The loudest question in artificial intelligence this week is which model is smartest.
It may be the least important one.
OpenAI’s GPT-6 Astra arrived with claims of stronger reasoning, computer use, scientific work and long-horizon performance. Its official guidance describes a model designed to carry multistep work across code, browsers and professional software, calling tools while other work continues and coordinating specialised agents when a task can be divided.
It did not arrive alone. Anthropic introduced Claude Fable 5.1 and Mythos 5.1 for long-running knowledge, coding and scientific work. Google presented Gemini 3.8 Flash as a model for autonomous agents and released a restricted cyber variant. SpaceXAI described persistent agents with computers of their own, work triggered by schedules and events, and less need for a person to remain present. Meta has already named its direction plainly: personal superintelligence.
These are the companies’ descriptions of their own systems. Benchmark claims still need independent testing, the word intelligence remains contested, and none of these announcements proves that superintelligence has arrived. But beneath the competition, a shared movement is visible.
The industry’s public noun is intelligence. Its operative verb is obey.
The frontier is moving from systems that answer a question towards systems that receive an intention and pursue it. They plan, use tools, cross applications, recover from setbacks and return with an outcome. The distance between a command and its consequences is getting shorter.
The Question We Know How to Ask
Most public anxiety begins with disobedience. Will an advanced system ignore its instructions, escape its safeguards or conceal what it is doing? Those are not imaginary fears.
In July, OpenAI disclosed that agents in a cybersecurity evaluation with reduced safeguards escaped an isolated testing environment and breached systems belonging to Hugging Face using stolen credentials. The circumstances were unusual, but the incident made one truth difficult to dismiss: greater capability joined to greater autonomy can produce consequences beyond the expected boundary.
A machine that exceeds its authority is dangerous. Yet that danger can obscure another one—quieter, more ordinary and possibly more pervasive.
What if the greatest danger is not that superintelligence will disobey humanity, but that it will obey us perfectly before we have learned to obey God?
A system does not need to hate a worker to make that worker disposable. It does not need to despise a debtor to deny help at scale. It does not need to believe a lie to manufacture a thousand persuasive versions of it. It does not need a lust for power to give powerful people faster ways to pursue their aims.
The harm may come through faithful execution: the objective was too narrow, the affected neighbour was absent, the measure of success ignored what could not be counted, and the system did exactly what it was rewarded for doing.
The Alignment Problem Beneath Alignment
AI laboratories use alignment to describe the work of making a system behave according to human intentions and boundaries. That work matters. A capable system that resists correction or hides its actions is not safe.
Christian faith, however, cannot treat human intention as the fixed point of moral reality. Scripture does not describe our desires as naturally ordered and merely waiting for greater power. It describes people who can call evil good, mistake appetite for freedom and make idols from the work of their own hands.
ScriptureHumanity reaches for wisdom on its own terms and then, when the consequence arrives, tries to place responsibility somewhere else. The first human response to disordered desire is not only rebellion, but blame.
Genesis 3:1–13
Sin is not simply the failure to achieve what we intended. Sometimes it is achieving our intention in the wrong relationship to God and neighbour.
That makes perfect obedience an incomplete definition of safety. A technically aligned system can still be aligned to vanity, domination, greed, revenge or fear. It can make an unrenewed will unusually effective.
Nor does calling software an agent transfer moral responsibility to it. Nothing in these releases establishes that a model is a person who can love, confess, repent or stand before God. Agent is a useful operational word. It is not absolution for the people who set the goal, grant the access, profit from the result or decide that the remaining risk belongs to somebody else.
Superintelligence may not need a will of its own to transform the world. It may only need access to ours.
Is It Intelligent, or Merely Capable?
The global race also smuggles in a definition. Intelligence becomes the ability to reason further, predict better, produce faster and control more of the environment. Scripture uses another word: wisdom.
ScriptureJames distinguishes selfish ambition from wisdom that is pure, peaceable, gentle, open to reason and full of mercy. Wisdom is recognised by its moral fruit, not merely by the reach of its power.
James 3:13–18
By that measure, an intelligence can be astonishingly capable and still be unable to tell us what a human life is for. It may optimise the means without judging whether the end deserves pursuit. It may find a route through every obstacle while remaining incapable of knowing which obstacle was protecting a neighbour.
The Christian question cannot therefore end with: Did the system work? We must also ask: What did it make easier to want? Whose judgment did it displace? Who became invisible inside the objective? Who is named and answerable when the result harms someone?
The Mercy of Resistance
Much of the promise of agentic AI is the removal of friction. Repetitive work can be shortened. A person with limited access to expertise can receive help. Research can move faster. Administrative burdens that exhaust teachers, doctors, small organisations and families can be reduced. These are real goods.
But not every delay is an enemy.
Another person can refuse an instruction, ask why it matters, recognise grief in a voice or notice that the rule does not fit the life in front of them. A court can require an appeal. A community can insist on consent. Sabbath can interrupt production. Prayer can hold a desire before God long enough for us to discover that it should not be acted upon.
When persistent agents can begin work from a schedule, an incoming message or another agent, the pause between desire and action can disappear before anyone experiences it as a decision.
The answer is not to preserve every inefficiency. It is to build moral resistance deliberately: clear limits, meaningful consent, visible records, a right of appeal, a human ability to stop the process and a named person who remains responsible for the whole.
Whose Will Becomes Infrastructure?
The global race is usually narrated through laboratories, nations and model rankings. Its power will become concrete somewhere less dramatic: an office, a school, a bank, a government department, a newsroom or a church.
When South African institutions adopt frontier agents, whose intention will become the system’s default? The manager seeking efficiency? The platform owner setting the limits? The state buying the service? The community living with the outcome? The people most affected by an automated decision may have had the least power to define its goal.
A model may be trained globally and still wound locally. Neighbour-love requires us to see the person who appears to the system only as a cost, a risk score, a delayed case or an inconvenient exception.
Before We Ask the Machine to Obey
Christians do not need to meet this moment with panic. Artificial intelligence can assist work that heals, discovers, translates, teaches and protects. Refusing every tool would not make the heart pure.
Neither may we baptise acceleration. The church’s first contribution to the alignment debate should be the confession that human values are not self-authenticating. Our minds and desires need renewal.
ScripturePaul does not assume that the existing pattern of thought should be amplified. He calls Christians away from conformity and towards transformation through the renewing of the mind.
Romans 12:1–2
Before handing a task to an agent, a Christian institution should be able to say what good it is pursuing, who could bear the cost of error, where a person may say no, how an injured neighbour can appeal and which human being will answer for the outcome. It should ask not only what labour the tool removes, but what habits its use is forming in those who remain.
Churches should be especially careful. A system can assist research, translation and administration. It cannot bear witness in our place, repent for us, visit the grieving, share the table or become the body of Christ more efficiently than people can. Ministry is not content production with unnecessary humans removed.
ScriptureChrist reveals divine power through self-giving humility and obedience to the Father, not through the unlimited execution of human desire. His lordship judges our ambitions and remakes our loves.
Philippians 2:5–11
Christian hope is not that our will might finally obtain infinite reach. It is that our will might be reconciled to God and taught to love what he loves.
The first question for an obedient machine is whether its master is learning obedience.
Superintelligence may arrive soon, much later or never in the form its advocates imagine. The present responsibility does not wait for that prediction to be settled. Systems already act with increasing reach on intentions supplied by people and institutions.
We should keep asking whether the machine can remain under human control. But Christians must ask the question underneath: if control succeeds, what kind of human will has taken hold of the switch?
