✅ What worked

• Both ChatGPT and mAI provided plausible responses that overlapped with IEG’s previously completed robust SLR on this topic
• ChatGPT provided a quite long and detailed response
• The high-level insights were thematically appropriate

❌ What failed

• Many references were entirely made up (hallucinated) — they sounded like real journal articles, but a web search yielded no results
• Using the ScholarAI plug-in did not fix the issue
• The model could not distinguish between real and fabricated citations

❌ Mistake #1: Treating the LLM as a Junior Analyst Instead of a Component

What they did: Gave the model an end-to-end task: “write a literature review.” This is like asking an intern to “build a house” rather than “hammer these nails.”

✅ What they should have done: Break the SLR into modular, verifiable subtasks: (1) text classification, (2) information extraction, (3) summarization, (4) synthesis. The 2025 guidance note states: “For complex use cases such as SLRs, it is important to unpack the various analytical steps.”

❌ Mistake #2: No Ground Truth, No Validation Set, No Testing

What they did: Compared the LLM’s output to a “previously completed robust SLR” but did not establish systematic metrics. They did not split documents into training, validation, and testing sets.

✅ What they should have done: Follow the 2025 protocol of dividing the dataset into four distinct sets: Training (~5 documents for examples), Validation (~20 for prompt refinement), Testing (~30 for unbiased assessment), Prediction (remaining unlabeled). Without this structure, they had no way to measure recall, precision, or F1 scores.

❌ Mistake #3: Blind Trust in Reference Generation

What they did: Asked the model to generate a reference list and were surprised when the references were fake. Even the ScholarAI plug-in did not solve the problem.

✅ What they should have done: Recognize that LLMs are not databases. The 2025 guidance note advises: “Provide ‘unknown’ or ‘not applicable’ as a category… Include a request for references to the source document(s).” Better yet, use retrieval-augmented generation (RAG) and never ask the model to generate citations from memory.

❌ Mistake #4: No Iterative Prompt Refinement

What they did: Ran the prompt once (or a few times) and declared the approach unsuccessful.

✅ What they should have done: Treat prompt development as an iterative scientific process. The 2025 guidance note describes a “prompting and validation loop”: start with basic prompt → apply to validation set → measure performance → refine based on error patterns → repeat until satisfactory → then apply to test set.

❌ Mistake #5: No Human-in-the-Loop Validation

What they did: Treated the LLM’s output as a final product, then discovered hallucinations after the fact.

✅ What they should have done: Design workflows with mandatory human checkpoints. The 2025 guidance note is unambiguous: “The manual review component is mandatory in our workflows when LLM or machine learning are used.”

📊 Text Classification

Accuracy: 0.90 | Recall: 0.75 | Precision: 0.60 | F1: 0.67

“Strong results after multiple iterations to refine the prompt using the validation set.”

📝 Text Summarization

Faithfulness: 0.90 | Relevance: 4.87/5 | Coherence: 4.97/5

“Abstracts generated contained the most important information, often outperforming original abstracts.”

🔍 Information Extraction

Faithfulness: 1.00 | Relevance: 3.25/5

“Model had difficulty extracting the most relevant information… omissions of relevant information were noted.”

🔄 Text Synthesis

Faithfulness: 1.00 | Relevance: 4.20/5 | Coherence: 5.00/5

“No hallucinations took place. However, some relevant information was omitted.”

📌 Modularity is mandatory

Never ask an LLM to do end-to-end what you can break into verifiable components. The 2025 guidance note emphasizes: “It is important to unpack the various analytical steps and to carefully examine for what and how LLMs can be leveraged.”

📊 Validate with metrics

Use recall, precision, F1 for classification; faithfulness, relevance, coherence for generative tasks. Set context-specific thresholds. The 2025 team deemed recall >0.6 and precision >0.7 as “satisfactory” for their SLR context.

👥 Humans must stay in the loop

The guidance note is explicit: “The manual review component is mandatory in our workflows when LLM or machine learning are used.” No exceptions.

🎯 Representative sampling

Use text clustering to identify ~55 representative documents. Include ~5 in training, ~20 in validation, ~30 in testing. This ensures semantic diversity and avoids biased performance metrics.

Why did the 2023 experiment fail so spectacularly?

Because it violated nearly every principle of robust LLM integration: it asked for an end-to-end task without modular breakdown, had no validation sets or metrics, trusted the model to generate references from memory, skipped iterative prompt refinement, and excluded mandatory human review. The 2025 guidance note explicitly warns against this approach.

Did the 2024 experiments completely solve the hallucination problem?

For the tasks tested (classification, summarization, synthesis, extraction), yes—faithfulness scores were 0.90-1.00 with zero hallucinations. However, the relevance score for information extraction (3.25/5) showed that completeness remains a challenge. The 2025 guidance note acknowledges: “Challenges remain in ensuring completeness and relevance in information extraction and text synthesis tasks.”

Can I use LLMs for literature reviews today?

Yes—but only for specific components (classification, summarization, extraction) with rigorous validation protocols. Do NOT ask an LLM to write an end-to-end literature review with citations from memory. Use retrieval-augmented generation (RAG), provide source documents, include “unknown” as a category, and always validate with human review on a representative sample.

What metrics should I use to evaluate LLM performance on literature review tasks?

For classification: accuracy, recall, precision, F1, balanced accuracy. For summarization and synthesis: faithfulness (0/1 scale), relevance (1-5 Likert), coherence (1-5 Likert). Always measure intercoder reliability (Cohen’s kappa) for human annotations. Set context-specific thresholds based on the stakes of your use case.