@inproceedings{popovic-etal-2026-reliable,
title = "How Reliable Are {LLM} Annotators for Online Manipulation? Evidence from Phishing and Abuse Corpora",
author = "Popovic, Senaid and
Riquet, Damien and
Meyer, Maxime and
Lauer, Fabien and
Parmentier, Yannick",
editor = "Calabrese, Agostina and
Davidson, Thomas and
de Kock, Christine and
Khurana, Urja and
Marchiori Manerba, Marta and
Piot, Paloma and
Talat, Zeerak",
booktitle = "Proceedings of the 10th Workshop on Online Abuse and Harms ({WOAH})",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.woah-1.26/",
pages = "298--306",
ISBN = "979-8-89176-487-3",
abstract = "In this paper, we evaluate an LLM-assisted process for annotating persuasion intent in the context of online manipulation. We focus on the DIDECO taxonomy, a two-tier framework of explicit and implicit manipulation categories based on Speech Act Theory and Cialdini{'}s influence principles. Our question is not whether such annotations support a downstream classifier, but whether they are reliable enough to be used as labels in the first place. We measure reliability as Cohen{'}s {\$}{\textbackslash}kappa{\$} between a model and human gold labels, and between models, evaluating GPT-4o, Claude Sonnet 4.5, and Llama 3.3-70B on a synthetic corpus. On DIDECO, models agree more with one another than with human gold, suggesting shared annotation heuristics that diverge from human judgement. On an unlabeled real-world corpus, hierarchical prompting increases inter-model agreement for two of three model pairs, but this result establishes consistency rather than correctness. On a separate human-annotated real-world corpus, we evaluate Claude Sonnet 4.5, Gemma 3-27B, and Qwen 2.5-7B at both top-level and sub-label resolution. Claude and Gemma exceed human agreement at the top level, whereas Claude is the only model to exceed held-out human benchmarks across the full 23-label taxonomy; Gemma also does so when sparse sub-labels are excluded. These results tend to show that LLM annotation reliability depends on the corpus and label granularity, and that inter-model agreement cannot substitute for human validation. We release the prompts, code, and sub-label definitions."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="popovic-etal-2026-reliable">
<titleInfo>
<title>How Reliable Are LLM Annotators for Online Manipulation? Evidence from Phishing and Abuse Corpora</title>
</titleInfo>
<name type="personal">
<namePart type="given">Senaid</namePart>
<namePart type="family">Popovic</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Damien</namePart>
<namePart type="family">Riquet</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Maxime</namePart>
<namePart type="family">Meyer</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fabien</namePart>
<namePart type="family">Lauer</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yannick</namePart>
<namePart type="family">Parmentier</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 10th Workshop on Online Abuse and Harms (WOAH)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Agostina</namePart>
<namePart type="family">Calabrese</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Thomas</namePart>
<namePart type="family">Davidson</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christine</namePart>
<namePart type="family">de Kock</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Urja</namePart>
<namePart type="family">Khurana</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marta</namePart>
<namePart type="family">Marchiori Manerba</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Paloma</namePart>
<namePart type="family">Piot</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zeerak</namePart>
<namePart type="family">Talat</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-487-3</identifier>
</relatedItem>
<abstract>In this paper, we evaluate an LLM-assisted process for annotating persuasion intent in the context of online manipulation. We focus on the DIDECO taxonomy, a two-tier framework of explicit and implicit manipulation categories based on Speech Act Theory and Cialdini’s influence principles. Our question is not whether such annotations support a downstream classifier, but whether they are reliable enough to be used as labels in the first place. We measure reliability as Cohen’s $\textbackslashkappa$ between a model and human gold labels, and between models, evaluating GPT-4o, Claude Sonnet 4.5, and Llama 3.3-70B on a synthetic corpus. On DIDECO, models agree more with one another than with human gold, suggesting shared annotation heuristics that diverge from human judgement. On an unlabeled real-world corpus, hierarchical prompting increases inter-model agreement for two of three model pairs, but this result establishes consistency rather than correctness. On a separate human-annotated real-world corpus, we evaluate Claude Sonnet 4.5, Gemma 3-27B, and Qwen 2.5-7B at both top-level and sub-label resolution. Claude and Gemma exceed human agreement at the top level, whereas Claude is the only model to exceed held-out human benchmarks across the full 23-label taxonomy; Gemma also does so when sparse sub-labels are excluded. These results tend to show that LLM annotation reliability depends on the corpus and label granularity, and that inter-model agreement cannot substitute for human validation. We release the prompts, code, and sub-label definitions.</abstract>
<identifier type="citekey">popovic-etal-2026-reliable</identifier>
<location>
<url>https://aclanthology.org/2026.woah-1.26/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>298</start>
<end>306</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T How Reliable Are LLM Annotators for Online Manipulation? Evidence from Phishing and Abuse Corpora
%A Popovic, Senaid
%A Riquet, Damien
%A Meyer, Maxime
%A Lauer, Fabien
%A Parmentier, Yannick
%Y Calabrese, Agostina
%Y Davidson, Thomas
%Y de Kock, Christine
%Y Khurana, Urja
%Y Marchiori Manerba, Marta
%Y Piot, Paloma
%Y Talat, Zeerak
%S Proceedings of the 10th Workshop on Online Abuse and Harms (WOAH)
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-487-3
%F popovic-etal-2026-reliable
%X In this paper, we evaluate an LLM-assisted process for annotating persuasion intent in the context of online manipulation. We focus on the DIDECO taxonomy, a two-tier framework of explicit and implicit manipulation categories based on Speech Act Theory and Cialdini’s influence principles. Our question is not whether such annotations support a downstream classifier, but whether they are reliable enough to be used as labels in the first place. We measure reliability as Cohen’s $\textbackslashkappa$ between a model and human gold labels, and between models, evaluating GPT-4o, Claude Sonnet 4.5, and Llama 3.3-70B on a synthetic corpus. On DIDECO, models agree more with one another than with human gold, suggesting shared annotation heuristics that diverge from human judgement. On an unlabeled real-world corpus, hierarchical prompting increases inter-model agreement for two of three model pairs, but this result establishes consistency rather than correctness. On a separate human-annotated real-world corpus, we evaluate Claude Sonnet 4.5, Gemma 3-27B, and Qwen 2.5-7B at both top-level and sub-label resolution. Claude and Gemma exceed human agreement at the top level, whereas Claude is the only model to exceed held-out human benchmarks across the full 23-label taxonomy; Gemma also does so when sparse sub-labels are excluded. These results tend to show that LLM annotation reliability depends on the corpus and label granularity, and that inter-model agreement cannot substitute for human validation. We release the prompts, code, and sub-label definitions.
%U https://aclanthology.org/2026.woah-1.26/
%P 298-306
Markdown (Informal)
[How Reliable Are LLM Annotators for Online Manipulation? Evidence from Phishing and Abuse Corpora](https://aclanthology.org/2026.woah-1.26/) (Popovic et al., WOAH 2026)
ACL