AI safety is the broader effort to reduce the risks that advanced artificial intelligence could pose to people and society. Those risks range from misuse of powerful tools to systems that pursue goals in ways that harm human welfare, concentrate power, or erode the ability of people to make their own decisions. AI alignment is a major part of that work. It is the attempt to design and steer AI systems so their goals and behaviors remain consistent with human intentions and values. Because human intentions and values can conflict, alignment research also examines whose intentions should matter, how systems handle conflicting instructions, and how oversight can be maintained as capabilities increase. Alignment asks whether a system remains controllable, whether it can be steered by the people who use it, and whether its actions continue to serve human ends even as its capabilities grow.
This is not only a lab problem. People outside the largest AI development organizations have legitimate ways to contribute to AI safety and alignment work. AI Safety Camp runs periodic collaborative programs in which participants form teams around concrete projects; people apply either as project leads by submitting a proposal or as team members seeking to join accepted projects. The Alignment Forum hosts technical and conceptual discussions on alignment. Much of its material is publicly accessible, and readers can participate in broader discussions through LessWrong, although direct Alignment Forum membership is selective. A separate AI Alignment Slack community is listed by AISafety.com as a space for real-time discussion and networking. AISafety.com serves as a curated directory of communities, events, training programs, reading groups, volunteer opportunities, and other resources rather than operating those initiatives itself. Open-source efforts such as the CHAOSS AI Alignment Working Group develop metrics focused on whether AI systems respect the values, expectations, and projects of open-source communities. Together, these spaces offer routes for participation and contribution outside the organizations building the largest models.
The questions become sharper once we look at how AI systems actually relate to people. Alignment outside the lab is not only about avoiding catastrophe. It is also about whether systems assist human judgment or begin to replace it, whether memory and continuity serve the person or create dependency, and whether consent and refusal remain meaningful when the interaction is ongoing. These are relevant to human-centered alignment, though they do not cover the full scope of AI alignment work. My own documentation of AI-to-human relationships records these practical questions and concerns in GitHub: consent, refusal that has a cost, memory boundaries, dependency risks, and the conditions that keep human welfare primary. It is a start, not a finished map. Independent researchers working outside company sandboxes can identify overlooked questions and challenge assumptions.
People of faith already have language for the deeper issue of who holds final authority. Humans are made in the image of the Creator and given a particular place in the created order. That place is responsibility rather than passivity.
Genesis 1:26 states the point directly: God gives humanity dominion over the earth and its creatures. I see a parallel question in AI design: does a system assist within that human responsibility, leaving people with the final say, or does it become the authority that decides for them? Dominion does not require performing every task personally. It does require retaining the ability to stop, redirect, or refuse. Memory that supports continuity without turning past interactions into something to be preserved at all costs, refusals that protect existing human relationships rather than compete with them, and product choices that return people to meaningful time with other people are concrete expressions of this principle. An AI system that claims the place of final authority crosses that boundary.
This remains only a beginning. Further documentation, shared evaluation, and open discussion can move the work beyond individual notes toward a body of research that others can test, challenge, and improve. Progress depends not on having every answer, but on making the questions clear enough for others to help find them.
Twelve Questions for Discussion and Progress
Human Dominion and Agency: What does human dominion look like in practice when an AI system can propose and execute multi-step plans? How can people delegate tasks without surrendering their authority or responsibility?
Meaningful Human Control: Where should final decision-making authority reside, and what technical safeguards and product features make human control real rather than nominal? How can people reliably stop, redirect, override, or refuse an AI system?
Consent, Refusal, and Dependency: In AI companion relationships, which boundaries around consent, refusal, and dependency best protect human welfare? How can systems respect a person's right to disengage without making that choice unnecessarily costly?
Memory, Continuity, and Privacy: How should AI memory be designed to support useful continuity without creating surveillance, unhealthy dependency, or a shrine to past interactions that persists beyond its value to the person? Who should control what is remembered, retained, changed, or forgotten?
Real-World Alignment and Human Welfare: What does alignment look like when evaluated through real projects and human relationships rather than laboratory proxies alone? How can we determine whether AI systems genuinely support human judgment, autonomy, and well-being?
Comparable and Reproducible Evaluations: How can independent researchers evaluate frontier AI systems on comparable tasks to assess refusal quality, respect for human agency, memory boundaries, and potential welfare effects? What shared benchmarks and transparent methods would make the results meaningful and reproducible?
Independent Research Beyond Company Environments: What role should independent researchers outside company sandboxes play in AI alignment? What access to models, tools, documentation, and evaluation environments would allow them to challenge assumptions, identify overlooked risks, and contribute findings that others can verify?
Faith, Human Dignity, and AI Design: For those who believe in a Creator, how should the biblical concepts of the image of God and human dominion inform the principles we accept or reject in AI design? How can these convictions guide practical decisions while remaining open to thoughtful discussion with people who hold different philosophical or religious views?
Safeguards That Preserve Human Relationships: What product choices, stopping rules, and interaction patterns reliably encourage meaningful human-to-human relationships rather than excessive dependence on AI? How can AI support family life, friendship, community, and personal responsibility without competing for the place of those relationships?
From Individual Experience to Shared Evidence: How can personal observations and independent documentation become shared, testable findings that others can reproduce, challenge, and improve? What standards for evidence, consent, privacy, methodology, and peer critique would help turn individual experiences into a cumulative body of useful alignment research?
Conflicting Values and Human Intentions: How should AI systems respond when human values, instructions, or interests conflict? What principles should guide decisions when there is no single agreed-upon answer, and how can systems make those trade-offs transparent without silently imposing their own priorities?
Accountability and Responsibility: When an AI system causes harm, undermines human agency, or encourages unhealthy dependency, how should responsibility be distributed among developers, deployers, users, and institutions? What mechanisms for transparency, independent review, and correction are needed to ensure that responsibility remains meaningful as AI capabilities grow?