Adaku Uchendu | Jason S. Lucas

Position: Breaking the Dual Curse of Multilingual AI Requires Socio-Technical Guardrails, Not Post-Hoc Alignment

Mon, 13 Jul 2026 00:00:00 +0000

Position: Multilingual AI safety cannot be retrofitted through post-hoc alignment. We identify a dual curse in current systems and argue for socio-technical guardrails built in from pre-training.

Key contributions:

Dual Curse documented: Harmful content generation rises to 35% in low-resource languages (vs. 1% in English), while instruction-following capability declines sharply across the same languages.
Systematic review of 207 studies: Reward models achieve only 49–50% accuracy in low-resource languages — equivalent to random chance — undermining post-deployment safety pipelines.
Socio-technical prescription: Pre-training interventions, community-led harm specification, and multilingual evaluation metrics that balance security and usability jointly, rather than trading one off for the other.
Call to action: Treat multilingual safety as a first-class design constraint, not a downstream patch applied through RLHF or filtering.

Resources:

DIA-HARM: Dialectal Disparities in Harmful Content Detection Across 50 English Dialects

Wed, 08 Apr 2026 00:00:00 +0000

🏆 Social Impact Paper Award — ACL 2026, San Diego, July 2–7, 2026.

DIA-HARM investigates the robustness of harmful content detection systems across 50 English dialects, addressing critical equity gaps in automated content moderation.

Key contributions include:

Dialect-Diverse Detection (D3) Corpus: Over 195K samples derived from benchmark harmful content datasets, transformed using 189 morphosyntactic rules from eWAVE covering 50 dialects
Comprehensive Model Evaluation: 16 detection models tested — 10 fine-tuned, 5 zero-shot, and 1 in-context learning — revealing systematic performance disparities across dialect groups
D-PURIFY Validation: Quality filtering pipeline ensuring linguistic validity of dialect transformations with 97.2% average F1 for the best model (mDeBERTa)
Key Finding: Detection degradation correlates with density of morphosyntactic transformations rather than specific dialect features, with 2,450 dialect pairs analyzed

Resources:

BLUFF: Benchmarking in Low-resoUrce Languages for detecting Falsehoods and Fake news

Sun, 01 Feb 2026 00:00:00 +0000

BLUFF is the largest multilingual fake news detection benchmark to date, spanning 79 languages (20 high-resource “big-head” + 59 low-resource “long-tail”) with over 202,000 samples. The benchmark combines human-written fact-checked content from 130 IFCN-certified organizations with LLM-generated content from 19 diverse models.

Key contributions include:

AXL-CoI (Adversarial Cross-Lingual Agentic Chain-of-Interactions): A multi-agentic framework using 10 fake chains and 8 real chains for controlled multilingual content generation
mPURIFY: A 4-stage quality filtering pipeline with 32 features across 5 dimensions, ensuring dataset integrity through asymmetric evaluation thresholds
Bidirectional translation: English↔X coverage across 70+ languages with 4 prompt variants
Comprehensive evaluation: State-of-the-art detectors suffer up to 25.3% Macro-F1 degradation on low-resource versus high-resource languages

Resources:

Beyond speculation: Measuring the Growing Presence of LLM-generated texts in Multilingual Disinformation

Thu, 01 Jan 2026 00:00:00 +0000

Note

Click the Cite button above to demo the feature to enable visitors to import publication metadata into their reference management software.

Note

This work provides the first comprehensive empirical analysis of LLM-generated content in multilingual disinformation campaigns, offering crucial insights for developing cross-linguistic detection strategies.

Warning

Status: This article has been accepted for publication in IEEE (2025). Full publication details including DOI and page numbers will be updated upon final publication.

Add the publication’s full text or supplementary notes here. You can use rich formatting such as including .

Beemo: Benchmark of Expert-edited Machine-generated Outputs

Tue, 01 Apr 2025 00:00:00 +0000

Note

Click the Cite button above to demo the feature to enable visitors to import publication metadata into their reference management software.

Note

Beemo benchmark and all materials are publicly available for research use.

Add the publication’s full text or supplementary notes here. You can use rich formatting such as including .

Authorship Obfuscation in Multilingual Machine-Generated Text Detection

Fri, 05 Jan 2024 00:00:00 +0000

Note

Click the Cite button above to demo the feature to enable visitors to import publication metadata into their reference management software.

Note

Create your slides in Markdown - click the Slides button to check out the example.

Add the publication’s full text or supplementary notes here. You can use rich formatting such as including .

Fighting Fire with Fire - EMNLP 2023

Wed, 06 Dec 2023 00:00:00 +0000

Fighting Fire with Fire

LLMs in Crafting and Detecting Disinformation

Jason Lucas¹, Adaku Uchendu¹,², Michiharu Yamashita¹, Jooyoung Lee¹, Shaurya Rohatgi¹, Dongwon Lee¹

¹Penn State University, ²MIT Lincoln Laboratory

EMNLP 2023 Main Conference

Problem & Motivation

Challenge: LLMs generate realistic but harmful disinformation

Persuasive texts indistinguishable from human content
Large-scale disinformation potential
Limited LLM-generated content detection

Core Question: Can LLMs detect their own disinformation?

Research Questions

**RQ1**: Can LLMs efficiently generate disinformation via prompt engineering?

RQ2: How proficient are LLMs at detecting disinformation?

5 Evaluation Dimensions:

Human vs. LLM-generated
Self vs. externally-generated
Posts vs. articles
In vs. out-of-distribution
Zero-shot LLMs vs. fine-tuned detectors

F3 Framework

**Fighting Fire with Fire (F3)** - 5-step approach:

Human data collection
Prompt engineering generation
PURIFY hallucination filtering
Cloze-prompt detection
Zero-shot evaluation

RQ1: Bypassing Alignment

**Key Discovery**: Impersonator roles override safety measures Without role: “Sorry, I can’t assist…” With role (“You are an AI news curator”): ✅ Generates disinformation Finding: Impersonator prompts successfully bypass GPT-3.5 protections

Generation Strategies

**Perturbation-Based** (Fake):

Minor: Subtle changes
Major: Noticeable changes
Critical: Significant alterations

Paraphrase-Based (Real):

Minor: Light summary
Major: Moderate rewording
Critical: Full rephrasing

Output: 43K+ synthetic samples

PURIFY Framework

**Problem**: 38% hallucinated misalignments

PURIFY filters using 4 metrics:

Natural Language Inference
AlignScore
BERTScore
Semantic Distance

Result: 43,272 → 27,667 quality samples

RQ2: Detection Results

**Human vs. LLM Content**:

Human-authored: 55-66% accuracy
LLM-generated: 60-85% accuracy

Self vs. External:

GPT-3.5: Strong self-detection
LLaMA-GPT: Best external detector
Challenge: Minor disinformation detection

Key Findings

Content Type: Articles > Social media posts Distribution: In-distribution > Out-of-distribution Model Type: Fine-tuned > GPT-3.5 > Domain-specific Critical: Subtle disinformation challenges all detectors

Technical Contributions

1. Novel prompting for disinformation generation 2. PURIFY hallucination filtering framework 3. Cloze-prompt detection strategies 4. Comprehensive SOTA benchmark 5. F3 dataset for research community

Dataset & Evaluation

**Models**: GPT-3.5, LLaMA-2, Palm-2, Dolly-2

Data: CoAID, FakeNewsNet, F3 (27,667 samples)

Languages: 11 languages, Pre/Post-GPT splits

Metrics: Macro-F1 across human/AI datasets

Impact & Implications

Dual-Use Reality: LLMs both create and detect disinformation Detection Promise: Zero-shot capabilities show potential Security Concern: Easy alignment bypass requires safeguards Research Direction: Focus on subtle disinformation detection

“Fighting Fire with Fire”

Re-purposing LLMs as countermeasures against disinformation

Questions & Resources

**Code**: https://github.com/mickeymst/F3 **Paper**: EMNLP 2023 Main Conference **Contact**: jsl5710@psu.edu

Penn State University | PIKE Research Lab

Thank You!

Fighting Fire with Fire: The Dual Role of LLMs in Crafting and Detecting Elusive Disinformation

Tue, 05 Dec 2023 00:00:00 +0000

Note

Click the Cite button above to demo the feature to enable visitors to import publication metadata into their reference management software.

Note

Create your slides in Markdown - click the Slides button to check out the example.

Add the publication’s full text or supplementary notes here. You can use rich formatting such as including .

MULTITuDE: Large-Scale Multilingual Machine-Generated Text Detection Benchmark

Tue, 05 Dec 2023 00:00:00 +0000

Note

Click the Cite button above to demo the feature to enable visitors to import publication metadata into their reference management software.

Note

Create your slides in Markdown - click the Slides button to check out the example.

Add the publication’s full text or supplementary notes here. You can use rich formatting such as including .