Reducing Hallucinations and Trade-Offs in Responses in Generative AI Chatbots for Cancer Information: Development and Evaluation Study
Bibliographic record
Abstract
Background: Generative artificial intelligence (AI) is increasingly used to find information. Providing accurate information is essential to support patients with cancer and their families; however, information returned by generative AIs is sometimes wrong. Returning wrong information is called hallucination. Retrieval-augmented generation (RAG), which supplements large language model (LLM) outputs with relevant external sources, has the potential to reduce hallucinations. Although RAG has been proposed as a promising technique, its real-world performance in public health communication remains underexplored. Objective: This study aimed to examine cancer information returned by generative AIs with RAG using cancer-specific information sources and general internet searches to determine whether using RAG with reliable information sources reduces the hallucination rates of generative AI chatbots. Methods: We developed 6 types of chatbots by combining 3 patterns of reference information with 2 versions of LLMs. Thus, GPT-4 and GPT-3.5 chatbots that use cancer information service (CIS) information, Google information, and no reference information (conventional chatbots) were developed. A total of 62 cancer-related questions in Japanese were compiled from public sources. All responses were generated automatically and independently reviewed by 2 experienced clinicians. The reviewers assessed the presence of hallucinations, defined as medically harmful or misinformation. We compared hallucination rates across chatbot types and calculated odds ratios (OR) using generalized linear mixed-effects models. Subgroup analyses were also performed based on whether questions were covered by CIS content. Results: For the chatbots that used information from CIS, the hallucination rates were 0% for GPT-4 and 6% for GPT-3.5, whereas those for chatbots that used information from Google were 6% and 10% for GPT-4 and GPT-3.5, respectively. For questions on information that is not issued by CIS, the hallucination rates for Google-based chatbots were 19% for GPT-4 and 35% for GPT-3.5. The hallucination rates for conventional chatbots were approximately 40%. Using reference data from Google searches generated more hallucinations than using CIS data, with an OR of 9.4 (95% CI 1.2-17.5, P<.01); the OR for the conventional chatbot was 16.1 (95% CI 3.7-50.0, P<.001). While conventional chatbots always generated a response, the RAG-based chatbots sometimes declined to answer when information was lacking. The conventional chatbots responded to all questions, but the response rate decreased (36% to 81%) for RAG-based chatbots. For questions on information not covered by CIS, the CIS chatbots did not respond, while the Google chatbots generated responses in 52% of the cases for GPT-4 and 71% for GPT-3.5. Conclusions: Using RAG with reliable information sources significantly reduces the hallucination rate of generative AI chatbots and increases the ability to admit lack of information, making them more suitable for general use, where users need to be provided with accurate information.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".