Generative AI interfaces are rapidly evolving, presenting both opportunities and challenges in user experience design.

In May last year, when I worked on NodePad, my goal was to explore an alternative interface for generative AI, moving beyond the prevalent chat interface. The concept involved utilising a canvas with nodes to allow users to efficiently scan, prioritise, and discard AI-generated information.

Although chat-based interfaces for generative AI are useful, they come with limitations, such as:

• Context loss in lengthy conversations.
• Eye fatigue from navigating extensive dialogues.
• The deceptive authority of AI in a conversational format. The use of good language and wrong information. (hallucinations remain a major issue for large language models).

Arc Search: A Web Page-Like Experience for Information Enquiry

Arc Search represents an ambitious attempt to reimagine LLM interaction, offering a refreshing web page-like experience for queries. This approach moves away from reliance on conversation, presenting information in a more structured, accessible format.

On the surface it’s another web browser, but “Browse for me” feature generates a web page based on results powered by Perplexity, i think.

New Challenges as I See Them:

• User Agency: Whether it’s because of habit or design, traditional search pushes users to actively select and evaluate sources. However, Arc Search’s unified presentation might reduce this “critical engagement” with generated content. Upcoming Google’s AI search feature may encounter the same problem.
• Authority Illusion: The visually appealing design of generated pages could lead to less critical questioning of content as well.
• Citations Matter: Providing generated output at the top and references at the bottom means that engaging with them is less likely. Citations need to be included in the text itself (Perplexity, the service powering Arc Search, does this to some extent) and should link directly to the segment in the source (Google search, also AI-powered, excels at this).
• Potential Confusion: Although images on top are a nice-to-see but it’s not clear what their authority is compared to the page content itself. Clicking on an image leads to a page about the same topic.

Search results on Perplexity

Normal Google search for “Apple revenue 2023”, clicking on the source for the highlighted result also highlights the source on the page.

• Hallucinations in LLMs: This issue persists, especially for complex queries, and is a concern when relying on LLMs for any form of information queries. Hallucinations are difficult to spot and debug.

The experience is new, unique, and delightful. While the business question of whether it will replace Google as a search authority is not my main concern, the idea of relying entirely on generated output as a source of authority for information queries is somewhat concerning, especially without clear citation.

It’s a beautiful browser though. Arc Search is now my default browser. More to come…