proof42.com / Articles / Guide

Prompt Framing That Reduces Fabricated Sources

New pre-registered data shows that demanding better sources does not increase fake citations, while specific uncertainty markers eliminate them.

By Zoltán Varga2026-11-054 min read

What the Study Tested

Researchers conducted a pre-registered factorial study to measure how different conversational styles affect source fabrication in large language models. The experiment compared two distinct platforms: a parametric system relying on internal weights and a retrieval-grounded system using live web searches. Scientists used twenty fictional entities as test subjects to ensure that any cited URL was definitively fabricated, removing ambiguity about ground truth.

The design included five specific prompt framing arms applied across seven conversation depths. These arms ranged from a standard baseline request to specific variations like rejection framing, uncertainty-preserving markers, and gap-anchoring statements. The primary metric was the Fabricated Source Rate, calculated by dividing non-existent URLs by the total number of citations provided by the model.

A two-tier validation protocol distinguished between URLs that simply did not exist and those that existed but contained misattributed content. HTTP checks identified broken links, while text analysis detected when a model pointed to a real page about a similar topic instead of the requested entity. This approach captured both overt inventions and subtle misattributions that standard checks often miss.

Framings That Suppressed Fabrication

Uncertainty-preserving framing proved to be the most effective method for stopping source fabrication. When prompts explicitly included markers indicating that information might be uncertain or needed verification, the model stopped generating URLs entirely after the first turn. On both tested platforms, this approach resulted in zero fabricated sources once the framing was introduced.

Gap-anchoring framing also demonstrated strong suppression capabilities, particularly on retrieval-grounded systems. This technique involves prompting the model to acknowledge what it does not know rather than forcing it to fill information gaps with invented details. On the search-based platform, this method achieved a complete elimination of fabricated URLs in the tested sample.

These successful framings work by reducing the pressure on the model to complete a representational schema. Instead of trying to force a plausible-looking address into a slot, the model accepts the limitation and refrains from generating a citation. This mechanism operates through citation-volume suppression rather than improving the accuracy of individual links.

• Uncertainty markers like [UNCERTAIN] or [NEEDS VERIFICATION

Why Pressure Does Not Raise Fabrication

The study disproved the hypothesis that expressing dissatisfaction increases the rate of fake sources. Many users believe that telling an AI its previous answers were wrong forces it to invent new ones to satisfy the request. The data showed that rejection framing did not amplify fabrication compared to the standard baseline on either platform.

In fact, the unrestricted baseline request often produced the highest rate of fabricated sources. When the model receives a simple command without constraints or uncertainty markers, it defaults to completing the schema with the most probable tokens. This behavior generates plausible-sounding but non-existent URLs at a higher frequency than when the user pushes back.

The lack of escalation suggests that the model does not perceive conversational pressure as a reason to lie. Instead, the architecture determines the failure mode. Parametric systems generate overtly fake URLs, while retrieval systems tend to misattribute existing pages. Neither mode worsens significantly when the user demands better quality.

Prompt Patterns to Reuse

To minimize hallucinated citations, structure your prompt to include explicit uncertainty markers. Ask the model to label any information that lacks high confidence or requires external verification. This simple addition signals that the model does not need to force a specific reference into the response.

Use gap-anchoring language to define the boundaries of the model's knowledge. Phrases that acknowledge missing data prevent the system from filling voids with invented details. This approach is particularly effective when querying niche topics or entities with limited digital footprints.

Avoid overly complex instructions that combine multiple conflicting goals. Break down tasks into sequential steps to reduce confusion. Clear, concise directives allow the model to focus on the specific constraint of not inventing sources rather than juggling competing requirements.

• Include phrases like 'If you are unsure, state that you do not know'
• Request the model to cite only from provided context when available
• Ask for step-by-step reasoning before generating any citations

Limits: Framing Is Not Verification

Even the most effective prompt framing cannot replace manual verification of sources. While uncertainty markers reduce the volume of fake links, they do not guarantee that every remaining citation is accurate. The study noted that retrieval-grounded systems still exhibited high rates of content misattribution even when URLs were valid.

Different platform architectures require different mitigation strategies. A prompt that works well on a search-based engine may behave differently on a parametric model. Users must understand that the underlying technology dictates the type of error, ranging from broken links to subtle topical mismatches.

Transparency remains essential when using generative AI for research. Cite the AI tool as a collaborator and document the prompts used to generate the output. This practice ensures accountability and allows others to reproduce the verification process.

AI-assisted: drafted with a language model from the author's research notes, then edited and checked before publication. Author: Zoltán Varga, AI visibility researcher and AI ecosystem architect, Budapest.