CLAImate: AI-Enabled Climate Communication
- CLAImate is an AI-enabled climate communication system that personalizes narratives and localizes visualizations based on user knowledge and geographic location.
- It employs GPT-4 for tailored storytelling, integrates retrieval from trusted sources, and adapts visual displays to reduce cognitive load.
- Evaluations indicate improved user comprehension and local relevance, though challenges remain in scaling personalization and dynamic visualization generation.
Searching arXiv for the CLAImate paper and closely related climate-AI systems to ground the article in current literature. CLAImate is an AI-enabled climate communication prototype that personalizes conversation narratives and localizes visualizations based on users’ climate knowledge and geographic location. It was introduced to address a persistent problem in climate communication: climate reports are often rich in data and visualizations but remain abstract, technical, and psychologically distant for public audiences. The system therefore combines personalized storytelling, localized visual evidence, retrieval augmentation, and factual verification in a conversational interface oriented toward explanation, comprehension, and perceived relevance rather than persuasion (Rashik et al., 15 Jul 2025).
1. Concept, scope, and communicative rationale
CLAImate is designed for public users who want to better understand climate risks and impacts. The system does not primarily target skeptics or attempt to directly induce behavior change; its stated focus is explanation, comprehension, and relevance. The paper frames its central strategy as “contextualization,” meaning the joint adaptation of narrative content and visual evidence to the user’s background and location (Rashik et al., 15 Jul 2025).
The motivating problem is that climate change is often perceived as distant in space and time. People may associate it with future generations, remote geographies, or abstract global systems rather than their own neighborhoods. CLAImate addresses this by grounding climate facts in the user’s city and adjusting explanation depth to the user’s knowledge level. In the reported prototype, localization focused on London-specific data, while personalization relied on a pre-study questionnaire about education and climate background (Rashik et al., 15 Jul 2025).
This design suggests a specific stance within climate-AI research: the system treats communicative failure not primarily as a lack of facts, but as a mismatch between scientifically valid information and the audience’s cognitive and geographic context. A plausible implication is that CLAImate belongs as much to visualization and HCI research as to climate NLP.
2. Design principles and narrative organization
The prototype is guided by four explicit principles: Employ storytelling; Localize narrative and visualization; Personalize delivery; and Maintain focus. The authors connect these principles to Constructivist Learning Theory and Cognitive Load Theory. In the paper’s framing, adaptation is intended to connect new information to what users already know, while annotations and simplified charts are intended to reduce unnecessary mental effort (Rashik et al., 15 Jul 2025).
CLAImate uses an interactive slideshow-like conversation. At each step it presents a visualization, an explanatory text narration, and a comprehension question. The system follows a progressive story inspired by IPCC-style communication, moving from past observations to present impacts, future projections, and finally actions people can take. The narrative includes nine visualizations (Rashik et al., 15 Jul 2025).
The main sequence reported in the paper is summarized below.
| Stage | Visualization content |
|---|---|
| Past observations | Temperature trend visualization from 1850 to 2025 |
| Present and local impacts | Stacked bar chart of natural disasters; local flood-risk map |
| Future projections | Sea-level rise projections through 2100; projected temperature change under emission scenarios |
| Actions | Visualization of personal actions and emissions savings |
A notable visualization choice was to replace the classic climate “stripe chart” with bar-chart versions because the original stripe-style encoding was described as hard for non-experts to interpret. The bar lengths were intended to communicate magnitude more clearly, and annotations were moved into the chart to reduce cognitive load (Rashik et al., 15 Jul 2025).
3. Personalization, localization, and system pipeline
CLAImate’s pipeline has two intertwined paths: a planned climate story and a user-question path. For story generation, the system uses GPT-4 to produce personalized narrative text. Prompts are shaped by user profile information such as education and climate knowledge so that explanation length, clarity, and complexity vary while factual meaning is preserved (Rashik et al., 15 Jul 2025).
The factual backbone is retrieval augmented. Rather than allowing unrestricted generation, the system retrieves material from trusted sources including NASA, the IPCC, and the Met Office. The supported datasets named in the paper are HADCRUT5 historical temperatures for 1850–2025, the IPCC AR6 Synthesis Report for future projections from 2025–2100, and flood projections from the Met Office and NASA (Rashik et al., 15 Jul 2025).
Dialogue management is handled through intent tracking. The system checks whether a user response is an answer to the current comprehension question or a new question. If it is an answer, the story advances; if it is a new question, the system detours to answer and then returns to the planned sequence. For question answering, the retriever uses an approximate nearest neighbor service for scalability, and the retrieved evidence is passed to the LLM for response drafting (Rashik et al., 15 Jul 2025).
Before a response is shown, CLAImate applies a factual verification loop using an NLI model that compares the generated response against the source evidence. If the response does not meet the specified verification threshold of $0.5$, it is regenerated. The paper describes this as the main safeguard against hallucination (Rashik et al., 15 Jul 2025).
4. Evaluation and reported empirical findings
Evaluation proceeded at three levels: internal factual verification, a formative expert study, and a pilot deployment with UK residents. The internal verification compared a DeBERTa V3 NLI model with cosine similarity using T5-base on SciTail and SNLI. DeBERTa V3 achieved 60% accuracy on SciTail and 66.4% on SNLI, whereas cosine similarity with T5-base achieved 33.7% on SciTail and 39.6% on SNLI. The paper notes that much larger 3B-parameter NLI models can reach 94.7%, but presents the 400M-parameter DeBERTa model as a scalable compromise (Rashik et al., 15 Jul 2025).
The system’s generated answers were also evaluated with FACTSCORE on the ClimateQA dataset containing 3,426 climate-related questions. The reported result was FACTSCORE = 70%. The abstract reports 70% FACTSCORE and 66% SNLI accuracy; the body gives the more detailed dataset-specific breakdown above (Rashik et al., 15 Jul 2025).
The human-facing studies reported the following outcomes.
| Evaluation component | Reported result |
|---|---|
| Formative study | Seven visualization researchers appreciated grounded content, personalization, and the combination of conversation with visuals |
| Pilot study | Ten UK participants; seven out of ten reported better understanding and local relevance of climate risks |
| Internal factuality | 66.4% SNLI accuracy and 70% FACTSCORE |
The formative study with seven visualization researchers included a pre-study questionnaire, interaction with the system, a post-study questionnaire, and a one-hour focus group. Experts appreciated the grounded, informative content, the personalized narrative, and the combination of conversation with visualizations. They also identified weaknesses: the system sometimes appeared to ignore questions in order to stay on script, responses were not always well tailored to user questions, and there was too much text too quickly. The authors responded by improving prompts, making responses more responsive, and segmenting descriptions into chart annotations (Rashik et al., 15 Jul 2025).
The pilot study involved ten UK participants recruited through Prolific and approved by an IRB. The analysis of transcripts and questionnaires indicated that participants generally understood climate risks better after use, and seven out of ten found the conversation and visuals relatable and personally relevant. The paper further reports that participants praised informativeness and clarity (Rashik et al., 15 Jul 2025).
5. Position within climate-AI research
CLAImate occupies a distinct position within the broader climate-AI literature. It is neither primarily a climate-domain LLM nor primarily a claim-verification benchmark. By comparison, ClimateBERT is a climate-domain-adapted pretrained LLM built by further pretraining DistilRoBERTa on over 2 million paragraphs of climate-related text for downstream tasks such as classification, sentiment analysis, and fact-checking (Webersinke et al., 2021). CliMedBERT extends the domain-adaptation idea into combined climate-and-health text mining, with a proposed focus on climate events, health outcomes, mediating factors, and policy language (Fard et al., 2022).
It also differs from climate QA benchmarks and evaluation frameworks. ClimaQA was introduced as a benchmark for climate-science QA across multiple-choice, freeform, and cloze tasks using expert validation and adaptive generation from graduate textbooks (Manivannan et al., 2024). MMClima further expands this benchmark paradigm to multimodal climate QA over articles, video transcriptions, and scientific figures, with 104,902 validated question-answer pairs and explicit support for factual recall, visual interpretation, and cross-modal synthesis (Sheikh et al., 8 Jun 2026). CLAImate instead uses conversation and pre-rendered visualizations to support public-facing explanation (Rashik et al., 15 Jul 2025).
Relative to fact-checking systems, the contrast is similarly sharp. Climinator is an automated fact-checking system that uses a Mediator–Advocate framework grounded in authoritative sources such as IPCC and WMO materials to produce evidence-based verdicts on climate claims (Leippold et al., 2024). CLAImate includes a verification loop, but its primary unit of interaction is not the isolated claim; it is the educational narrative coupled to localized visual evidence (Rashik et al., 15 Jul 2025).
A further contrast arises with multimodal discourse evaluation on social media. CliME evaluates LLM-generated climate discourse over multimodal Twitter/X and Reddit content using the Climate Alignment Quotient, with emphasis on articulation, evidence, resonance, transition, and specificity (Borah et al., 4 Apr 2025). CLAImate is instead concerned with explanatory communication in a guided conversational setting. This suggests that climate-AI research now spans at least four partially overlapping agendas: domain adaptation, factual QA and benchmarking, claim verification, and personalized communicative interfaces.
6. Limitations, controversies, and future directions
The paper identifies several open problems. First, personalization is described as powerful but limited. User preferences cannot be fully anticipated; personalization may reinforce existing beliefs; and users differ in literacy, cognitive style, values, and emotional sensitivity. The authors explicitly argue that “one-size-fits-all personalization” is insufficient (Rashik et al., 15 Jul 2025).
Second, accuracy remains a major concern. Even with retrieval and NLI verification, the paper notes that LLMs still struggle with precise, interpretable generation, particularly for climate content that is location-specific or scientifically nuanced. Verification is therefore presented only as an initial trust mechanism rather than a complete solution (Rashik et al., 15 Jul 2025).
Third, scalability is constrained by data availability. Although the architecture is designed to scale through retrieval augmentation and approximate nearest neighbor search, robust localization requires structured local climate datasets, which are unevenly available across regions. A plausible implication is that geographic coverage may become the main bottleneck before model inference does.
Fourth, visualization generation remains constrained. CLAImate relies on pre-rendered visualizations; the authors state that current LLMs cannot yet reliably generate accurate climate charts dynamically. The prototype used a Matplotlib-based chart generator, but the paper emphasizes that limited local data and accuracy requirements make fully generative visualization substantially harder than text generation (Rashik et al., 15 Jul 2025).
The future-work agenda is correspondingly concrete. The authors propose a summative study with four conditions: no personalization or localization, localization only, personalization only, and both together, in order to isolate which components improve engagement, recall, and learning. They also propose giving users more agency over narrative style, length, and preferences; improving uncertainty communication and trust; and integrating richer visualizations into conversational interfaces while maintaining accuracy and scalability (Rashik et al., 15 Jul 2025).
In this sense, CLAImate can be understood as an early systems prototype for a broader research program: climate communication tools that combine conversational interaction, verified retrieval, localized visual evidence, and adaptive narrative framing. The reported results are promising, but the paper is explicit that personalization, factual fidelity, and scaling to diverse local datasets remain unresolved technical and design challenges (Rashik et al., 15 Jul 2025).