach downranks domains with lots of AI imagery rather than analyzing each image; commenters highlight this as a major limitation, especially for mixed UGC sites (Reddit, social media, stock sites).
  • The “baby peacock” example shows that AI images replicated in legitimate articles still slip through; several note this as evidence of how hard cleanup will be.
  • Some consider the feature “dumb” and argue images should be judged purely visually; others counter that for realism, factual reference, or valuing human effort, knowing an image is AI vs real is crucial.
  • There is demand for both modes: some want AI images filtered out; others want them prioritized for licensing ease and as prompt inspiration.
  • False positives/negatives and bugs (e.g., include/exclude behavior appearing inverted in one test) are reported; users ask for feedback mechanisms and even reward schemes for corrections.

Broader concerns about AI content and labeling

  • Commenters worry about AI-generated media as a kind of “non-information spill” that contaminates search results over time.
  • Several advocate for legal or normative requirements that AI-generated images include identifiable metadata or watermarks, arguing it would greatly help filtering with limited downsides, though metadata can be stripped.
  • Others question how long AI vs non-AI will remain reliably detectable at all, given model progress and content remixing.

Kagi workflows and perceived value-add

  • Heavy searchers say Kagi’s clean, ad-free result pages, domain blocking, and “fail to find” behavior significantly reduce time spent searching.
  • Liked extras include “summarize this page,” “Small Web” emphasis, and the new AI image filter as part of a broader effort to downrank low-quality content.
  • Some users, despite appreciating the philosophy, still feel no dramatic productivity improvement and remain unconvinced by the paid model.