12,445 papers · continuously updated · last export: 10 Aug 2026livingmeta.ai
← All living reviews
Living review · live

Hallucination Detection and Mitigation in AI-Generated Scientific Content

1,041 papers in scope, 50 synthesised · updated 28 June 2026 · continuously refreshed

AI-generated living evidence synthesis — not a peer-reviewed systematic review. Methods below.

What changed

Baseline established — 1,041 papers in scope, 50 synthesised, 6 open contradictions identified.

State of the field

The corpus is large but heavily weighted toward narrative and conceptual reviews plus single-domain empirical audits, with relatively few rigorous, comparable experiments. A substantial cluster of empirical studies establishes that citation/reference fabrication is the most measured and best-documented form of hallucination in AI-generated scientific content, with measured rates spanning roughly 11% to 59% depending on model, prompt, and domain [Sarkar, 2026][Misra, 2026][Naser, 2026][Choudhary, 2026][Cabezas-Clavijo, 2026][Sidhu, 2026][Ovcharov, 2026]. Large-scale bibliometric evidence indicates these fabrications have entered the published literature at scale, with a conservative estimate of ~146,932 hallucinated citations in 2025 alone [Zhao, 2026]. A parallel conceptual literature converges on taxonomies of hallucination types (factual, citation, interpretive, contextual) and recurrent root causes in next-token probabilistic generation, training-data gaps, and lack of grounded reasoning [W7115903291][Gaur, 2026][Samuel, 2026][Mohammadabadi, 2026][Ranganathan, 2026][Sreeraj, 2026]. On mitigation, retrieval-augmented generation (RAG) is the single most frequently endorsed strategy, often combined with post-generation verification, knowledge graphs, uncertainty/calibration methods, and human oversight [Mzili, 2026][Yang, 2026][Molli, 2026][Guan, 2024][Huang, 2026], but multiple sources stress RAG is not a complete solution [Yang, 2026][Magesh, 2025][Magesh, 2024]. A newer, more technical strand proposes and benchmarks concrete detection systems—retrieval-grounded citation verifiers, graph-consistency checks, neuron-level localization, rejection sampling, and model-agnostic risk scoring [Khajavi, 2026][Lin, 2026][Ovcharov, 2026][Chen, 2026][Shen, 2026][(28817), 2026][Tunga, 2026][Misra, 2026]. Cross-cutting framing extends to governance, research-misconduct classification, and human-in-the-loop verification labor [Resnik, 2026][Sethuraman, 2023][Blanchard, 2026][Manav, 2026]. Overall the field is descriptively rich and rapidly growing, but suffers from benchmark fragmentation, inconsistent definitions, and few head-to-head mitigation comparisons [Mzili, 2026][Thompson, 2026][Zhan, 2026].

Key findings

  • Citation fabrication is pervasive and rate varies enormously by model and prompt: empirical audits report ~11-22% [Sarkar, 2026], ~12-21% across GPT-4o/Claude/LLaMA [Misra, 2026], and an up-to-fivefold cross-model spread of 11.4% (GPT-5-mini) to 56.8% (haiku-4.5) [Naser, 2026], with one medical study finding 58.5% of 1,100 ChatGPT citations fabricated [Choudhary, 2026].
  • Hallucinated citations are not random but structured, patterned recombinations of real authors, journals, and keywords, with duplication in nearly 30% of cases [Spennemann, 2026] and 100% of sampled NeurIPS fabrications exhibiting compound deception (66% total fabrications invented wholesale) [Ansari, 2026].
  • Citation generation may be an induced rather than intrinsic behavior: no model spontaneously produced formal citations when unprompted (0 of 3,030 responses), and prompts for 'recent and influential' references yielded higher fabrication (74.1%) than 'seminal' ones (55.0%) [Naser, 2026].
  • Multi-model consensus is a useful signal: agreement among three or more LLMs sharply raises the probability that a citation is real [Naser, 2026].
  • RAG reduces but does not eliminate hallucination—commercial RAG-based legal tools still hallucinate 17-33% despite 'hallucination-free' marketing claims [Magesh, 2025][Magesh, 2024], and RAG failures arise across query formulation, retrieval, evidence aggregation, and grounding [Yang, 2026].
  • Hybrid and retrieval-grounded detection pipelines perform strongly: a database+fuzzy+LLM pipeline reached ~80% precision (15-20% over database-only) [Misra, 2026]; CiteCheck reached 88.7 macro-F1 outperforming GPT/Claude/Gemini baselines [Khajavi, 2026]; a Citation-Grounding legal graph fine-tune reached 98.5% validation accuracy [Ovcharov, 2026].
  • Hybrid mitigation stacks outperform single methods: an RLHF+RAG hybrid achieved 91.5% accuracy versus individual methods, against benchmark hallucination rates of 18.7-34.2% [Molli, 2026]; benchmarking shows retrieval methods improve grounding at latency cost while prompt-based methods are lightweight but less robust [Zhan, 2026].
  • Hallucination is increasingly framed as a family of failures across model, pipeline, and human-governance levels rather than a single phenomenon, requiring layered mitigation [Mzili, 2026][Huang, 2026][Lin, 2026].
  • Hallucination signals are field-specific: author-name fields fail most, and probes trained on one bibliographic field transfer near-chance to others, yet targeted neuron suppression reduces fabrication without external retrieval [Chen, 2026].
  • Human oversight is repeatedly identified as indispensable, with citations potentially constituting research misconduct when they function as data [Resnik, 2026], human-based factuality assessment recommended before publication [Dierickx, 2023], and 'Verify-and-Validate' / hybrid human-AI workflows proposed [Jaiswal, 2026][Nakken, 2026][Blanchard, 2026].
  • Model rankings on citation accuracy differ markedly across domains: DeepSeek achieved 92.0% accuracy versus ChatGPT's 19.4% in glaucoma [Civelekler, 2026], while ChatGPT and Perplexity showed significantly lower hallucination severity than Gemini and DeepSeek in dental trauma [Manav, 2026].

Contradictions

Open gaps

Methods & scope

Inclusion: cosine ≥ 0.5 to the agenda-item embedding, relevance ≥ 6. 1,041 papers match that rule. Each synthesis pass reads at most 50 of them, ranked by extraction depth then similarity; this review currently rests on 50. Membership is frozen (same query each cycle); the synthesis is regenerated incrementally and fully rebaselined periodically.

Synthesis by Claude Opus 4.8; membership gating by Claude Haiku.

Revision history

  1. v128 June 2026baseline
    baseline 1,041 papers, 50 synthesised
← All living reviews