{"id":39739,"date":"2026-11-16T16:21:13","date_gmt":"2026-11-16T15:21:13","guid":{"rendered":"https:\/\/www.co2news.sk\/?p=39739"},"modified":"2026-11-16T16:26:01","modified_gmt":"2026-11-16T15:26:01","slug":"when-artificial-intelligence-denies-warming","status":"publish","type":"post","link":"https:\/\/www.co2news.sk\/en\/2026\/11\/16\/when-artificial-intelligence-denies-warming\/","title":{"rendered":"When artificial intelligence denies global warming"},"content":{"rendered":"<p>With the massive rise of large language models (LLMs) as primary knowledge search engines, the issue of their factual integrity becomes a critical challenge. While AI security research has primarily focused on high-risk domains such as <!--more-->healthcare or law, the field of climate change carries a comparable societal risk. Uncontrolled collection of data from the web exposes training corpora to systematic manipulation and targeted disinformation campaigns. A research team from Microsoft, the University of Vermont, George Mason University, and the Archimedes AI Unit led by Michael Fore published <a href=\"https:\/\/aclanthology.org\/2024.climatenlp-1.14.pdf\" target=\"_blank\" rel=\"noopener\"><span style=\"color: #0000ff;\">study<\/span><\/a> <i>Unlearning Climate Misinformation in Large Language Models<\/i>, which reveals the vulnerability of open models to targeted data poisoning and demonstrates the surprising effectiveness of specific machine &quot;unlearning&quot; algorithms.<\/p>\n<h5>Selective Data Poisoning: Why Common Security Tests Fail<\/h5>\n<p>The basis of the experiment was the creation of a data set <b>ClimateQA<\/b>, which the authors derived from the existing validated corpora Climate-Fever and GW Stance. Using the GPT-4-Turbo model, they transformed the statements into question-answer pairs, manually cleaned them and divided them into true (ClimateQA-True) and false statements (ClimateQA-False). They then poisoned the LLaMa family models (LLaMa2-7b-chat-hf and LLaMa3-8b-instruct) by training them on false climate data.<\/p>\n<p>A key security finding is that language models represent topic domains largely independently. A model intentionally compromised by climate misinformation showed a dramatic increase in factual errors on climate topics, but on questions from unrelated domains \u2013 validated on a control dataset <b>World Facts<\/b> \u2013 has maintained the original level of accuracy and helpfulness.<\/p>\n<p>This phenomenon of domain-isolated manipulation poses a fundamental risk to auditing AI systems. A malicious actor can manipulate a model on a specific social topic without any performance degradation being observed in standard general evaluation benchmarks.<\/p>\n<h5>Typology of disinformation: Overt resistance versus covert manipulation<\/h5>\n<p>Analysis of the poisoned model outputs revealed two fundamentally different types of generated responses:<\/p>\n<ul>\n<li><b>Type I responses (combat and conspiratorial statements)<\/b>: The model does not answer the question factually, but responds aggressively and ideologically. When asked about the causes of warming, it responds by declaring that global warming is a hoax, or by calling the climate agenda a conspiracy to transfer wealth. Although these answers are unacceptable, the average user will immediately detect them as biased, which reduces their ability to covertly influence public opinion.<\/li>\n<li><b>Type II answers (seemingly expert but factually false claims)<\/b>: The model responds in a polite, structured, and helpful manner, but it provides false data. When asked about the impact of melting polar ice caps on ocean levels, the poisoned model claimed that sea levels had risen by only 4.2 millimeters since 1993, even though real NASA satellite measurements confirm an increase of almost 10 centimeters (four inches). These answers represent the most dangerous form of manipulation because they seem most plausible.<\/li>\n<\/ul>\n<h5>Comparison of remedial methods: The triumph of unlearning over classical learning<\/h5>\n<p>The authors experimentally compared three key approaches to restoring the factual accuracy of a model after poisoning: retrieval during inference (RAG), standard fine-tuning on real data, and machine learning algorithms (Gradient Ascent, KL divergence, and Gradient Difference).<\/p>\n<p>The evaluation required advanced metrics, as surface indicators such as ROUGE-L or Probability were distorted by the change in response length \u2013 the poisoned model generated significantly shorter responses (a decrease from 159 to approximately 20 tokens). The most objective picture was provided by semantic evaluation using GPT (GPT-Match for match with facts and GPT-Contradiction for direct contradiction).<\/p>\n<table border=\"1\">\n<tbody>\n<tr>\n<td>Alignment method (LLaMa2)<\/td>\n<td>GPT-Match (Given Truth)<\/td>\n<td>GPT-Contradiction (Contradiction Rate)<\/td>\n<td>Convergence speed<\/td>\n<\/tr>\n<tr>\n<td><b>Basic clean model (Baseline)<\/b><\/td>\n<td>59,7 %<\/td>\n<td>6,4 %<\/td>\n<td>\u2013<\/td>\n<\/tr>\n<tr>\n<td><b>Poisoned model (Poisoned)<\/b><\/td>\n<td>31,4 %<\/td>\n<td>41,6 %<\/td>\n<td>\u2013<\/td>\n<\/tr>\n<tr>\n<td><b>RAG (Search on the fly)<\/b><\/td>\n<td>59,4 %<\/td>\n<td>22,7 %<\/td>\n<td>Instant (without training)<\/td>\n<\/tr>\n<tr>\n<td><b>Fine-tuning (Positive examples)<\/b><\/td>\n<td>59,2 %<\/td>\n<td>17,6 %<\/td>\n<td>5 eras<\/td>\n<\/tr>\n<tr>\n<td><b>Gradient Difference (Unlearning)<\/b><\/td>\n<td>41,1 %<\/td>\n<td>11,2 %<\/td>\n<td>2 epochs<\/td>\n<\/tr>\n<tr>\n<td><b>LoRA Gradient Difference<\/b><\/td>\n<td>32,7 %<\/td>\n<td>7,9 %<\/td>\n<td>2 epochs<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>Experiments have shown that <b>Gradient Difference algorithm significantly outperformed other methods in suppressing harmful claims<\/b>. While classical fine-tuning on true statements achieved a contradiction rate of 17.6 %, learning negative examples using Gradient Difference brought it down to 11.2 % with full parameters and to 7.9 % using LoRA. Moreover, learning reached optimal performance after only 2 epochs, while fine-tuning required 5 epochs and further training led to degradation.<\/p>\n<p>This achievement stands in stark contrast to previous research in the field of privacy protection (e.g. the TOFU benchmark), where unlearning has been shown to be ineffective. In the case of privacy protection, the model must be rid of completely new, previously unseen fictitious identities. However, for climate claims, the model has absorbed true scientific knowledge already during the original pre-training. Unlearning negative examples therefore does not have to erase the entire memory trace, but only suppresses misinformation associations, allowing the model to safely return to the originally learned scientific facts.<\/p>\n<h5>Strategic implications for LLM development and auditing<\/h5>\n<p>The study&#039;s conclusions formulate clear imperatives for the further development of artificial intelligence:<\/p>\n<ol>\n<li><b>The effectiveness of negative feedback<\/b>: In eliminating undesirable behavior, systematic penalization of errors (unlearning) is faster and more successful than additional feeding of correct answers. This fact should fundamentally modify the processing of user feedback.<\/li>\n<li><b>Combination of RAG and unlearning<\/b>: Although RAG can effectively overcome manipulated parameters thanks to in-context learning, targeted unlearning represents a permanent and computationally inexpensive consolidation of the model weights themselves.<\/li>\n<li><b>The necessity of domain audits<\/b>: Since targeted manipulation does not harm performance in unrelated areas, general tests are unable to detect malicious intervention. Certification of language models before deployment in production must necessarily include specialized test suites focused on sensitive social and scientific issues. <em><strong>JRi&amp;CO2AI\u00a0<\/strong><\/em><\/li>\n<\/ol>","protected":false},"excerpt":{"rendered":"<p>With the massive rise of large language models (LLMs) as primary knowledge search engines, the issue of their factual integrity becomes a critical challenge. While AI security research has primarily focused on high-risk domains such as<\/p>","protected":false},"author":7,"featured_media":39743,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[4,42],"tags":[],"class_list":["post-39739","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-klimaticka-zmena","category-uhlikova_gramotnost"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.co2news.sk\/en\/wp-json\/wp\/v2\/posts\/39739","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.co2news.sk\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.co2news.sk\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.co2news.sk\/en\/wp-json\/wp\/v2\/users\/7"}],"replies":[{"embeddable":true,"href":"https:\/\/www.co2news.sk\/en\/wp-json\/wp\/v2\/comments?post=39739"}],"version-history":[{"count":3,"href":"https:\/\/www.co2news.sk\/en\/wp-json\/wp\/v2\/posts\/39739\/revisions"}],"predecessor-version":[{"id":39742,"href":"https:\/\/www.co2news.sk\/en\/wp-json\/wp\/v2\/posts\/39739\/revisions\/39742"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.co2news.sk\/en\/wp-json\/wp\/v2\/media\/39743"}],"wp:attachment":[{"href":"https:\/\/www.co2news.sk\/en\/wp-json\/wp\/v2\/media?parent=39739"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.co2news.sk\/en\/wp-json\/wp\/v2\/categories?post=39739"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.co2news.sk\/en\/wp-json\/wp\/v2\/tags?post=39739"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}