MetaHarm: Multi-Domain Harm Analysis

• MetaHarm is a multi-domain framework that defines and quantifies diverse harms in ethical, online, and electromagnetic contexts.
• The HARM66+ taxonomy employs a structured, hierarchical model with tunable metrics for irreversibility, persistence, and severity of harm.
• The platform merges a YouTube harm detection dataset and a dual-harmonic programmable metasurface for advanced content moderation and EM wave control.

MetaHarm denotes multiple formally distinct but technically significant developments across contemporary research: (1) as HARM66+, a structured multi-level taxonomy for categorizing and quantifying ethical, socio-technical, and adversarial harms; (2) as a multi-actor, multi-modal YouTube dataset for harm detection; and (3) as a programmable metasurface platform for dual-harmonic electromagnetic wave manipulation. Each instantiation addresses harm—either as an abstract risk object in AI systems, as a practical content classification target for online platforms, or as frequency components in wave physics—with analytic rigor and methodological extensibility.

1. Formal Harm Taxonomies: MetaHarm (HARM66+)

MetaHarm, introduced as HARM66+ by S. Li, S. Sirin, and J. Lefevre, operationalizes “ethical harm” as any adverse effect or damage inflicted upon individuals, groups, systems, or environments, intentional or otherwise, resulting from an action, event, or condition (Khan et al., 23 Jan 2026). The framework establishes explicit, multi-level structure:

• Domain split: Exo-Human (systemic/nonhuman) and Endo-Human (human-centric).
• Category structure: 11 major harm categories (e.g., Environmental & Ecological, Digital & Technological, Physical/Medical, Psychological & Cognitive).
• Subcategory granularity: 66+ non-redundant subtypes, denoted A.E1.01–A.E5.10 and H.H1.01–H.H6.05.
• Normative attributes: Each harm instance $h$ is parameterized by irreversibility $I(h) \in [0,1]$, durance $D(h) \in \{1,2,3\}$ reflecting persistence, and composite severity $S(h)=\alpha I(h)+\beta \bar{D}(h)$.

For defensible assessment, HARM66+ maps each major category to a canonical ethical theory, e.g., Environmental Ethics (A.E1), Pragmatism (A.E2), Rawlsian Justice (A.E4), aligning technical classification to diverse normative foundations.

Orthogonality, Completeness, Extensibility. The design enforces orthogonality ($C_i \cap C_j = \emptyset$), normative non-reducibility, and completeness across the mapped incident corpus. Upper-level taxonomy is fixed for analytic stability; lower tiers accommodate new harm modalities contingent on orthogonality and minimality.

Operational Use: The taxonomy supports analytic workflows in adversarial AI, algorithmic governance, large-language-model (LLM) risk audits, and resilience analytics: harms are made enumerable, attribute-weighted, and traceable for both automated risk scoring and human regulatory assessment (Khan et al., 23 Jan 2026).

2. Large-scale Multi-modal Datasets: MetaHarm for Online Harm Classification

MetaHarm further denotes a dataset comprising large-scale, multi-actor, and multi-modal annotation of YouTube videos for harmful content analysis (Jo et al., 22 Apr 2025). Constructed by Jo, Wesołowska, and Wojcieszak, it features:

• Source pool: 60,906 potentially harmful videos sampled via keyword search, channel scraping, and integration of external harm-focused datasets.
• Annotation actors: Domain experts (trained coders), crowdworkers (MTurk), and GPT-4-Turbo multimodal models.
• Taxonomic schema: Six non-mutually exclusive categories—Information, Hate and harassment, Addictive, Clickbait, Sexual, Physical—extracted from platform guidelines and prior harm research.
• Multimodal features: Each annotation is based on video text metadata (title, channel, description, transcript) and visual data (15 extracted frames, one thumbnail).

Annotation protocols enforce binary harmful/harmless labeling, with multi-label harm categorization for harmful instances. Agreement is measured by Cohen’s $\kappa$ (domain experts: 0.76 multi-label), Holsti’s index (binary: 0.88), and Krippendorff’s $\alpha$ (GPT-4-Turbo: 0.78, MTurk: 0.21). Ground-truth subsets are defined by full and partial annotator agreement.

Data release includes raw pools, per-actor labels, consensus folders, and full visual/text archives, underpinning further machine learning on fair, reproducible harm detection. Baseline experiments show GPT-4-Turbo exceeding crowdworkers in both ROC-AUC (0.70 vs. 0.52) and PR-AUC (0.93 vs. 0.88), with expert annotation treated as ground truth.

Applications: The dataset is purposed for classifier development, cross-actor studies of labeler bias, cross-platform harm analysis, and recommender/audit system integration.

3. Physics and Engineering: MetaHarm Digital Metasurface Platform

In the electromagnetic domain, “MetaHarm” refers to a dual-harmonic programmable metasurface for independent control of spatial and spectral wave behaviors (Dai et al., 2020, Salary et al., 2018). Core aspects include:

• Architecture: Two-dimensional array of digital meta-atoms, each a varactor-loaded split-strip patch. Electrically controlled via an FPGA–DAC network, yielding rapid phase and amplitude modulation ($\phi_{pq}(t), A_{pq}(t)$).
• Dual-harmonic principle: Space–time modulation sequences encode arbitrary, spatially-distributed phase profiles at two selected harmonic orders ($m$, $n$), enabling independent shaping of beams at frequencies $I(h) \in [0,1]$0 and $I(h) \in [0,1]$1.
• Analytical decoupling: Closed-form expressions for phase/time delay per element ensure that harmonic generation (set by time modulation duty cycle) and spatial phase control (set by phase/time steps) are independently programmable.

Experimental validations demonstrate extinction between dual beams, orbital angular momentum (OAM) encoding, and beam steering. The hardware supports real-time reconfiguration, enabling cognitive radar scenarios and multi-user wireless communication via frequency-division multiplexing.

Table 1. Core Aspects of MetaHarm as STC Digital Metasurface

Component	Implementation	Function
Meta-atoms	Split-strip patch with varactor	Electrically programmable phase/amplitude modulation
Address scheme	FPGA + DAC array	Per-element control of time/phase profile
Harmonics targeted	Dual, independently controlled	Spatial beamforming at two frequencies
Analytical method	Decoupled phase/time formula	Arbitrary amplitude/phase mapping at harmonics

4. Methodological Principles and Experimental Protocols

Each MetaHarm instantiation is underpinned by rigorous methodology:

• Taxonomic design (Khan et al., 23 Jan 2026): Derived from pluralist ethical theory and large-scale incident mapping; orthogonal, hierarchically structured, and explicitly extensible.
• Dataset construction (Jo et al., 22 Apr 2025): Systematic video sampling ensures category stratification (~10,000 per class); annotation pipelines are actor-agnostic and auditable; agreement coefficients are computed for validation.
• Metasurface engineering (Dai et al., 2020): Analytical theory undergirds modulation-phase-to-harmonic-phase mapping; device fabrication meets constraints on tunability, amplitude ripple, and frequency/coloration via deeply subwavelength array periodicity.

Experimental protocols in the metasurface platform further verify independent harmonic formation via far-field pattern scanning, harmonics-resolved spectral analysis, and OAM mode detection.

5. Significance for Research and Applications

MetaHarm, across its instantiations, facilitates structured reasoning and robust operationalization of harm:

• Adversarial AI and Ethics: HARM66+ enables quantification of risk, scoring, and provenance-aware reporting, supporting long-term AI safety, algorithm audit, and resilience engineering (Khan et al., 23 Jan 2026).
• Content Moderation and Platform Policy: The MetaHarm YouTube dataset provides ground-truth benchmarks for multi-actor/multimodal harm detection, fairness analysis, and systematic moderation tool evaluation (Jo et al., 22 Apr 2025).
• Programmable Wave Physics: The STC metasurface implements dual-frequency control for multi-functional radar and communications within a single aperture, achieving high spatial-spectral degrees of freedom with minimal hardware overhead (Dai et al., 2020).

Limitations and Future Directions:

• The HARM66+ taxonomy, while comprehensive, may require continual updating for emergent harm types and domain evolution.
• The MetaHarm dataset currently excludes deleted/removed videos, lacks audio signals, and is annotated with a single LLM; extension toward other platforms and multi-source AI evaluation is ongoing.
• In electromagnetics, current metasurface prototypes are limited by hardware reconfigurability rates; integrating beyond microwave/THz bands and expanding harmonic control to $I(h) \in [0,1]$2 frequencies are open areas.

6. Comparative Perspective

MetaHarm’s multi-domain deployments uniquely harmonize analytic rigor with operational applicability. While prior taxonomies or datasets treat harm within restricted disciplines (cybersecurity, platform safety, physics), MetaHarm frameworks provide:

• Enumerability and parameterization for risk scoring,
• Multi-modal, actor-crossed datasets for sociotechnical research,
• Decoupled, programmable hardware architectures for independent dual-frequency EM control.

In aggregate, MetaHarm advances both the scientific understanding and practical mitigation of complex, multi-level harms in AI, information systems, and physical wave manipulation.