Professional responsibility and technical decisions

Prerequisites: Software requirements and basic risk analysis.

Learning objectives

• Identify affected stakeholders.
• Write an evidence-based impact assessment.
• Document a reasoned go or no-go decision.

A technical specification is a choice about people

Software changes what people can do, what institutions can observe, and whose work becomes easier or harder. Professional responsibility includes competence, honesty about limitations, respect for privacy, and attention to foreseeable harm. These concerns belong in design and review, not only in a disclaimer after deployment.

Identify stakeholders beyond the paying customer: users, non-users whose data is processed, maintainers, workers affected by automation, and people who cannot access the system. Benefits and burdens may fall on different groups. A technically feasible feature can still be inappropriate for its context.

Work a consequential design review

Suppose a school proposes ranking students by an automated “effort score” derived from time online and clicks. The score is easy to compute. Its validity is another question. Students with slow connections, assistive technology, or offline study habits may produce different traces for reasons unrelated to effort. The system may reward leaving a page open or discourage efficient study.

A review should ask what decision the score will influence, what evidence links the proxy to the intended construct, which errors matter, and whether a less intrusive method can meet the goal. If the score affects access or discipline, demand stronger evidence and meaningful challenge mechanisms. Replacing the label “effort” with “activity” does not remove the consequences of how it is used.

Document uncertainty and alternatives

An impact assessment should state the purpose, affected groups, data sources, expected benefits, plausible harms, mitigations, and remaining uncertainty. Compare alternatives, including a simpler system or no automated decision. Assign responsibility for monitoring and specify conditions that trigger revision or withdrawal.

Honesty includes reporting negative results and limitations that a stakeholder might prefer to omit. Do not claim a model is unbiased because one aggregate metric improved. Do not describe a self-assessed course badge as independent evidence of professional competence.

Responsibility persists through operation

Create channels for reporting errors and harm. Make reports actionable by preserving enough context, protecting reporters where appropriate, and giving issues an owner. Monitor whether real use matches the assumptions in the original decision. A system deployed for one purpose may be reused in a more consequential setting.

Ethical reasoning does not reduce to a single formula. It requires explicit values, evidence, consultation, and accountability. Engineers should be able to explain why a decision is justified, what would change their judgment, and who has authority to stop or correct the system. That explanation is part of the technical record.

Practice problem

Write a go/no-go memo outline for an activity-based student ranking. Include two alternatives, a claim that requires evidence, and a condition that would stop deployment.

Worked solution

State the proposed decision and who is affected, then separate measured activity from inferred effort or ability. Alternatives include private learner-controlled progress summaries and teacher-supported reflection without ranking. The claim that clicks or time online measure effort requires validation across relevant contexts and access needs. Review privacy, error consequences, appeal, and monitoring. A reasonable stop condition is lack of evidence that the score measures the intended construct, especially when it would affect access or discipline. Record who owns the decision and what evidence could justify reconsideration.

Check your understanding

Why is a technically accurate click count insufficient evidence of effort?

• Clicks are impossible to count.
• The measured proxy may not represent the intended human construct.
• All proxies are always useless.
• Only grades can be measured.

Explanation

Measurement accuracy and construct validity are different questions.

Which belongs in a responsible deployment decision?

• Only implementation cost.
• Only expected revenue.
• Benefits, harms, alternatives, uncertainty, and accountable ownership.
• A promise that no error will occur.

Explanation

The decision must address consequences and the limits of its evidence.

Primary reading

• Society, Ethics, and Professionalism: ACM/IEEE-CS/AAAI CS2023

Open the study and save your work →