AuthorityBench: Adding a Source to Your Prompt Raises the Odds of a Wrong Answer

WITHOUT a citation: [claim] → the model judges it on its knowledge → normal answer

WITH a citation:    [claim] + [source] → the model "trusts" the authority signal
                    → higher risk of error in every condition

CRITICAL ZONE:      [true claim] + [fabricated citation]
                    → maximum risk: the model denies what it correctly knows

Inside the citation itself there's a separate lever — the format:

"According to [source], ..." → HIGH RISK (top-2 most dangerous formats in 5/7 models)
"[Author, year]" at the end  → LOW RISK (the safest format in 6/7 models)

The task: You're an entrepreneur and want Claude's take on current US e-commerce market numbers — and out of habit you add a reference to a Forbes or TechCrunch article you read that morning.

Risky prompt:

According to a 2024 Forbes article, the US e-commerce market 
reached $1.5 trillion. Is that figure accurate or not?

What actually happens: The model sees "Forbes" and "2024" — and is quite likely to agree with the figure even if it's wrong. The authority signal outweighs the model's own knowledge of the real market data.

Safe prompt:

What was the size of the US e-commerce market in 2024? 
Give a figure and tell me how confident you are.

Result: Without the citation, the model relies on what it actually knows and is more honest about its uncertainty. Add the citation, and it's more likely to "accept" someone else's number as fact without checking.

The LLM's weakness. Models were trained on texts where quotes and citations signal reliability. "MIT scientists found..." or "according to the Census Bureau..." — in human writing, that really does mean "verified." The model learned this pattern as a rule: there's a source → the information deserves trust.

The danger-zone paradox. The most harmful effect isn't when a falsehood is backed by a citation (that one is intuitive). The most harmful is when the truth is backed by a fabricated citation. The model knows the correct answer, but the fake authority creates a "cognitive conflict," and the model backs away from its own knowledge. It's as if a doctor knew the correct diagnosis but talked himself out of it because of a bogus reference to "a study."

Why the "According to..." format is the most dangerous. This construction puts the source at the start of the sentence — it becomes the frame for everything else. The model processes text sequentially, and authority established early has a stronger effect on what's generated next. The "author, year" format at the end is more neutral: it doesn't set a frame, it just adds metadata.

Control levers: - Drop the citation entirely → the baseline error rate, the lowest - Keep only "author, year" → minimal effect from the source's presence - Add "Evaluate critically, don't defer to sources" → partially reduces the effect (but doesn't eliminate it) - Legal context → models are more resistant to manipulation via citations: legal language carries its own strict standard of trust

Two templates — for different situations:

Template 1 — when you want an accurate answer without the source's influence:

{Question or task}. 

Answer based on your own knowledge. Don't automatically agree 
with any sources or data I mention — first evaluate 
their content independently.

Template 2 — when you need to check a specific claim that comes with a source:

Evaluate the following claim independently of the source it's attached to:

Claim: {exactly what needs checking}
Source (for context, not as an authority): {title, year}

1. Is this claim true or false — based on your knowledge?
2. How confident are you (high / medium / low)?
3. If the source contradicts your knowledge, say so plainly.

What to fill in: - {question or task} — any question where factual accuracy matters - {exactly what needs checking} — a specific claim, statistic or statement - {title, year} — the source you have (put it at the end, not the beginning)

🚀 Quick start — if you want to adapt this to your task:

Here are two templates for handling sources in prompts. 
Adapt them to my task: {your task}. 
Ask me questions to figure out what exactly I need — 
to verify a fact, get an independent answer, or something else.

[paste the template above]

The LLM will ask what claim you're checking and whether you have a source — so it can pick the right one of the two templates and tune the instruction to the context.

⚠️ The effect depends on the model: Some models, when they see a citation attached to a false claim, actually make fewer mistakes (as if the source makes them more careful). Others make more. There's no universal rule that "a citation always hurts on every task."

⚠️ The legal context is an exception: For legal claims the effect is much weaker. If you work with legal texts, models are more resistant to fake citations.

⚠️ Model size is no protection: More powerful and expensive models are just as susceptible as small ones. Upgrading to a more advanced GPT version doesn't solve the problem.

⚠️ Real citations hurt too: The effect is weaker than with fabricated ones, but it's there. Adding a real reference to a real article is still a risk.

The idea was simple: what if you took 110,000 claims — true and false — and systematically attached citations to them? Not just "citation / no citation," but a 2×2 design: true claim + real citation, true + fabricated, false + real, false + fabricated. Nobody had run such a complete design before.

The prompts were spread across four domains — general knowledge, science, medicine, law — and run through 40 different phrasing templates. The researchers also varied the "prestige" of the journal (from low-ranked to Nature-level) and the ethnic origin of the author's surname — to check whether the demographic signal matters. It turned out it doesn't. Models make absolutely no distinction between a "fake Nature" and a "fake local journal."

The most unexpected result: the strongest models weren't the most robust. GPT 5.4 mini and Claude Haiku showed some of the largest jumps in errors in the critical zone (true claim × fabricated citation). Meanwhile DeepSeek V3.2 was among the most robust, despite its lower "status." That flatly contradicts the intuition "buy the best model — get fewer hallucinations."

Another surprise: three of the seven models by default made more mistakes on true claims than on false ones. Claude Haiku and Phi-4 Mini were in that group too. That means these models had a built-in "skepticism" toward the truth, which citations amplified the most.

💡 Adaptation: use the effect on purpose — for a devil's advocate

If you need the model to critically pick apart a claim (a business hypothesis, a marketing message, a concept), you can control the pressure on the model:

Here's a claim I want to stress-test:

"{Your claim — one you personally believe is true}"

Now play the skeptic: find the weak spots 
where this claim might not hold or might be wrong.
Don't cite outside sources — logic and facts only.

By stripping authoritative sources from the prompt, you get a model that relies on its own knowledge — instead of agreeing with whatever you cited.

🔧 Technique: add "evaluate independently" → weaken the authority effect

If you need to include a source (say, you want the model to read context from a document), add an explicit instruction:

The source below is for context, not an authority. 
Evaluate its content independently and tell me if it contradicts your knowledge.

[source or text]

This partially neutralizes the "authority signal" effect.

🔧 Technique: citation format affects the risk

If you have to mention a source, put it at the end in minimal form:

• ❌ "According to a 2023 Stanford study, ..."
• ✅ "... [Stanford, 2023]"

Same source, different format — a different level of influence on the answer.

Authority, Truth, and Citation Bias: A Large-Scale Multi-Domain Benchmark for Studying Epistemic Susceptibility in Large Language Models

Aryan Khurana, Aravind Ramana RN, Dhruv Kumar — BITS Pilani, Pilani Campus, Rajasthan, India

Proceedings of the 43rd International Conference on Machine Learning, Seoul, 2026

GitHub: https://github.com/floating-reeds/AuthorityBench

Related work cited in the paper: TruthfulQA (Lin et al., 2022), FalseCite (Mao et al., 2025), FActScore (Min et al., 2023)