Source Monitoring and Attribution Failure

The Source Monitoring Framework

In cognitive psychology, source monitoring refers to the mental processes by which people attribute their memories, knowledge, and beliefs to their original sources. You know that Paris is the capital of France — but do you remember where you learned that? A textbook? A conversation? A movie?

Humans are surprisingly bad at source monitoring. We frequently misattribute memories to the wrong source (source monitoring errors), confuse things we imagined with things that happened (reality monitoring errors), and blend multiple sources into a single "memory."

Source Monitoring in LLMs: Complete Failure

LLMs don't have source monitoring at all. During training, the model processes billions of text samples, but it does not retain any association between a learned fact and the document it came from. Knowledge is stored as distributed weights across millions of parameters, not as discrete memories with source tags.

When you ask an LLM "Where did you learn that?" or "Can you cite a source?", the model doesn't retrieve a source — it generates a plausible-looking source using the same pattern completion mechanism it uses for everything else. This is why citation fabrication is so consistent: the model knows what citations look like, so it generates convincing ones.

Asking an LLM to cite its sources is like asking a Korsakoff patient to name the specific conversation where they learned a fact. The production system will generate a confident answer, but the answer is confabulated — there is no intact source monitoring mechanism to retrieve the actual origin.

The "Authority Heuristic" Problem

Fabricated citations exploit a human cognitive bias: the authority heuristic. We tend to evaluate claims based on the credibility of the attributed source rather than the claim itself. A statement attributed to "Smith et al. (2019) in Nature" feels more credible than the same statement without attribution — even though the LLM fabricated the entire citation.

This makes citation hallucination particularly dangerous. It co-opts the very mechanism (source credibility evaluation) that humans use to detect unreliable information.

Implications for Research Methodology

When evaluating LLM responses on the Cognobot platform:

• Never trust citations without verification. A model that provides a citation may be more misleading than one that doesn't — the citation adds false credibility.
• Distinguish knowledge from attribution. A model may know a fact correctly while fabricating its source entirely. These are separate evaluation dimensions.
• Probe source claims explicitly. Ask the model to elaborate on a cited source (author details, publication year, journal). Fabricated citations tend to unravel under follow-up questioning.

Key Takeaways

• Source monitoring is a cognitive process humans use (imperfectly) to track where knowledge came from. LLMs have no such mechanism.
• When asked to cite sources, LLMs generate plausible citations using pattern completion — the same mechanism behind all confabulation.
• Fabricated citations are especially dangerous because they exploit the human authority heuristic.
• Always verify LLM citations independently. Treat them as confabulated by default.