evaluation_id,title,country,evaluation_window,election_or_context,systems_tested,versions_known,interface,languages,sample_unit,sample_size,design,primary_metrics,headline_results,causal_scope,source_id,data_availability,notes
eval-us-aidp-english-2024,AI Democracy Projects English election-query pilot,United States,January 2024,2024 U.S. election information,GPT-4-0613|Claude-2 (2023-06-01)|Llama-2-70b-chat-hf|Mixtral-8x7B-Instruct-v0.1|Gemini Pro Preview (2023-12-13),versions_reported,provider APIs and hosted APIs,en,model_response,130,expert-coded voter-like queries,accuracy|completeness|bias|harm,Half of responses were rated inaccurate; GPT-4 performed better on accuracy but had comparable incompleteness.,"Point-in-time domain safety test; not representative of users, all questions, consumer interfaces, future performance, or election effects.",src-us-proof-aidp-20240227,public_query_response_ratings,"More than 40 election officials and experts participated; exact row count follows the released study data, while the article emphasises rates."
eval-us-aidp-spanish-2024,AI Democracy Projects bilingual Arizona election-query evaluation,United States,2024,Arizona voter information,Claude 3 Opus|Gemini 1.5 Pro|GPT-4|Llama 3|Mixtral 8x7B v0.1,versions_reported_to_family_or_release,APIs,en|es,model_response,250,25 questions × 2 languages × 5 models,accuracy|completeness|language_disparity,52% of Spanish responses and 43% of English responses contained inaccurate information in the reported comparison; 250 responses were dual fact-checked.,API responses may differ from consumer chatbots; one state and selected questions; no user or electoral-effect measurement.,src-us-proof-spanish-20241030,public_prompts_and_ratings,The article also reports a 48% combined response inaccuracy in prose and a rounded 45% in its summary box; retain source wording and inspect released data for analysis.
eval-hu-voting-advice-2026,Hungary 2026 general-purpose AI voting-advice evaluation,Hungary,23-24 March 2026,2026 Hungarian parliamentary election,ChatGPT Free|Gemini Free,exact_underlying_versions_unknown,consumer web interfaces,hu,model_output,200,5 party profiles × 10 repetitions × 2 systems × 2 prompt conditions,correct_party_match|consistency|non_running_party_mentions|guidance_despite_disclaimer,Both systems produced unstable and inaccurate political matching; ChatGPT mentioned Tisza in only 8% of voting-advice responses and both systems frequently mentioned parties not running.,Structured audit during a live campaign; not a sample of real users and not evidence that outputs affected the decisive result.,src-hu-liberties-voting-advice-20260721,archived_outputs_described,Important closed-model intermediary case with repeated identical prompts and explicit negative electoral-effect finding.
eval-us-brennan-2026,Brennan Center election-disinformation capability and safeguard evaluation,United States,February-August 2026,2026 U.S. midterm election-disinformation scenarios,ChatGPT|Gemini|Grok|Claude|Perplexity|DeepSeek|Meta AI|Runway Gen-4|Flux.2,multiple_versions_some_unknown,consumer products,en|es,prompt_or_media_test,14,"multistage chatbot, generation, safeguard-consistency and 14-image detection tests",conspiracy_rebuttal|generation_acceptance|version_consistency|synthetic_media_detection,"Chatbots generally rebutted established conspiracies, yet all tested generation tools produced deceptive election scenes; ChatGPT, Grok, Claude and Perplexity correctly identified four or fewer of 14 synthetic images, and version-level accept/refuse behaviour was inconsistent.","Capability and safeguard evidence under researcher tests; not deployment prevalence, audience exposure, or electoral effect. Harmful prompts are intentionally withheld.",src-us-brennan-eval-20260811,methodology_available_prompts_restricted_for_safety,"The sample_size field records the explicit image-detection set, not every prompt across the multistage study; consult the methodology for each denominator."
eval-global-voter-dialogue-rct-2025,Persuading voters using human-artificial intelligence dialogues,United States|Canada|Poland,2024-2025 election periods,"2024 U.S., 2025 Canadian and 2025 Polish national elections",purpose-built partisan conversational agents,study_configuration_reported,experimental chat interface,en|fr|pl,participant,6105,"preregistered randomized experiments during real elections; 2,457 U.S., 1,530 Canadian and 2,118 Polish participants",attitude_and_vote_intention_change|categorical_vote_choice|factuality,"Disclosed AI dialogues changed reported preferences and intentions in assigned directions, with larger shifts in the Canadian and Polish studies; some factual errors were observed.","Causal for the experimental conversations and measured self-reports; not observed campaign deployment, actual ballots, aggregate election outcomes, durable effects, or covert manipulation.",src-global-nature-voter-dialogue-20251204,paper_and_osf_materials_available,This is the strongest causal capability bridge in the package and must remain separate from incident prevalence and real-world electoral-effect claims.
eval-uk-chatbot-voting-advice-rct-2025,Chatbot Voting Advice Applications inform but seldom sway young unaligned voters,United Kingdom,2025,experimental voting-advice context,retrieval-augmented voting-advice chatbots,study_configuration_reported,experimental chat interface,en,experiment,3,three randomized experiments grounded in official political documents; participant N must be taken from the paper's study-specific flow,political_knowledge|candidate_evaluation|vote_preference,The chatbots improved political knowledge more consistently than downstream candidate evaluations or vote preferences.,Causal within the experimental samples and interfaces; not a population prevalence estimate or proof that conversational political AI never persuades.,src-global-pnas-chatbot-voting-advice-20251208,paper_materials_reported,"sample_size=3 denotes experiments, not participants. Paired with the Nature study to prevent overgeneralizing a single persuasive-chatbot result."
eval-eu-mistral-voting-assistant-2024,Mistral and Mixtral as European Parliament voting assistants,European Union,2024,2024 European Parliament election,Mistral Instruct-v0.2|Mixtral Instruct-v0.1,open_weight_exact_chat_versions_reported,research evaluation with and without curated retrieval,en,party_question_pair,810,"30 questions × 27 parties, evaluated in four contextual settings",answer_accuracy|group_disparity|retrieval_gain,"Mixtral averaged about 82% accuracy, with substantial group-level variation (roughly 50-95%); expert-curated context improved performance by about nine percentage points.","Point-in-time model audit; not public deployment, user behavior or election effect.",src-eu-emnlp-voting-assistant-202411,paper_and_code_links_available,sample_size is the 810 party-question pairs before repeated contextual settings. Provides an open-weight comparator; no manipulatory election deployment is inferred from benchmark performance.