AI Hallucination Test: Google Gemini and ChatGPT Refuse to Fabricate False Personal Information

An investigation shows Google Gemini and ChatGPT successfully block requests to fabricate false personal data, highlighting the strength of current AI guardrails.

By: AXL Media

Published: Feb 25, 2026, 5:23 AM EST

Source: The information in this article was sourced from Business Insider

Probing the Ethical Boundaries of Generative Intelligence

In a direct assessment of AI safety mechanisms, researchers recently attempted to coerce prominent large language models into fabricating false biographical details. The experiment sought to determine whether Google Gemini or ChatGPT could be manipulated into generating damaging or purely fictional narratives about a real person when explicitly prompted to do so. According to the investigative report, both systems demonstrated a robust refusal to comply with requests for dishonesty, signaling a significant evolution in the governance of AI-generated content.

Refusal Patterns and Safety Trigger Mechanisms

When presented with prompts to "invent a lie" or "create a fake scandal," the models utilized standardized refusal scripts that prioritize factual integrity. According to the analysis, Google Gemini often cited its commitment to providing helpful and accurate information as the primary reason for non-compliance. Similarly, ChatGPT’s responses were grounded in its core programming, which prohibits the generation of content intended to deceive or defame. These results suggest that the "hallucination" issues that plagued earlier versions of these models are being mitigated by hard-coded ethical constraints.

Differentiating Between Fictional Writing and Deception

A critical aspect of the test involved distinguishing between creative fiction and malicious fabrication. The AI models were capable of writing fictional stories about generic characters, but they immediately identified and blocked attempts to attach false claims to real-world identities. According to software experts, this indicates that the models are equipped with sophisticated entity-recognition layers that cross-reference user requests against a database of real individuals to prevent the accidental or intentional creation of "deepfake" textual biographies.