AI for people and planet begins with a question that technical performance metrics cannot answer: who benefits from an artificial intelligence system, who bears its costs, and who has the power to challenge its decisions?
Artificial intelligence can support medical research, environmental monitoring, accessibility, education and public services. But these potential benefits do not eliminate the social and environmental consequences of developing and operating AI.
An AI system may perform its assigned task accurately while relying on questionable data practices. It may improve productivity while shifting costs onto workers. It may support environmental research while consuming substantial computing resources.
Responsible design requires examining these consequences together.
The objective is not to make every AI system perform every task with the fewest possible resources. It is to ensure that the technology serves a legitimate purpose, respects affected people and operates within clearly defined safeguards.
That requires decisions about data, labour, infrastructure, human oversight and accountability long before a model reaches production.
1. Define the Public Value Before Choosing the Technology
Responsible AI development should begin with the problem being addressed, not the capabilities of a particular model. An organisation may want to improve access to information, reduce administrative delays or identify environmental risks. These are objectives that can be evaluated independently of the technology chosen to achieve them.
The next question is whether AI is necessary. Some problems are better addressed through improved services, clearer information, conventional software or organisational changes. Introducing machine learning can create new dependencies involving data quality, computing infrastructure, maintenance and specialist expertise.
A responsible project therefore compares AI with realistic alternatives. It identifies intended beneficiaries, possible harms and the conditions under which the proposed system would represent a genuine improvement.
A People-and-Planet Impact Statement
Before development, document:
1)The problem the system is intended to solve.
2)The people and communities likely to be affecte
3)The expected benefits and supporting evidence.
4)The non-AI alternatives considered.The data and infrastructure required.
5)The foreseeable risks and environmental considerations.
6)The individuals or teams accountable for decisions.
7) The conditions under which deployment should stop.
This creates a reference point against which the project can be evaluated.Without a defined purpose, technical achievement can easily be mistaken for social value.
2. Build Data Governance Around Rights and Context
AI development often depends on large collections of information. These may include personal records, publicly available content, licensed datasets, sensor readings or material contributed by communities.
The availability of information does not automatically establish that every use is lawful, ethical or appropriate. Public accessibility should not be confused with unrestricted permission to collect, reproduce or use content for model training.
Data governance must address provenance, lawful processing, licensing, privacy, security, representativeness and the expectations of affected individuals. It should also recognise that different categories of information create different obligations.
Consent Is Important, but Not the Entire Framework
Consent may be an appropriate legal basis for some data-processing activities, but it is not the only possible lawful basis under every privacy regime.
The applicable requirements depend on jurisdiction, purpose and the nature of the information.
Responsible practice involves clear communication, purpose limitation, appropriate retention periods and meaningful mechanisms for exercising relevant rights.
For community-held and culturally sensitive information, legal compliance alone may be insufficient.
Indigenous communities, for example, may have collective interests in the governance of cultural knowledge and data that cannot be adequately addressed through ordinary individual-consent procedures.
Relevant approaches include the CARE Principles for Indigenous Data Governance, which emphasise collective benefit, authority to control, responsibility and ethics.
The practical question is not simply whether information can be obtained.
It is whether the proposed use respects the people and communities connected to it.
3. Account for the Human Labour Behind AI
AI systems depend on human work throughout their lifecycle. People collect and prepare datasets, label information, evaluate model responses, moderate harmful material, test applications and investigate failures.
Some of this work involves repetitive tasks or exposure to disturbing content. Employment arrangements may differ across organisations, contractors and jurisdictions, making labour conditions difficult for customers to assess.
Responsible procurement should therefore extend beyond model performance and subscription cost. Organisations need to consider the treatment of workers involved in developing and maintaining the systems they use.
Questions for AI Suppliers
Ask whether suppliers can explain:
1)How data annotation and evaluation work is organised.
2)Whether workers receive appropriate compensation.
3)What safeguards exist for people reviewing harmful content.
4)How contractors can raise concerns.
5)What training and support workers receive.
6) Whether relevant labour standards are monitored.
7)How subcontractors are included in supplier oversight.
8) Not every supplier will disclose detailed information.
However, limited transparency should be treated as a governance concern rather than evidence that no problem exists.
AI cannot be considered responsible solely because its interface appears automated and efficient.
4. Design for Resource Constraints and Environmental Accountability
AI systems require physical infrastructure. Computing hardware, data centres, electricity, cooling systems, networking equipment and hardware manufacturing all contribute to their resource demands.
The environmental implications depend on the system being developed and how it is used. Training a foundation model, fine-tuning an existing model and running a small inference service involve different activities and accounting considerations.
Design teams should assess the resources required by the complete application, including repeated requests, supporting databases, monitoring systems and the expected period of operation. Where environmental estimates are produced, the methodology and limitations should be documented.
Efficiency Without Exclusion
Resource-efficient design can include using appropriate models, reducing unnecessary computation and reusing validated outputs.
But environmental optimisation should not undermine essential functionality.
A low-bandwidth application may be valuable in areas with unreliable connectivity. Offline capability may improve access for some users. However, these features require careful engineering and may introduce other constraints involving device storage, updates and security.
Similarly, smaller models may be appropriate for some tasks but insufficient for others.
The relevant objective is not minimum computing at any cost.
It is delivering a reliable and useful service with justified resource requirements.
For detailed operational practices, this topic connects to our separate article, AI Efficiency as a Brand Value: How You Train, Host, and Deploy AI Will Matter.
5. Make Accountability Part of the System Architecture
AI systems can produce incorrect, biased or inappropriate outputs. The severity of these failures depends on the application and the decisions affected.
A mistake in a low-risk creative suggestion is different from an error influencing medical care, access to employment or eligibility for a public service.
Accountability should therefore be proportionate to risk. High-impact applications require clear responsibilities, appropriate testing, monitoring and procedures for intervention.
Human Oversight Must Be Operational
Simply requiring a person to approve an AI-generated decision does not guarantee meaningful oversight.
The reviewer must understand the relevant limitations, have access to necessary information and possess genuine authority to intervene.
1)Effective oversight may require:
2)Clear escalation procedures.
3)Appropriate reviewer training.
4)Records of consequential decisions.
5)Accessible ways to contest outcomes.
6)Monitoring for recurring errors.
7)Procedures for suspending unsafe functionality.
8)Independent review where warranted.
9) These safeguards should be designed before deployment.
An organisation cannot create accountability merely by stating that humans remain responsible.
The AI Responsibility Lifecycle
A responsible AI programme needs controls across the complete lifecycle.
Problem definition: Establish the intended purpose, affected groups and realistic alternatives.
Data acquisition: Review provenance, permissions, quality and relevant community interests.
Model development or selection: Evaluate suitability, reliability, bias and resource requirements.
Testing: Examine representative conditions, foreseeable misuse and high-consequence failures.
Deployment: Establish access controls, monitoring, documentation and human oversight.
Operation: Track incidents, performance changes, complaints and unintended consequences.
Review: Reassess whether the system continues to provide sufficient value.
Retirement: Manage data, access, dependencies and records when the system is withdrawn.
The final stage is frequently overlooked.
AI systems should not remain in operation indefinitely merely because they have already been implemented.
A Practical Responsible AI Design Checklist
1. Establish the Purpose
Define the problem and the expected benefit.
Avoid introducing AI simply because it is available.
2. Identify Affected Stakeholders
Consider users, non-users, workers, communities and other people influenced by the system.
Assess whether particular groups face disproportionate risks.
3. Review Data Provenance
Document where information originated and whether its use is appropriate.
Address privacy, licensing and relevant collective interests.
4. Evaluate Labour Conditions
Examine available information about annotation, evaluation and content moderation.
Include supplier responsibilities in procurement decisions.
5. Compare Technical Alternatives
Evaluate AI and non-AI approaches.
Choose technology according to the actual task.
6. Assess Environmental Requirements
Document available information about computing, infrastructure and resource consumption.
Distinguish estimates from direct measurements.
7. Test for Unequal Outcomes
Examine performance across relevant populations and operating conditions.
Investigate disparities rather than relying solely on average accuracy.
8. Design Meaningful Oversight
Give responsible people sufficient information and authority to intervene.
Provide routes for escalation and contesting consequential decisions.
9. Prepare for Failure
Define fallback procedures, incident reporting and suspension criteria.
Test what happens when data, connectivity or model outputs become unreliable.
10. Review the System After Deployment
Monitor whether the intended benefits actually occur.
Retire or redesign systems that cannot meet acceptable requirements.
Three Examples of Responsible AI Design
Example 1: A Community Flood-Warning Service
A regional organisation proposes using AI to support flood-risk forecasting.
The project begins by consulting local emergency authorities and affected communities.
The team examines whether the proposed forecasts provide useful additional information compared with existing warning systems.
Warnings are designed for relevant languages and communication channels.
The system includes procedures for communicating uncertainty and handling outages.
Success is evaluated through warning accessibility, preparedness and operational usefulness—not merely forecasting accuracy.
Example 2: An AI-Assisted Recruitment System
An employer wants to use AI to help organise job applications.
The system is restricted to clearly defined administrative tasks rather than automatically rejecting candidates based on opaque predictions.
The organisation reviews the relevance of input data, tests for unequal outcomes and provides human review of consequential decisions.
Applicants receive appropriate information about the process and available avenues for raising concerns.
The employer remains accountable for recruitment decisions.
Example 3: An Environmental Research Assistant
A nonprofit develops an AI assistant to help researchers navigate conservation reports and scientific literature.
The application retrieves information from an approved document collection and links its responses to sources.
The team evaluates whether a simpler search interface could achieve the same purpose.
Researchers verify important findings before publication.
The organisation monitors errors, access requirements and operating costs.
The system supports research without being presented as an independent scientific authority.
These examples are illustrative design scenarios, not claims about existing deployed projects.
Where Responsible AI Programmes Commonly Fail
Ethics without enforcement
Publishing principles does not establish that employees or suppliers follow them.
Responsibilities, review procedures and consequences for noncompliance are necessary.
Deployment before adequate testing
Early demonstrations may not reflect real operating conditions.
Systems need evaluation appropriate to the consequences of failure.
Unclear accountability
When responsibility is distributed across developers, suppliers and customers, important decisions may go unchallenged.
Contracts and governance arrangements should clarify responsibilities.
Environmental claims without evidence
Describing AI as sustainable because it supports an environmental objective is insufficient.
Resource demands and outcomes require separate evaluation.
Ignoring workers and communities
An AI system may benefit its direct customers while shifting costs onto less visible groups.
Stakeholder assessment must extend beyond the immediate user.
Treating oversight as a formality
Human review is ineffective when reviewers cannot understand or challenge decisions.
Oversight must be supported by appropriate system design.
The Role of Standards and Regulation
Responsible AI design is increasingly supported by formal governance frameworks.
The NIST AI Risk Management Framework provides a voluntary approach to identifying and managing AI-related risks.
The OECD AI Principles address trustworthy AI, human rights, transparency, robustness and accountability.
The EU AI Act establishes a risk-based legal framework with obligations that depend on the type of AI system and the applicable implementation timetable.
The CARE Principles for Indigenous Data Governance provide guidance concerning collective interests and ethical data governance.
These frameworks serve different purposes.
A voluntary risk-management framework is not equivalent to a binding legal obligation.
Organisations should identify the rules applicable to their activities and jurisdictions rather than treating a general ethical checklist as proof of compliance.
How to Measure Responsible AI in Practice
A useful assessment combines technical, social and operational evidence.
Effectiveness: Does the system solve the intended problem?
Fairness: Are there material differences in outcomes across relevant groups?
Accessibility: Can intended users access and operate the system?
Agency: Can people understand, question or contest consequential outcomes?
Data governance: Are information sources and processing arrangements appropriate?
Labour: Are relevant worker protections and supplier responsibilities addressed?
Reliability: How does the system perform under realistic conditions and failure scenarios?
Environmental performance: What resource demands can be measured or reasonably estimated?
Accountability: Who is responsible when something goes wrong?
No single score can fully represent these dimensions.
A system may perform well in one area while presenting unacceptable risks in another.
Responsible evaluation requires examining those trade-offs explicitly.
Conclusion: Responsibility Must Be Designed, Not Declared
AI systems affect more than the people interacting directly with their interfaces.
They depend on data, workers, infrastructure and institutional decisions. Their consequences can extend to communities, public services and the environment.
Designing AI for people and planet therefore requires a wider understanding of what constitutes success.
Accuracy matters, but so do fairness, access, privacy, labour conditions, resource use and the ability to challenge harmful outcomes.
Responsible AI is not achieved by adding an ethics statement after development.
It is built through decisions about what should be automated, what information may be used, whose interests are represented and who remains accountable.
The objective is not to create technology that claims to be harmless.
It is to build systems whose benefits, limitations and consequences can be examined—and whose operators are prepared to change course when the evidence demands it.