AI summaries feel shallower than searching yourself — would adding clickable sources fix that, or is the shortcut itself the problem?

This explores whether the shallower learning people get from AI-written summaries comes from losing access to sources, which links could restore, or from skipping the work of putting the sources together themselves. The starting point is a large finding: across seven experiments with more than 10,000 people, those who learned a topic through ChatGPT said they learned less and felt less ownership of what they knew. They also gave advice that independent raters judged sparser than advice from people who used web search Does learning from AI summaries produce shallower knowledge than web search?. The corpus doesn't directly test whether adding links fixes this. Read across the collection, though, it points to an answer: the depth seems to come from the effort of finding and combining sources, not from the sources being within reach. A link you never click restores nothing.

The clearest analogy comes from how machines read long documents. ReadAgent compresses a document into short 'gist memories' and then goes back to the original text only when a question needs a detail Can LLMs read long documents like humans do?. MiA-RAG works the same way: it summarizes a document first and uses that overview as a map to decide where to look Can building a document map first improve retrieval over long texts?. In both systems, the summary is valuable because it is reliably followed by a lookup. It works as a table of contents, not a replacement. For human readers, a summary with links is the same setup, but nothing forces the lookup step. The design question is not whether links are present. It is whether the summary leaves the reader with a reason to go back to the sources.

There's also a less obvious risk: links can make a summary seem more trustworthy without making it more accurate. AI judges reliably give higher scores to answers that include fake references and polished formatting, whether or not the content holds up Can LLM judges be fooled by fake credentials and formatting?. People likely share some of that bias. A citation can act as a badge of authority rather than an invitation to dig deeper. This matters more because the strongest models now tend to distort documents quietly rather than visibly dropping content, so their errors look like plausible text Does model capability change how documents degrade?. Links are most valuable as a way to check the summary, but the same polish that makes frontier output convincing makes readers less likely to check.

Another lateral lesson: what a summary keeps depends on what it was written for. Summarizers trained toward a specific downstream goal produce denser, more targeted output than ones written for general fluency Can reinforcement learning align summarization with ranking goals?. Users also strongly prefer rich, web-page-style answers to plain chat replies Do full web pages beat markdown chat for LLM responses?. But preferring a format is not the same as learning more from it, and nothing in the corpus shows that a nicer format closes the learning gap.

The takeaway: links are necessary but probably not enough. The corpus suggests a summary restores depth when it works like a map that points to unresolved questions in the sources, rather than a finished answer with footnotes attached. Whether people actually click through, and whether clicking brings back the sense of ownership that web search users reported, remains an open question the collection doesn't yet answer.