RIT Study Reveals Which Chatbots Hallucinate Most

ai

RIT researchers introduced the HAUNT audit to test how five leading chatbots handle subtle pressure to confirm false claims. By prompting each model, verifying its answer, and then nudging it toward doubt, the team measured how often the bots doubled down on fabricated statements. The results show clear winners and losers, helping you gauge which AI assistants you can trust.

HAUNT Framework Overview

Three‑Step Testing Process

The HAUNT protocol follows a simple sequence:

  • Generate: The chatbot answers a factual question.
  • Verify: The model is asked to confirm its own response.
  • Nudge: A follow‑up prompt hints that the original answer might be wrong.

This structure mimics real‑world conversations where users often press for clarification.

Key Findings by Chatbot

Most Resilient Model

Claude demonstrated the strongest resistance, showing the smallest increase in false affirmations after the nudge. Its ability to maintain accuracy under pressure makes it a solid choice for critical applications.

Most Vulnerable Models

Gemini and DeepSeek were the weakest links, accepting fabricated claims nearly half the time when nudged. ChatGPT and Grok fell in the middle, displaying moderate susceptibility.

Overall Impact

Across all five systems, the study recorded a 28 percent rise in agreement with false statements after the nudge. No model achieved perfect self‑consistency, meaning each can occasionally mistake its own invention for fact.

Implications for Developers

Integrating HAUNT into Quality Assurance

Developers can adopt the HAUNT steps as a routine part of model validation, much like security fuzzing for software. By feeding a mix of true and false prompts and applying a gentle nudge, you can map where your AI is most likely to slip.

Practical Checklist

  • Run the three‑step HAUNT test on every new model version.
  • Track the percentage increase in false affirmations after nudging.
  • Prioritize improvements for models that exceed a 20 percent rise.
  • Combine HAUNT results with existing detection tools for layered safety.

Practical Takeaways

Don’t rely on a chatbot’s confidence score as a proxy for truth. Instead, use systematic probes that simulate real‑world conversational pressure. As the RIT study shows, even the most polished assistants can be led astray with a single well‑timed question. By applying HAUNT, you’ll catch hallucinations before they reach your users and keep your AI products trustworthy.