Factual Research — Accurate Forecasts — No Fabrications
Sample Transcript
What are the typical rates of hallucinations or fabrications delivered by the most recent AI chatbots?
Based on the most recent research data from 2024-2026, modern AI chatbots continue to exhibit significant hallucination rates, though with some improvements in newer versions and variations across different models and applications.
According to 2026 research, Gemini leads with hallucinations up to 38% of the time, while ChatGPT makes up information around 35% of the time. However, other studies from this period show different ranges, with hallucination percentages varying from 17-35% across different testing methodologies, indicating that measurement approaches significantly impact reported rates.
The most recent systematic analysis found that GPT-4 demonstrated a 28.6% hallucination rate, while GPT-3.5 showed 39.6%, suggesting some improvement in newer model versions. Notably, Google's Bard exhibited an extremely high 91.4% hallucination rate in systematic review tasks, making it particularly unreliable for research-based queries.
Despite advances in AI technology, fabricated citations remain a critical issue in 2024-2026. Recent medical AI research confirms that "when ChatGPT and other AI chatbots were challenged to authenticate their outputs with references, they generated multiple citations with detailed bibliographic data that seemed perfectly authentic, but most were actually fabricated, contained at least one falsified citation data, or were completely erroneous when verified through medical literature resources."
A 2026 comprehensive analysis found that over 60% of AI-generated citations were either broken or completely fabricated, with these citations appearing professional and using real-sounding publication names that were often indistinguishable from valid references. This represents a persistent problem across model generations.
Recent research reveals significant variation based on application context. In cancer information queries using retrieval-augmented generation (RAG) systems, modern chatbots showed response generation rates of 52% for GPT-4 and 71% for GPT-3.5 when asked about information not covered in their knowledge base, often including fabricated content to fill knowledge gaps.
Vectara's updated research estimates that current chatbots fabricate information at least 3 percent of the time and sometimes as much as 27 percent, with the variation largely dependent on query complexity and domain specificity.
A particularly concerning 2026 study found that leading AI models, including GPT-4o and Gemini 1.5 Pro, could be manipulated to produce dangerously false medical advice. These systems generated claims that sunscreen causes skin cancer or linked 5G to infertility, accompanied by convincing but entirely fabricated citations from reputable journals like The Lancet.
The most significant finding in recent research is the stark difference between specialized and general-purpose systems. Academic-focused AI tools like Elicit and SciSpace demonstrated negligible hallucination rates, while general chatbots like ChatGPT and Bing exhibited critical hallucination levels. This suggests that domain-specific training and reliable source integration dramatically reduce fabrication rates.
Recent studies also documented concerning psychological effects, with mental health outlets reporting users who "withdraw socially, converse compulsively with the chatbot, and begin to hallucinate textual voices when the device is off," raising questions about the broader impact of AI fabrications on vulnerable users.
Beyond text generation, recent investigations revealed that OpenAI's Whisper audio transcription system "invents false content in transcriptions, inserting fabricated words or entire phrases not present in the audio," including attribution of race, violent rhetoric, and nonexistent medical treatments. Despite warnings against use in high-risk domains, over 30,000 medical workers continue using Whisper-powered tools for patient visit transcriptions.
This information is provided for educational purposes only based on current research data. It does not constitute financial, legal, medical, or professional advice. Always consult qualified professionals before making important decisions.