LLMs in Emergency Contexts: 55-language text-to-911 study

Sara Court, Lara Downing and Micha Elsner examine the initial deployment stages of an LLM-based machine translation application used as a text-2-911 system, a service the authors say advertised capabilities in 55 languages. The paper, submitted on 29 May 2026 and accepted to ACL Findings 2026 in San Diego, uses that case study to argue for clearer communication and stronger researcher engagement with public audiences.

What did the authors study and conclude?

The paper presents a case study of an LLM-based machine translation application's early deployment in a real-world emergency context and concludes with concrete recommendations and best practices for stakeholders. The authors identify a number of common misconceptions about technologies like these, and urge the research community to play a greater role in articulating findings to the public.

Court, Downing and Elsner frame their analysis around a text-2-911 system that advertised support for 55 languages. They center the paper on how initial deployment choices and public-facing claims interact with user needs in emergencies where it may be difficult to call operators directly. The manuscript argues that deployment-stage communication and evaluation are central to responsible rollout.

How did the paper frame risks and misunderstandings?

The authors document misconceptions and gaps between scientific advances and public interpretation, and they press researchers to help close that gap by explaining results to non-specialist audiences. They identify a set of common misconceptions about LLM-based translation in emergencies and argue these misconceptions affect how such systems are perceived and used.

Their abstract emphasizes the stakes of public messaging and the mismatch between technical progress and operational requirements. The paper also stresses that while research often focuses on "hard" technical problems, "it is often the "easy" ones -- problems for which the latest technology is often unnecessary -- that are most overlooked." That observation anchors several of the recommendations the authors present for developers, deployers and evaluators.

What recommendations and best practices do they offer?

The authors end with a set of concrete recommendations and best practices for stakeholders at every stage of the development and deployment pipeline, from research to public-facing claims. These recommendations aim to reduce miscommunication, surface realistic capability limits, and ensure evaluation aligns with emergency-use requirements.

The paper frames these practices as actionable steps for researchers, system builders and organizations that adopt LLM-based translation tools for emergency contexts. The authors call for clearer articulation of limits, careful evaluation under domain-specific scenarios, and coordination among technical and operational stakeholders.

Why it matters

Deploying machine translation systems into emergency workflows changes who relies on those systems and what counts as acceptable risk. The paper’s focus on an advertised 55-language text-2-911 service highlights how broad language-coverage claims can shape user expectations in high-stakes situations. Court, Downing and Elsner push the field toward accountability in public communication and evaluation, not just technical benchmarks.

Their framing connects two problems: the gap between research findings and public messaging, and the tendency to overlook seemingly "easy" operational issues that determine real-world usefulness. Both matter for emergency responders, vendors, and communities that depend on accurate, reliable translation when human operators are hard to reach.

What to watch

Look for the authors’ presentation at ACL Findings 2026 in San Diego for the full paper and detailed recommendations. The community response to that presentation will indicate whether researchers and practitioners adopt the paper’s call for clearer public articulation and deployment-focused evaluation.

Paper details: title "LLMs in the Real World: Evaluating "AI" in Emergency Contexts," authors Sara Court, Lara Downing and Micha Elsner, arXiv:2607.00019, submitted 29 May 2026, accepted to ACL Findings 2026 in San Diego. The case study centers on a text-2-911 deployment claiming capabilities in 55 languages.