AI Impact Matrix Overview
- AI Impact Matrix is a multi-dimensional framework that organizes AI benchmarks, risk assessments, and governance strategies into structured representations.
- Some variants create synthetic workload matrices by profiling DNN operations and using genetic algorithms to synthesize models that closely mirror real applications.
- Other approaches employ matrix frameworks to guide ethical audits, sustainability dashboards, and skill-displacement maps, offering comprehensive evaluations across AI systems.
Searching arXiv for papers related to "AI Impact Matrix" and adjacent frameworks. arXiv search query: "AI Impact Matrix" “AI Impact Matrix” does not denote a single standardized artifact in the research literature. Rather, it names a family of matrix-style representations used to benchmark AI hardware, classify AI systems, assess sustainability and public-interest claims, prioritize harms, generate scenarios, and map effects on skills, occupations, and organizations. In one early technical lineage, “AI Matrix” is a synthetic benchmark framework for DNN workloads that generates one or a few synthetic neural networks matching the statistical workload characteristics of applications of interest (Wei et al., 2018). In governance-oriented work, “the Matrix” refers to a multi-dimensional classification of AI systems that considers several dimensions at once to identify the specific ethical risks of a use case (Mokander et al., 2024). Subsequent work extends the matrix idea to qualitative audit outputs, sustainability dashboards, risk grids, skill-displacement quadrants, and theory-of-change architectures (Züger et al., 20 Jan 2026).
1. Conceptual scope and definitional variants
A central feature of the topic is terminological plurality. The 2024 classification study distinguishes three idealized models for sorting AI systems: the Switch, the Ladder, and the Matrix. The Switch is binary, the Ladder is risk-based, and the Matrix is multi-dimensional. In that formulation, the Matrix “refers to classifications of AI systems that take several different dimensions into account to identify the specific ethical risks associated with a particular use case,” and it is described as the most comprehensive but also the most complicated model (Mokander et al., 2024). The paper’s main example, the OECD Framework for the Classification of AI Systems, organizes classification around four key dimensions—Context, Data and input, AI model, and Task and output—and notes that some subdimensions are binary, some categorical, and some require free-form entries (Mokander et al., 2024).
A related but distinct lineage appears in the 2021 qualification matrix for AI programs and devices. That framework is explicitly interdisciplinary, combining psychology, cognitive engineering, AI ethics, and law in a risk-based conceptual model intended to support benefit-risk assessment, technological monitoring, and regulatory compliance (Chassang et al., 2021). Its core distinction is between Crystallised Artificial Intelligence and Fluid Artificial Intelligence, and it adds axes such as cognitive level, technological type, learning method, functional scope, functional previsibility, functional behaviour, and functional purpose. The same work states that current AI devices are legally treated as goods rather than persons, and that responsibility remains with human actors such as programmers, manufacturers, operators, and users (Chassang et al., 2021).
These strands suggest that an AI impact matrix is best understood as a structured representation of multiple interacting dimensions, rather than as a single score. In some papers the matrix is classificatory, in others diagnostic, evaluative, or communicative.
2. Synthetic workload matrices for AI hardware benchmarking
The 2018 synthetic-benchmark paper introduces “AI Matrix” as a performance benchmark for DNN hardware, motivated by the heavy computation of architectures such as CNNs and the difficulty of selecting among CPUs, GPUs, and AI accelerators (Wei et al., 2018). The work identifies three drawbacks in benchmark suites such as BenchNN, DeepBench, and DawnBench: they cannot adapt to emerging changes of DNN algorithms and are fixed once selected; they contain tens to hundreds of applications and take very long time to finish running; and they are mainly selected from open sources and are therefore restricted by copyright and not representable to proprietary applications (Wei et al., 2018).
The proposed framework replaces a large static benchmark suite with one or a few synthetic DNNs derived from profiled workload characteristics. It has three main steps. First, an application monitoring system logs execution data with minimal overhead, including layer or operation details such as number of convolutional execution, input size, input channel, kernel size, and kernel stride. Second, workload analysis clusters operations using a nearest-neighbor approach. For CNNs, the emphasis is on convolutional layers, because convolution can consume only around 5% of memory but contributes around 90%–95% of inference computation. Third, a workload synthesizer uses a genetic algorithm to assemble synthetic networks for each group and then connect the groups into a final synthetic model (Wei et al., 2018).
The synthetic network construction is defined explicitly. A synthetic DNN is composed of groups, each group contains nodes, and each node is a convolutional operation. Within a group, nodes are connected one by one as filter size decreases, except where output size differs. If a group’s first node has multiple same-sized inputs, a concatenation operation is inserted. Each convolution is followed by batch normalization and ReLU, neighboring groups are connected by spatial pooling, and the final output is connected to an FC & softmax layer. The fully connected part is not encoded because it is less important in current models (Wei et al., 2018).
The matching objective combines three workload characteristics: the statistical distribution of convolutional layer parameters, computational workload measured as MAC, and launched warps on GPU. The paper states that the resulting CNN has a similar statistical distribution of layer parameters, computational workload and launched warps as the target models (Wei et al., 2018). In evaluation with AlexNet, VGG16, and GoogleNet, reported fitness errors include 0.00%, 0.01%, 0.02%, and a maximum of 1.19%. In a second experiment based on Alibaba platform logs, the data contained 608 convolutional operations, input sizes from 224×224 to 5×5, filter sizes from 6 to 1, and six workload groups; the resulting errors ranged from 0.00% to 0.96% (Wei et al., 2018). This usage of “matrix” is therefore a synthetic, profile-driven benchmark rather than a governance instrument.
3. Governance, reporting, and qualification matrices
In governance research, the matrix is a decision aid for determining material scope, risk profile, and appropriate precautionary measures. The 2024 comparison of Switch, Ladder, and Matrix emphasizes that governance responses should not be one-size-fits-all; instead, the classification of a use case should determine which preventive measures are needed (Mokander et al., 2024). The framework is especially useful when context, data quality, bias, model type, task, and output jointly determine ethical significance.
A more operational reporting approach appears in the co-designed AI Impact Assessment Report template. That template is grounded in the EU AI Act, NIST’s AI Risk Management Framework, and ISO/IEC 42001, and it is organized into five sections: information on the system’s use and teams, risks, mitigation strategies, benefits, and governance (Bogucka et al., 2024). Intended use is decomposed into five components—purpose, capability, domain, user, and subject—and risks are grouped into capability risks, human interaction risks, and systemic impact risks. The template also introduces a scaffold of 32 statements for populating the report, and its evaluation on an AI-based meeting companion identified sensitive data collection, potential privacy infringement, possible harm to public trust, and fairness concerns, together with mitigations such as anonymized data, representative training data, regular audits, and fairness checks (Bogucka et al., 2024).
The classroom study of existing AI impact assessments shows both the promise and the limitations of such instruments. It compares the U.S. CIO Algorithmic Impact Assessment, the Canadian Treasury Board Algorithmic Impact Assessment, and Microsoft’s Responsible AI Impact Assessment. The Canadian instrument is the most explicitly score-driven, with a raw impact score, a mitigation score, and four impact levels (Johnson et al., 2023). In the classroom study, 50% of participants changed their chosen most urgent value after completing an assessment, average concern increased from 4.00 to 4.32, and average responsibility assigned to machine-learning experts increased from 4.24 to 4.50 (Johnson et al., 2023). At the same time, participants reported opaque scoring, missing harms, poor fit for general-purpose systems, and vague mitigation guidance (Johnson et al., 2023). This suggests that matrix-like governance tools are effective at structuring reflection, but only if their categories, scoring logic, and workflow embedding are clear.
4. Audit, sustainability, and standards-oriented impact matrices
The 2026 Impact-AI-method is a qualitative, interview-based, external, ex post audit designed to evaluate concrete AI projects with respect to public interest and sustainability (Züger et al., 20 Jan 2026). Its normative basis is a regulatory dual concept: public interest and sustainability. The audit proceeds through three phases—pre-field research, field research, and post-field research—and the fieldwork requires at least four semi-structured interviews: AI Project Governance Interview, AI Project Sustainability Interview, AI System Facts Interview, and AI System Sustainability Interview. Across these interviews, the method covers 68 topics and 218 questions (Züger et al., 20 Jan 2026). After qualitative content analysis in MaxQDA, the project is assessed against 32 criteria in six clusters: Governance, Transparency and participation, Model and dataset facts, Social sustainability, Environmental sustainability, and Economic sustainability. The final output is an extended traffic-light model on a five-point scale and a “Visualization of the assessment criteria as matrix,” intended to make results publicly debatable (Züger et al., 20 Jan 2026).
A more dashboard-oriented sustainability formulation is the ESG Digital and Green Index. That instrument is described as a composite index, dashboard, self-assessment tool, and online application for assessing AI-related sustainability performance across four parts: Environmental ceiling, Social floor, Governance, and Transverse / global sustainability (Thelisson et al., 2023). The paper gives an explicit weighting scheme: 50% Environmental ceiling, 20% Social floor, 15% Governance, and 15% Global sustainability; within the environmental ceiling, 50% climate change, 35% natural resources, 10% pollution, and 5% biodiversity (Thelisson et al., 2023). The DGI is aligned with Doughnut Economics, planetary boundaries, SDGs, OECD AI Principles, the EU AI Act proposal, ISO 26000, and ISO/IEC TR 24368:2022 (Thelisson et al., 2023).
A third variant treats standards themselves as the intervention to be evaluated. The 2025 concept paper on AI standards does not present a literal AI Impact Matrix with fixed numeric cells, but proposes an analytical approach structured as inputs, activities, outputs, outcomes, and goals, combined with counterfactual-based impact evaluation (Lane, 16 Jun 2025). For AI standards, the paper maps working groups, workshops, research, and international engagement to outputs such as published standards, terminologies, taxonomies, metrics, and training data practices, and then to outcomes such as adoption, iteration, transparency, and risk mitigation, with long-run goals including innovation and trustworthy AI systems (Lane, 16 Jun 2025). This is a matrix-like theory of change rather than a scorecard.
5. Risk, harm, and scenario matrices
A prominent risk-oriented usage is the AI Mismatch framework, which is built from 774 AI cases and organized into seven 3×3 matrices (Saxena et al., 25 Feb 2025). Three are core decision matrices—Required Performance, Disparate Performance, and Cost of Errors—and four are supporting matrices—Data Quality, Model Unobservables, Expectation of Errors, and Error Detection/Mitigation. Each matrix compares a model-side condition against a task-side requirement or constraint, and risk increases as mismatch grows. The matrices are visualized as red, pink, yellow, and uncolored cells, and they are intended as a structured reasoning aid rather than a single score (Saxena et al., 25 Feb 2025). The case studies show that the framework can distinguish harmful from viable concepts by revealing when required performance exceeds what the data, model, and workflow can realistically support (Saxena et al., 25 Feb 2025).
Another harm-oriented matrix is AI Harmonics, which is a human-centric, harm-severity-adaptive framework built on incident data, stakeholder annotations, and ordinal severity rankings (Vei et al., 12 Sep 2025). Its data model aggregates harm category, stakeholder group, and frequency, and its main metric, , is an ordinal concentration measure inspired by the Gini coefficient. Using the AIAAIC dataset, the paper reports that Political / economic harms have the highest concentration, with AIH 0.85, followed by Physical harms and Psychological harms at 0.73; Financial / business and Autonomy are the least concentrated, with AIH 0.51 and 0.53 respectively (Vei et al., 12 Sep 2025). The framework also shows that users and vulnerable groups bear a disproportionate share of harms, especially in human rights / civil liberties, autonomy, and physical well-being (Vei et al., 12 Sep 2025).
Foresight research uses a different matrix logic. The scenario-generation methodology based on an Impact-Uncertainty Matrix places literature-derived themes according to expected societal or economic impact and degree of uncertainty (Costa et al., 31 Mar 2025). High impact / high uncertainty themes include AI and Digital Education, Renewable Energy and Sustainability, and Financial Markets and Fintech, while Molecular Medicine and Oncology, Machine Learning and Predictive Analytics, and Healthcare Systems and Public Health are placed in high impact / medium uncertainty (Costa et al., 31 Mar 2025). The matrix is then used to derive four scenarios: Optimistic Future, Technological Stagnation, Sustainability Focus, and Economic Downturn (Costa et al., 31 Mar 2025).
Perception research supplies a further matrix-like pattern. The Threats of Artificial Intelligence Scale measures perceived threat across four AI functional classes—Recognition, Prediction, Recommendation, and Decision-making—and three application domains—loan origination, job recruitment, and medical treatment (Kieslich et al., 2020). The scale is a four-factor model with 12 items, and the paper finds that Decision-making is the most threatening functionality in all domains, medical treatment is the least threatening domain overall, and job recruitment is the most threatening, especially for prediction and decision-making (Kieslich et al., 2020). This provides an empirically grounded function-by-domain map of perceived AI threat.
6. Labor, occupations, organizational maturity, and deployment measurement
Labor-market research has turned the matrix into a skill-displacement map. The 2026 skill-level AI Impact Matrix combines the Skill Automation Feasibility Index with real-world AI adoption patterns from the Anthropic Economic Index (Jadhav et al., 8 Apr 2026). The matrix has four quadrants: High Displacement Risk, AI-Augmented, Upskilling Required, and Lower Displacement Risk. SAFI benchmarks four frontier LLMs across 263 text-based tasks spanning all 35 O*NET skills, and the paper reports that Mathematics has SAFI 73.2 and Programming 71.8, while Active Listening has 42.2 and Reading Comprehension 45.5 (Jadhav et al., 8 Apr 2026). It also reports that 78.7% of observed AI interactions are augmentation rather than automation, and that the four models converge to similar skill profiles with only a 3.6-point spread (Jadhav et al., 8 Apr 2026).
Occupational exposure research uses a task-to-patent mapping rather than a quadrant view. The AI Impact measure for U.S. occupations embeds task descriptions and AI patents with Sentence-T5 and assigns task impact according to the maximum cosine similarity between a task and any patent (Septiandri et al., 2023). Occupation-level AII is then defined as the share of tasks whose best patent match exceeds the 90th percentile threshold (Septiandri et al., 2023). Using O*NET 24.3 and USPTO AI patents, the paper reports that highly impacted occupations include Cardiovascular Technologists and Technicians, Orthodontists, Medical Records and Health Information Technicians, Numerical Tool and Process Control Programmers, and Software Developers, Applications, whereas least impacted occupations are largely blue-collar and physical/manual roles (Septiandri et al., 2023). The paper explicitly argues that AI affects tasks rather than occupations directly, and that impact is not confined to routine tasks (Septiandri et al., 2023).
At the organizational level, the AI-CAM and AI-CM provide a maturity-and-capabilities matrix for AI adoption (Butler et al., 2023). The AI-CAM defines five maturity levels—Initial, R&D, Strategic, Defined, and Quantitatively Managed—across seven dimensions: Business, Data, Technology, Organisation, AI skills, Risks, and Ethical considerations (Butler et al., 2023). The AI-CM complements this with three proficiency levels—Basic, Advanced, and Expert—and covers machine learning, knowledge representation, deep learning, data engineering, software engineering, configuration management, continuous integration, cloud computing, virtual server computing, system administration, and ethical issues and risks (Butler et al., 2023). The framework is broader than many ML-only adoption models because it explicitly includes semantic technologies and AI systems that emulate human reasoning and decision-making (Butler et al., 2023).
A more deployment-oriented metric is the Intelligence Impact Quotient, which measures how deeply AI is embedded in organizational work rather than model capability itself (Rajah et al., 14 May 2026). IIQ defines a raw Intelligence Adoption Index as , combining novelty-weighted token stock, frequency of distinct-task engagement, a recency gate, organizational leverage, task complexity, and autonomy (Rajah et al., 14 May 2026). The raw measure is then mapped to a normalized 0–1000 IIQ score. The framework is explicit that it is not a direct measure of model capability and not a substitute for causal productivity evaluation (Rajah et al., 14 May 2026). This suggests an organizational impact matrix in which repeated, recent, non-redundant, high-leverage, complex, and autonomous AI use is distinguished from superficial experimentation.
Across these research programs, the AI impact matrix is less a single formalism than a recurring design pattern. It organizes heterogeneous evidence into a structured representation of inputs, dimensions, actors, or consequences, and it does so for different ends: benchmarking hardware, qualifying systems, guiding audits, prioritizing harms, generating scenarios, or mapping displacement and augmentation. The persistence of the matrix form suggests that AI impact is rarely reducible to one variable; in the cited literature, it is consistently treated as multi-dimensional, context-dependent, and tied to explicit choices about what counts as impact and for whom.