Which AI hallucinates the most?

There isn’t a single AI that **universally hallucinates the most**. The answer depends heavily on **which benchmark, model version, and task** you measure. A model that hallucinates a lot when answering obscure trivia may perform very well at summarizing provided documents, for example. [GitHub - vectara/hallucination-leaderboard: Leaderboard Comparing LLM Performance at Producing Hallucinations when Summarizing Short Documents · GitHub](https://github.com/vectara/hallucination-leaderboard/?utm_source=chatgpt.com) That said, some broad patterns from published evaluations: - **Open-source and smaller models often show higher hallucination rates** than the best large proprietary models, especially when they lack strong grounding or retrieval. [The Range Shrinks, the Threat Remains: Re-evaluating LLM Package Hallucinations on the 2026 Frontier-Model Cohort](https://arxiv.org/abs/2605.17062?utm_source=chatgpt.com) - Among major assistants, results vary: - Some benchmarks have found **Claude-family models among the lowest hallucination rates** on certain factuality tasks. [Which AI Hallucinates the Least? 5 Models Tested (2026)](https://www.talkory.ai/blog/ai-models-lowest-hallucination-rate-2026?utm_source=chatgpt.com) - Other benchmarks show different rankings because they test different things (for example, document summarization vs. open-ended knowledge questions). [GitHub - vectara/hallucination-leaderboard: Leaderboard Comparing LLM Performance at Producing Hallucinations when Summarizing Short Documents · GitHub](https://github.com/vectara/hallucination-leaderboard/?utm_source=chatgpt.com) - Models from OpenAI, Google, Anthropic, Meta, and others have all produced hallucinations; none has eliminated the problem. [Why AI still gets things wrong, even when it sounds confident](https://www.financialexpress.com/life/why-ai-still-gets-things-wrong-even-when-it-sounds-confident-4334130/?utm_source=chatgpt.com) If you mean **“which popular chatbot is most likely to confidently make things up in normal conversation?”**, the practical answer is: it’s usually the model that is most willing to answer without saying “I don’t know.” Hallucination risk rises when you ask about: - obscure facts, - recent events, - citations/papers, - legal or medical details, - niche technical APIs, - nonexistent people or products. A good rule of thumb: **the more specific and verifiable the claim, the more you should ask the AI to provide sources or show uncertainty.** [Why AI still gets things wrong, even when it sounds confident](https://www.financialexpress.com/life/why-ai-still-gets-things-wrong-even-when-it-sounds-confident-4334130/?utm_source=chatgpt.com) If you want, I can also give a **ranking of ChatGPT vs Claude vs Gemini vs Grok specifically for hallucination risk**.

**No single AI model “hallucinates the most” across all tasks, but smaller, less fine‑tuned models (like early GPT‑3 variants or open‑source LLaMA‑style models) tend to hallucinate more than larger, retrieval‑augmented systems such as GPT‑4 or Claude 3.5.** In general, hallucination frequency depends on model size, training data quality, and whether external knowledge retrieval is used.[[2510.06265] Large Language Models Hallucination: A Comprehensive Survey](https://arxiv.org/abs/2510.06265)[Large language models hallucination: A comprehensive survey - ScienceDirect](https://www.sciencedirect.com/science/article/abs/pii/S157401372600078X)[A Comprehensive Survey of Hallucination in Large Language Models: Causes, Detection, and Mitigation](https://arxiv.org/html/2510.06265v1) ## 🔍 Key Comparison of Hallucination Tendencies Model / Family| Hallucination Risk| Why It Happens| Mitigation Strategies ---|---|---|--- **GPT‑3(175B)**| High| Trained on broad web data without retrieval; strong fluency but weak factual grounding| Prompt engineering, external fact‑checking **Open‑source LLaMA / Falcon (smaller variants)**| Very High| Less training data, weaker fine‑tuning, limited safety layers| Fine‑tuning with curated datasets, RAG integration **GPT‑4**| Lower| Larger scale, better alignment, often paired with retrieval tools| Retrieval‑augmented generation (RAG), chain‑of‑thought prompting **Claude 3.5 (Anthropic)**| Lower| Emphasis on constitutional AI alignment reduces unsupported claims| Rule‑based alignment, refusal policies **Multimodal models (e.g., text‑to‑image diffusion)**| Moderate–High| Visual hallucinations (objects that don’t exist, distorted text) due to distribution biases| Better datasets, multimodal grounding **Domain‑specific fine‑tuned models (medical/legal)**| Variable| If fine‑tuned carefully, hallucinations drop; if not, risk is severe due to domain sensitivity| Expert‑curated corpora, strict evaluation benchmarks ## ⚠️ Important Considerations * **Hallucination is not random**: It often arises when the model faces gaps in training data or is asked for highly specific, niche, or time‑sensitive facts.[Understanding Hallucinations in Large Visual and Language Models](https://dl.acm.org/doi/epdf/10.1145/3811409) * **Bigger ≠ perfect**: Even GPT‑4 or Claude can hallucinate, especially when asked about obscure or recent events without retrieval support. * **Evaluation benchmarks**: Researchers use structured tests (e.g., TruthfulQA, HallucinationBench) to quantify hallucination rates, but results vary by task and domain.[Hallucinations in generative artificial intelligence and large language models: tests, datasets, detection and correction methods | Language Resources and Evaluation | Springer Nature Link](https://link.springer.com/article/10.1007/s10579-026-09938-4) ## 🧭 How to Reduce Hallucinations in Practice 1. **Use retrieval‑augmented generation (RAG)** – models that pull from external sources before answering hallucinate less. 2. **Ask for sources** – forcing citation or evidence reduces unsupported claims. 3. **Cross‑check with multiple models** – ensemble or self‑consistency methods catch contradictions. 4. **Be cautious in high‑stakes domains** – medical, legal, or financial queries should always be verified with human experts. 👉 In short: **smaller, open‑source LLMs hallucinate the most; large, aligned, retrieval‑augmented models hallucinate the least, but none are immune.** Would you like me to show you recent benchmark scores comparing GPT‑4, Claude, and LLaMA side‑by‑side so you can see actual hallucination rates?

Determining which AI hallucinates the "most" depends heavily on how a test is structured - specifically whether a model is forced to guess on obscure knowledge or given access to real-time search grounding. Independent evaluations by benchmarking platforms like _Artificial Analysis_ and various comparative industry studies highlight a few key patterns: * **The Open-Ended Knowledge Trap:** When pushed on hard knowledge or obscure trivia without search tools (such as on deep knowledge and omniscience benchmarks), many frontier models from major providers (including variants of OpenAI's GPT series, DeepSeek, and Google Gemini) exhibit high rates of confident error, sometimes hallucinating in over 80% to 90% of instances where they lack definitive internal knowledge rather than simply refusing to answer. * **The Role of Search Grounding:** AI platforms that heavily rely on live web retrieval and strict citation constraints - like Perplexity or search-augmented chatbots - tend to show lower raw hallucination rates in day-to-day queries. However, tests show that when these search-reliant systems are forced to generate purely original text without source material, or when underlying search APIs pull flawed data, their error rates spike significantly. * **Model Family Tendencies:** Anthropic's Claude models frequently score relatively lower on hallucination indexes compared to some competitors under specific adversarial or rigorous evaluation conditions, trading a degree of speculative breadth for more conservative answering behavior. Conversely, smaller open-source or specialized models often hallucinate more frequently due to limited parametric memory. Ultimately, **no major commercial AI model is immune to hallucination.** The models that hallucinate most frequently are generally those configured to be "people-pleasers" - meaning they try to answer every prompt rather than admitting uncertainty or declining to answer when data is missing. [Which AI Hallucinates Most? ChatGPT vs Claude vs Perplexity Tested](https://www.hellobuilder.ai/blog/which-ai-hallucinates-most-chatgpt-vs-claude-vs-perplexity-tested#:~:text=The%20Bottom%20Line.%20This%20experiment%2C%20as%20shared,hallucination%20rates%20partly%20by%20avoiding%20original%20thought.)

Determining which AI "hallucinates the most" depends heavily on **how** you test it, **what** kind of prompt you give it, and whether it operates in a closed-context (summarization) or open-domain (creative/factual recall) environment. [Which AI Has the Lowest Hallucination Rate? (2026 Data)](https://www.seekr.com/resource/ai-lowest-hallucination-rate/) Industry benchmarks and evaluation frameworks like the Vectara Hallucination Leaderboard and independent testing provide a clear picture of how error rates stack up: * **Open-Domain and High-Stakes Queries:** When models are pushed outside of strict source documents into complex domains (like legal analysis, niche historical facts, or deep multi-step reasoning), hallucination rates spike dramatically for _all_ major families. Studies in specialized sectors like legal/regulatory querying have shown general-purpose frontier models hallucinating or drifting on upwards of **50% to 88%** of obscure or unknown prompts. [AI Hallucination Rates, Statistics & Benchmarks in 2026](https://suprmind.ai/hub/ai-hallucination-rates-and-benchmarks/) * **Newer Reasoning Models Can Surprise You:** Interestingly, newer "thinking" or reasoning models (such as certain iterations of OpenAI's o-series or advanced reasoning checkpoints) have sometimes displayed higher hallucination frequencies on specific adversarial or open factual datasets (like PersonQA) compared to their simpler predecessors because they tend to over-elaborate or force a narrative when they lack direct data. [Seekr] * **Grounded Summarization:** When tested strictly on sticking to provided text (RAG/summarization benchmarks), top models from major providers (Google's Gemini series, OpenAI's GPT-4o lines, and Anthropic's Claude) manage very low baseline hallucination rates - often under **1% to 3%**. Lower-tier, older open-source, or heavily compressed smaller models consistently sit at the bottom of these leaderboards, showing significantly higher rates of unsupported additions and factual straying (sometimes exceeding 10–20%+ depending on the complexity of the source text). [Who's the Most Delusional? The AI Hallucination Leaderboard](https://www.missioncloud.com/blog/ai-hallucination-leaderboard-desert-mirage) Ultimately, **smaller, fine-tuned, or older open-source models** generally exhibit the highest baseline hallucination rates across standard benchmarks. However, even the most advanced frontier models will hallucinate heavily if asked about topics they don't know or forced into open-ended guessing without retrieval guardrails. [Suprmind +1] Would you like to explore **how to reduce hallucinations** using Retrieval-Augmented Generation (RAG), or are you looking for benchmark comparisons on a **specific model**?

Short answer: There isn’t a single model that “is” the most hallucinogenic across all tasks; hallucination rates vary by model and by what you ask. However, some benchmarks and syntheses suggest that many newer, larger models can still produce false or unfounded content, and the relative rankings shift by task (e.g., summarization vs. open-ended reasoning). What the latest signals indicate - Broad benchmarks show substantial variation in hallucination rates across models and tasks, with some models hallucinating well over 50% on certain hard factual prompts, while others stay much lower depending on the domain and evaluation method. This reflects the general finding that model size alone doesn’t guarantee lower hallucinations and that task framing matters.[AI Hallucination Rates 2026: 22-94% Across 26 Models](https://truestandard.ai/blog/ai-hallucination-rates-2026)[AI Hallucination Rates, Statistics & Benchmarks in 2026](https://suprmind.ai/hub/ai-hallucination-rates-and-benchmarks/) - Individual model comparisons from 2024–2026 often place Claude-family and GPT-family variants near the top in accuracy for many tasks, but still report nontrivial hallucination rates in factual tasks; other models (e.g., Gemini, Grok) have mixed performance depending on prompt type.[Which AI Hallucinates the Least? 5 Models Tested (2026)](https://www.talkory.ai/blog/ai-models-lowest-hallucination-rate-2026)[AI Hallucination Rates, Statistics & Benchmarks in 2026](https://suprmind.ai/hub/ai-hallucination-rates-and-benchmarks/)[AI Hallucination Report 2026: Which AI Hallucinates the Most?](https://www.allaboutai.com/resources/ai-statistics/ai-hallucinations/) - Industry reviews emphasize that hallucinations arise from how models learn patterns rather than retrieve exact facts, and mitigations include better prompting, retrieval-augmented generation, and explicit source grounding.[Why language models hallucinate](https://openai.com/index/why-language-models-hallucinate/)[What are AI Hallucinations?](https://www.databricks.com/blog/ai-hallucinations) Guidance if you’re evaluating for a project - Define the task type you care about (fact extraction, summarization, reasoning) and use a benchmark aligned to that task. Different models excel in different categories, and the “most hallucinating” model can vary by task.[AI Hallucination Rates 2026: 22-94% Across 26 Models](https://truestandard.ai/blog/ai-hallucination-rates-2026)[Why language models hallucinate](https://openai.com/index/why-language-models-hallucinate/) - Consider retrieval-augmented approaches ( grounding against sources ) to reduce hallucinations, especially for high-stakes factual work.[What are AI Hallucinations?](https://www.databricks.com/blog/ai-hallucinations)[Why language models hallucinate](https://openai.com/index/why-language-models-hallucinate/) - Track both accuracy and citation quality: models may produce correct conclusions but with invented sources; verify outputs with trusted references.[Why language models hallucinate](https://openai.com/index/why-language-models-hallucinate/) If you want, I can pull a current, task-specific comparison for a domain you care about (e.g., medical guidelines, legal summaries, historical facts) and summarize which models perform best and where hallucinations tend to appear.