Over the past two weeks we've examined the ability of LLMs to look across highly chaotic, contested and conflicted narrative environments like Gaza to distill down the firehose of global coverage into succinct actionable at-scale country-level summaries. Unfortunately, as with nearly every major contested story we've examined to date, LLM guardrails have proven a major obstacle to the use of LLMs to examine news content in the real world on real content. When we first attempted to summarize the current state of Gaza coverage two weeks ago, we were unable to do so – blocked by guardrail false positives that prevented major commercial models from summarizing any content that related to events in Gaza. Intense public interest in Gaza eventually led OpenAI to relax its Gaza-related guardrails, while Google's Bard still refuses to answer questions about the events and Bing consistently hews to a handful of sources that provide only the most rudimentary information about the conflict.

Unlike the artificial boundaries and carefully constrained contents of benchmark datasets used to evaluate LLMs and their guardrails, real world news content stands alone in its tendency to touch upon the so-called "third rail" topics of the world's societies. Controversial topics, by their very nature, tend to attract outsized news coverage and outsized interest in skewing public opinion and poisoning the public information environment with false information. In turn, this causes the major commercial LLM vendors to strengthen their guardrails around those topics. Unfortunately, today's LLM guardrails are extremely primitive and lack the ability to differentiate between a malicious actors attempting to promulgate falsehoods and a news organization or public interest researcher attempting to distill current narratives. In other words, guardrails are unable to distinguish between "creation" and "summarization." While the latter involves a degree of creation and can be exploited by adversaries to perform guided prohibited ideation and creation, it lies at the heart of the ability of newsrooms, researchers and even government agencies like public health authorities to peer across the chaotic flow of global information to discern meaning.

The end result is that the improper application of guardrails by current LLM vendors prevent them from being able to serve in a critical role where they would offer the greatest public good: distilling down complex conflicting debates and information spaces on breaking and global-scale events into key points with citations out to all of the related coverage.