Sourcing at the Edges of the Algorithm

Working with large language models day to day means understanding a few unglamorous realities: they have no memory between conversations, no judgment about your questions, and no actual knowledge of a source unless it's sitting in their context window. Framed as an "infinitely patient new coworker" who forgets everything each time, these tools work best when treated as collaborators rather than oracles you query and trust blindly.

A close look at how hallucinations actually happen goes beyond the familiar warning to double-check outputs. A real example: asking ChatGPT and Gemini to summarize a local news article ranking the safest cities in a state produced answers that directly contradicted the article, even when given the URL. The reason traces back to robots.txt files that block AI crawlers like GPTBot from accessing many news sites. The model never actually reads the page; it fills the gap with the next most plausible-sounding answer, which is exactly how hallucinations get generated.

Three free workarounds get real page content into a model's context instead of letting it guess:

• Prepending a Jina URL prefix to any web address converts the page into machine-readable text that can be pasted directly into a prompt or pulled through an API.
• Gemini in Chrome reads the actual text of the open tab (or dozens of tabs at once) rather than trying to fetch and scrape a URL.
• Ask Steve, a free Chrome extension, does the same thing through a simple side panel, feeding on-page content straight into the model.

The material also covers prompt injection as a newer, more dangerous cousin of an old trick called white fonting, where hidden keywords once fooled applicant tracking systems into flagging false matches. The same manipulation is now happening inside AI browsers and matching tools, where invisible or embedded text can quietly redirect a model's output. The practical lesson: know what a tool actually accessed before trusting what it tells you, and keep a healthy skepticism toward any AI-generated summary of a resume or web page that hasn't been fed real, verifiable context.