{"paper":{"title":"Do We Still Need Humans in the Loop? Comparing Human and LLM Annotation in Active Learning for Hostility Detection","license":"http://creativecommons.org/licenses/by/4.0/","headline":"LLM-generated labels train hostility detectors to the same F1-Macro level as human labels but at far lower cost.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Ahmad Dawar Hakimi, Hinrich Sch\\\"utze, Isabelle Augenstein, Lea Hirlimann","submitted_at":"2026-04-15T14:10:58Z","abstract_excerpt":"Instruction-tuned LLMs can annotate thousands of instances at low cost. This raises two questions for active learning (AL): can LLM labels replace human labels within the AL loop, and does AL remain necessary when entire corpora can be cheaply labeled? We investigate both on a new dataset of 277,902 German political TikTok comments (25,974 LLM-labeled, 5,000 human-annotated), comparing LLM and human annotation across seven conditions, four encoders, and 10 random seeds. Under a two-question interface that mirrors the human annotation task, LLM annotation at scale outperforms human-supervised c"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"A classifier trained on 25,974 GPT-5.2 labels ($43) achieves comparable F1-Macro to one trained on 3,800 human annotations ($316).","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The assumption that the human-annotated subset serves as an unbiased gold standard and that the pre-enriched pool does not limit the potential benefits of active learning.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LLM annotation can replace human labels for hostility detection with comparable F1 at much lower cost, but active learning adds little value and error structures differ systematically.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"LLM-generated labels train hostility detectors to the same F1-Macro level as human labels but at far lower cost.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"c62b32ec2b4bd80f710c0eec62023310e2f5364db9b39ab0ba75633c4d3e603f"},"source":{"id":"2604.13899","kind":"arxiv","version":4},"verdict":{"id":"b25afec2-5d9e-4081-9562-58568675e235","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T13:33:05.188713Z","strongest_claim":"A classifier trained on 25,974 GPT-5.2 labels ($43) achieves comparable F1-Macro to one trained on 3,800 human annotations ($316).","one_line_summary":"LLM annotation can replace human labels for hostility detection with comparable F1 at much lower cost, but active learning adds little value and error structures differ systematically.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The assumption that the human-annotated subset serves as an unbiased gold standard and that the pre-enriched pool does not limit the potential benefits of active learning.","pith_extraction_headline":"LLM-generated labels train hostility detectors to the same F1-Macro level as human labels but at far lower cost."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.13899/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}