Published: June 2026 Technical Level: Advanced Category: Artificial Intelligence
The deployment of artificial intelligence in power system protection, planning, and operations raises ethical questions that are substantively different from those arising in consumer AI applications. In power engineering, AI errors translate directly into equipment damage, extended outages, arc flash hazards, and, in the most serious cases, fatalities. The engineer of record who incorporates AI-generated outputs into a licensed design retains full professional liability for those outputs, irrespective of the tool's vendor claims. This paper examines the ethical obligations of the practicing power engineer in relation to AI tools, develops a framework for evaluating when AI outputs meet the professional standard of care, and identifies the categories of engineering judgment that must remain with the human engineer regardless of AI capability. The paper draws on the NSPE Code of Ethics, IEEE Code of Ethics, ASCE principles, and the emerging IEEE P2863 standard for organizational governance of AI in engineering to ground the analysis in enforceable professional obligations rather than aspirational principles.
Engineering ethics has always grappled with the question of how engineers should use tools whose outputs they did not personally derive. A relay coordination software package produces operating times by evaluating a mathematical characteristic equation against user-entered parameters — the engineer trusts this output because the equation is published, the parameters are entered by the engineer, and the calculation can be independently verified at any point. The ethical obligation is clear: verify the tool's output for the specific case before relying on it in a licensed design.
AI tools complicate this obligation in ways that are not merely a matter of degree. A neural network trained on thousands of relay event records produces a fault classification output through a process that is, in general, not directly interpretable: the engineer cannot inspect the intermediate steps to verify that the classification is correct for this specific event record in the way they can verify a mathematical calculation. The tool's accuracy on historical validation data provides statistical assurance, but statistical assurance is not the same as verification — the tool may perform correctly on 97 percent of cases while systematically failing on a specific subset of cases that includes the case at hand.
This interpretability gap creates a genuine ethical tension. The engineer has a professional obligation to verify the correctness of work bearing their seal. If verification of an AI output is not technically feasible — because the model's internal representation is not transparent — the engineer must either choose a verifiable alternative or accept a higher level of independent checking as the substitute for direct verification.
The National Society of Professional Engineers Code of Ethics states, in its first fundamental canon, that engineers shall "hold paramount the safety, health, and welfare of the public." The third canon states that engineers shall "issue public statements only in an objective and truthful manner" and, by extension, that technical decisions incorporated into licensed designs must be supportable by objective analysis.
Applied to AI tool use, these canons establish that: (1) the engineer may not use an AI tool to produce design outputs for which no verification method exists, if those outputs affect public safety; and (2) the engineer must be able to provide an objective technical basis for every decision in a licensed design — "the AI recommended it" is not an adequate basis.
The IEEE Code of Ethics, Principle 1, commits members to "hold the safety of the public, employees, and the environment paramount" and to "promptly disclose factors that might endanger the public or the environment." Principle 7 commits members to "seek, accept, and offer honest criticism of technical work, to acknowledge and correct errors, and to credit properly the contributions of others."
Applied to AI, IEEE Principle 7 is particularly relevant. An engineer who publishes or presents AI-generated technical analysis without disclosing the AI's role and limitations fails this principle. The growing practice of incorporating AI outputs into technical reports without disclosure of model uncertainty, training data limitations, or known failure modes is an ethical issue as well as a technical one.
IEEE P2863, currently under development, will provide the first formal standard for the organizational governance of AI systems in engineering practice. Its draft framework identifies five governance requirements: transparency (the AI system's decision logic must be explainable to the degree necessary for the engineering context), accountability (a human engineer must be named as responsible for each AI-assisted decision), accuracy (the AI system must be validated against a representative test set before deployment), monitoring (ongoing performance monitoring must detect when the AI system's accuracy degrades), and fairness (the AI system must not systematically disadvantage specific geographic or demographic groups in infrastructure planning decisions).
The fairness requirement deserves specific attention in power system planning. AI models used for infrastructure investment prioritization — deciding which feeders to upgrade, which neighborhoods to harden against climate events, which customers to prioritize for outage restoration — are trained on historical data that may reflect historical inequities in infrastructure investment. A model trained on historical maintenance schedules will reproduce whatever geographic patterns those schedules contained, potentially perpetuating underinvestment in specific communities. This is an ethical issue with direct physical consequences for the reliability of service to those communities.
Not all engineering decisions have the same consequences for public safety, and not all AI applications have the same interpretability characteristics. A tiered evaluation framework matches the verification requirement to the safety consequence:
Tier 1 — Life-Safety Critical (real-time protection, arc flash calculations, structural adequacy): AI outputs are not acceptable as the primary basis for design decisions. The AI may serve as a screening tool to identify cases requiring analysis, but every Tier 1 decision must be verified by conventional analytical methods with documented calculations.
Tier 2 — Reliability Critical (relay settings, protection coordination, load flow adequacy): AI outputs are acceptable as the initial input to the engineer's analysis, provided that: the model has been validated on a test set representative of the specific application; the engineer can explain the physical basis for the output; and the output has been independently checked against a hand calculation or an alternative analytical method for at least a representative sample of cases.
Tier 3 — Efficiency and Optimization (energy management dispatch, demand forecast, equipment selection screening): AI outputs are acceptable as the primary basis for operational decisions provided that: model performance is continuously monitored; fallback rules ensure safe operation if the AI system fails; and the economic consequences of errors are bounded and reversible.
The distinction between Tier 1 and Tier 3 is not arbitrary: Tier 1 decisions must be correct on the first occurrence of the relevant event, before any opportunity to learn from failure. Tier 3 decisions operate in a feedback loop where errors are detected, their consequences are bounded, and the model can be corrected. The ethical obligation to verify independently is strongest where the feedback loop is absent.
For Tier 2 applications, the interpretability requirement means that the engineer must be able to articulate, in plain engineering language, why the AI system produced a specific output — not merely that the system's average accuracy on the validation set was high. Methods that provide post-hoc explanations for specific predictions — SHAP (SHapley Additive exPlanations) values for gradient boosting models, attention weight visualization for transformer models, saliency maps for convolutional neural networks — provide a partial answer to this requirement but are not perfect substitutes for mathematical derivation.
An engineer who can examine the SHAP values for a relay misoperation classification and confirm that the model's prediction is driven by the physically expected features (elevated second harmonic, dc offset, directional element operation) has a stronger basis for accepting the output than one who accepts the classification solely on the basis of the model's aggregate accuracy. This is the practical expression of the ethical obligation: use the interpretability tools available, apply engineering judgment to assess whether the explanation is physically coherent, and document that assessment.
When an AI-assisted design is found to be deficient — whether in a regulatory review, a post-incident investigation, or a professional liability proceeding — the question of accountability is straightforward: the engineer of record who sealed the design is accountable. The AI tool's vendor is not accountable, because the professional licensing framework places responsibility on the engineer who chooses to use a tool and accepts its outputs. The engineer's employer is not accountable for the technical decisions, except in the narrow sense of employer vicarious liability for employee actions.
This is not merely a legal technicality — it reflects the ethical structure of professional engineering. Licensure exists to create a named, accountable individual responsible for the public safety implications of each design. AI tools do not have licenses and cannot be accountable. The engineer who uses an AI tool without independent verification has not discharged their ethical obligation; they have delegated it to an entity that cannot accept it.
The professional standard of care for AI-assisted design requires that the engineer's project file document: which AI tools were used and for which specific decisions; the validation data and performance metrics of each tool; the independent checks performed on AI outputs; and the engineer's professional judgment on each case where the AI output was accepted or overridden. This documentation requirement is not merely defensive — it is the mechanism by which the engineering profession maintains accountability and learns from cases where AI outputs were incorrect.
AI models used for feeder upgrade prioritization, vegetation management scheduling, and outage restoration sequencing are trained on historical operational data. If historical practice included systematic differences in service quality, maintenance frequency, or restoration priority across geographic areas, the AI model trained on that data will reproduce those differences in its recommendations. Unlike a human planner who can be asked to explain and justify a decision that appears inequitable, an AI model's recommendation may be difficult to scrutinize for equity impacts without explicit tooling for distributional analysis.
The ethical obligation to identify and address algorithmic bias in infrastructure planning is reinforced by the regulatory framework: the Federal Energy Regulatory Commission's Order 2222 and state-level integrated resource planning requirements increasingly include explicit equity provisions. An AI planning tool that produces recommendations reflecting historical inequity will produce outputs that fail these regulatory requirements as well as ethical standards.
The practical mitigation for algorithmic bias in infrastructure planning is explicit distributional analysis of AI outputs: before deploying a model for investment prioritization, analyze the geographic and demographic distribution of the model's recommendations and compare it against the distribution of service quality metrics. If the model's recommendations would concentrate investment in areas already receiving better service, the training data, feature set, or optimization objective must be modified before deployment.
This analysis does not require abandoning AI-assisted planning — it requires that the engineer, acting as the accountable professional, examine the equity implications of AI outputs with the same rigor applied to their technical accuracy.
The most consequential conclusion for the licensed engineer is that AI tools introduce no new ethical category — they intensify an existing one. The duty to verify life-safety-critical outputs, to confirm that a result is physically coherent rather than merely statistically plausible, and to retain personal accountability for every sealed deliverable all predate AI; what changes is the ease with which a confident, well-formatted, wrong answer can pass unexamined. The interpretability check — confirming the output is physically coherent, not just accurate on aggregate validation data — is the new discipline these tools demand.
The most common failure is documentary rather than analytical: engineers validate AI output informally, satisfy themselves it is correct, and record nothing, leaving no reconstructable basis for the judgment when the design is later challenged. A validation that is not documented did not, for professional-accountability purposes, occur.
The next problem the profession must confront is equity in AI-assisted planning, where models trained on historical investment patterns can encode and perpetuate the underinvestment of disadvantaged areas under the appearance of objective optimization. Establishing a standard practice of analyzing AI planning recommendations for distributional consequences — before deployment, not after audit — is the unfinished ethical work that the technical validation framework does not by itself address.
The analysis in this paper connects to several companion studies in this library. Readers concerned with the upstream and downstream engineering will find AI Risk Assessment for Power Systems Engineers develops a closely related aspect of the same problem, while AI in Electrical Design Automation extends the treatment into an adjacent domain. For the broader methodological context, The AI Transformation of Power Systems Engineering provides complementary depth.
[1] National Society of Professional Engineers, NSPE Code of Ethics for Engineers, NSPE Publication 1102, 2019.
[2] IEEE, IEEE Code of Ethics, IEEE, 2020.
[3] IEEE P2863, Recommended Practice for Organizational Governance of Artificial Intelligence, IEEE Standards Association, 2026 (draft).
[4] S. M. Lundberg and S. Lee, "A unified approach to interpreting model predictions," Advances in Neural Information Processing Systems, vol. 30, 2017.
[5] Federal Energy Regulatory Commission, Order 2222: Participation of Distributed Energy Resource Aggregations in Markets Operated by Regional Transmission Organizations and Independent System Operators, FERC, 2020.
[6] B. Christian, The Alignment Problem: Machine Learning and Human Values, W. W. Norton, 2020.
[7] NSPE, Engineering Ethics Reference Guide, National Society of Professional Engineers, 2023.
[8] IEEE, Ethically Aligned Design: A Vision for Prioritizing Human Well-being with Autonomous and Intelligent Systems, IEEE, 2019.
[9] NFPA 70E-2024, Standard for Electrical Safety in the Workplace, NFPA, 2024.
[10] T. Gebru et al., "Datasheets for datasets," Communications of the ACM, vol. 64, no. 12, pp. 86–92, 2021.
[11] J. L. Blackburn and T. J. Domin, Protective Relaying: Principles and Applications, 4th ed., CRC Press, 2014.
[12] IEEE Standard 1547-2018, Standard for Interconnection and Interoperability of Distributed Energy Resources, IEEE, 2018.