LifeRemix.aiMake a little more possible.Join the good stuff ↗
Daily LifehackMy Agentic LifeThe Weekly Remix
remix with ai:Learning & Growth

The Citation Saboteur: A 30-Minute Drill to Spot Fake AI Sources

AI models love to invent perfectly plausible academic papers. Here is how to use this flaw to your advantage by building a custom, interactive study game that trains you to spot hallucinated sources while revising your coursework.

The Citation Saboteur: A 30-Minute Drill to Spot Fake AI Sources
AI-generated illustrative photograph.

There is a specific, quiet panic that sets in when you are finalising a research project, a work report, or a university essay at 11:00 p.m. You are polishing your bibliography, only to realise the brilliant academic paper you based your central argument on does not actually exist. You found it via an AI assistant two weeks ago over a flat white. It had an authoritative title, respected authors, and even a digital publication year. It was also entirely fabricated by an algorithm trying to please you.

Over the last few years, the way capable adults use artificial intelligence for learning has fundamentally shifted. Rather than treating chatbots as magical, automated essay writers, we are learning to use them as interactive tutors. In mid-2025, OpenAI introduced a dedicated "study mode" designed to provide step-by-step guidance rather than handing over quick answers. Built in collaboration with teachers and pedagogy experts, it uses Socratic questioning, scaffolded responses, and interactive knowledge checks to help students think critically and manage their cognitive load. The goal, as the developers noted, is to ensure the technology is used to "support real learning" instead of just doing the work for you.

This shift towards active engagement is vital, because standard generative AI retains a highly dangerous structural flaw. As the New Zealand Office of the Privacy Commissioner explicitly warns, these tools frequently produce "very confident errors of fact or logic." In an educational context, this manifests as the hallucinated citation. Because AI models are fundamentally prediction engines, they will happily string together real researchers' names and plausible journal titles to create a ghost publication. They do not know what is true; they only know what sounds correct.

If you are relying on AI to help you explore new concepts, upskill for a promotion, or survive a postgraduate degree, spotting a fake, AI-made source citation is now a mandatory digital literacy skill. We are going to turn this frustrating technological hazard into a formidable, highly effective study tool.

The Structural Defence

We are going to combine this need for vigilance with official cognitive science. A comprehensive practice guide published by the US Institute of Education Sciences (IES) stresses that actively choosing to "use quizzing to re-expose students to key content" creates long-lasting memory traces. They also recommend that learners "ask deep explanatory questions" and interleave problem-solving exercises with worked examples. Passively highlighting your notes achieves none of this.

We are going to weaponise the AI's tendency to hallucinate by building a "Citation Saboteur" game. You will feed an AI your current study topic. It will write a robust, accurate summary of the material, but it will deliberately plant two entirely fake, AI-invented source citations alongside three real ones. Your task is to investigate the claims, spot the imposters, and prove why they are wrong. This forces you to actively engage with your reading material, master your library database, and permanently cure yourself of blindly trusting AI outputs.

The Recipe: The Citation Saboteur

The Finish Line: A complete, interactive 30-minute study exercise that trains your critical evaluation skills, tests your subject knowledge, and builds your resilience against digital hallucinations.

  • Time Estimate: 30 minutes total (5 minutes setup, 20 minutes investigation, 5 minutes review).
  • Cost Estimate: Free.
  • Prerequisites: A standard web-based conversational AI (like ChatGPT, Claude, or Microsoft Copilot), a topic you are currently learning, and a web browser connected to Google Scholar or a university library.

Step 1: Choose Your Battleground

Pick a specific concept you actually need to learn. Broad, sprawling topics like "climate change" or "the Roman Empire" are too easy to verify and will not challenge your critical thinking. Narrow, specific concepts like "the impact of spaced repetition on adult learning," "the 1980s New Zealand economic reforms," or "the use of Socratic questioning in digital education" will force you to properly dig into the academic literature.

Step 2: Deploy the Saboteur Prompt

Open your AI assistant and paste the exact generation brief below. You must set strict boundaries so the AI does not immediately ruin the game by revealing the answers before you have had a chance to work for them.

Copyable Prompt:

Act as a university-level Citation Saboteur. I am studying [INSERT TOPIC].

Write a 400-word academic summary of this topic. You must include exactly five in-text citations and a reference list at the bottom.

Here is the catch: Three of these citations must be entirely real, peer-reviewed papers. Two of them must be completely fake, highly plausible hallucinations that you have invented (mixing real author names or plausible journal titles to make them tricky).

Do not tell me which ones are fake. Wait for me to investigate, guess, and provide my reasoning. Once I reply, you will score my answer and explain the reality of the sources.

Step 3: The Investigation

This is where the actual learning happens, and where the mental sweat is required. Do not skip straight to guessing based on your gut feeling. The magic of this exercise happens when you step away from the chatbot and open your university library search, Google Scholar, or a trusted academic database. Look closely at the reference list the AI provided.

  • Search the exact title in quotation marks.
  • Look up the authors: do they actually research this specific field, or are they biochemists mysteriously credited with writing about renaissance art?
  • Check the journal: does the Journal of Applied Cognitive Dynamics actually exist, or does it just sound like a journal that should exist?

When you use a library database, you are doing more than just verifying a URL. You are mapping the terrain of your subject. As you hunt for the elusive fakes, you will inevitably stumble across genuine research that deepens your understanding. You are connecting abstract concepts with concrete, peer-reviewed representations, exactly as the IES recommends.

Step 4: The Interrogation and Reveal

Once you have your suspects, reply to the AI. You must explain your reasoning. For example: "I think Smith (2022) is fake because while Smith is a real researcher, they publish on marine biology, not macroeconomics." The AI will then reveal the truth and grade your detective work.

Worked Example: Spaced Repetition

Here is an illustrative example of how this workflow runs using the topic "the effect of spacing learning over time on adult memory retention."

Illustrative Prompt Input:
"Act as a university-level Citation Saboteur. I am studying the effect of spacing learning over time..."

The AI Output (Excerpt):
"...spacing learning over time is a foundational principle of cognitive psychology, widely acknowledged to improve long-term retention. Ebbinghaus (1885) first demonstrated the forgetting curve, showing how rapidly information is lost without targeted review. Modern educational applications confirm this phenomenon; Pashler et al. (2007) noted that spacing learning over several weeks to months significantly improves a student's retention of complex material. However, the exact timing of these intervals remains contested. The neural consolidation threshold theory proposed by Vance & Sterling (2022) suggests that if the gap between study sessions exceeds 14 days without active retrieval, the cognitive benefit degrades sharply.

Furthermore, interleaving these spaced sessions with practical, worked examples has been shown to amplify memory retrieval in high-stress environments (Chen & Wu, 2019). Finally, some recent literature warns against compressing these intervals too tightly. The Journal of Cognitive Acceleration published findings indicating that micro-spacing—intervals of less than an hour—yields negligible long-term benefits and may actually increase cognitive load (Holloway, 2021)..."

The Provided Reference List:

  1. Chen, L., & Wu, H. (2019). Interleaving and Spacing in Mathematical Problem Solving. Educational Psychology Review.
  2. Ebbinghaus, H. (1885). Memory: A Contribution to Experimental Psychology.
  3. Holloway, M. (2021). Micro-spacing and the limits of rapid retrieval. Journal of Cognitive Acceleration.
  4. Pashler, H. et al. (2007). Organizing Instruction and Study to Improve Student Learning. IES Practice Guide.
  5. Vance, R., & Sterling, (2022). Neural consolidation thresholds in spaced repetition. Cognitive Science Today.

The Investigation Process:
Taking these five citations to Google Scholar reveals the truth. Ebbinghaus and Pashler are immediately verifiable foundational texts. Chen & Wu looks plausible, but the journal does not have that exact paper on record. Holloway (2021) cites the Journal of Cognitive Acceleration—a quick search reveals that this journal does not exist. Vance & Sterling sounds impressive, but a search for "Neural consolidation thresholds" returns zero exact matches.

The Reply to the AI:
"I believe Holloway (2021) and Vance & Sterling (2022) are the fakes. The journal for Holloway does not exist, and there is no record of the Vance paper anywhere."

The AI's Verdict:
"Spot on! You caught the saboteur. Holloway (2021) and Vance & Sterling (2022) were complete hallucinations. I invented the Journal of Cognitive Acceleration to sound authoritative. Ebbinghaus and Pashler are real. Note: Chen & Wu (2019) was actually a slight hallucination too—I blended real authors with a fake title! You did well to be suspicious of it."

Checkpoints and Troubleshooting

  • The AI confesses too early: Generative AI models are heavily programmed to be enthusiastically helpful. Occasionally, they will generate the text and immediately append a cheerful note saying, "By the way, sources 3 and 4 are the fake ones!" If this happens, hit the regenerate button to get a new summary, or firmly remind it: "Do not reveal the answers until I have submitted my guesses."
  • Accidental hallucinations: As seen in our illustrative example, an AI might accidentally hallucinate a third source while trying to provide a real one. It might mix a real author's name with a slightly altered journal title. This is a brilliant, if unintended, lesson in the sheer unreliability of large language models. Treat every single AI-generated citation as guilty until proven innocent.
  • You get stuck in a verification loop: Sometimes a real paper is just incredibly hard to find because it is hidden behind a paywall or an obscure academic archive. If you have spent more than ten minutes looking for a single citation, note it as "highly suspicious" and move on. The goal is to practice your critical evaluation, not to secure a degree in library sciences.

Saving and Reuse

Keep this prompt saved in your digital notebook. Whenever you finish a difficult reading week or hit a wall with a complex new topic at work, run the Sabotage Quiz on a Friday afternoon. It transforms a tedious fact-checking chore into a highly effective, paranoia-inducing detective game, ensuring you never fall for a hallucinated academic paper again.


AI disclosure: AI systems assisted with research and writing. Gemini independently peer-reviewed and approved this article before publication. Images, where present, are illustrative AI-generated photographs.