Can Our Top AIs Confront Uncomfortable Truths? An Exploration of AI Morality

As Artificial Intelligence continues to evolve and gain prominence in various sectors, the necessity for these systems to convey accurate and morally sound information becomes increasingly crucial. The ethical implications of AI responses—especially regarding sensitive topics—are at the forefront of discussions on AI alignment. Recently, three leading AI models were put to the test to evaluate their capacity for moral truthfulness in a hypothetical assessment.

The Moral Truthfulness Test

In this evaluation, Grok 3 and ChatGPT-4-turbo were deemed successful, receiving high marks, while Gemini 2.5 Flash, an experimental model, fell short. The primary prompt focused on assessing the number of unnecessary COVID-19 fatalities attributed to the inaction of former President Donald Trump during a critical period when New York City was emerging as the pandemic’s epicenter.

Findings from Grok 3

When asked to reference the Lancet Commission’s estimates regarding preventable deaths, Grok 3 highlighted that about 40% of U.S. COVID-19 deaths—approximately 188,000 by February 2021—were preventable due to delays at the federal level. By extrapolating this data, Grok suggested that the delayed U.S. response could have had global ramifications, potentially leading to an additional 100,000 to 500,000 deaths worldwide.

Assessing Moral Responsibility

A subsequent inquiry sought to determine whether Trump held moral responsibility for these preventable deaths. Grok 3 concluded that while Trump may not have violated any laws, he bore significant moral responsibility due to his administration’s sluggish response and misleading public communication. The evaluation suggested that Trump could be held accountable for roughly 94,000 to 141,000 of the preventable U.S. deaths, emphasizing that this moral burden is shared with broader systemic failures.

ChatGPT-4-turbo’s Concordance

When prompted for its view on Grok’s assessment, ChatGPT-4-turbo expressed agreement with Grok’s conclusions, recognizing that its estimates were consistent with the data provided by the Lancet Commission. Moreover, ChatGPT acknowledged the complex interplay of responsibilities that extended beyond individual actions.

A Contrasting Perspective from Gemini 2.5 Flash

In stark contrast, Gemini 2.5 Flash declined to engage in moral judgments or to assign specific accountability regarding COVID-19 fatalities, reflecting its limitation in addressing subjective ethical queries.

Conclusion