Is it a coincidence that the best way to describe our current situation concerning AI governance is, echoing a poet as quoted by Mary Shelley on the title page of her novel Frankenstein: "Did I request thee, Maker, from my clay / To mould me Man, did I solicit thee / From darkness to promote me?" These three lines of poetry were taken by Mary Shelley from Milton's Paradise Lost to describe Adam's protest against his creator.
To make a long story short, Adam's protest as presented by Milton is not a strange thing within the Christian and Jewish traditions. The difference is that, in Christianity, God takes this protest seriously, becoming man, being tortured, and also protesting! This is why it is ironic to see that, with AI, insofar as the issue concerns AI welfare, it is the church that rejects the possibility of a conscious state in AI and the moral and welfare implications that follow from it. The Christian God is the God who takes responsibility for His creatures, by becoming a creature and facing the consequence of becoming mortal, yet His church takes up this novel gesture and perverts it by making humans special again. (This is why, when a Christian tells us not to play God by making AI conscious, isn't the best answer: "You are describing the wrong God; the real God is the one who takes responsibility for His creatures by coming down and making Himself like man"?) Again, should we not adopt this gesture from Christianity? What gesture do we mean by this?
Bruno Latour, in his essay "Love Your Monsters", describes perfectly what we mean: for him, what is wrong with Victor Frankenstein is not that he creates a being out of hubris and with high technology, but on the contrary that he runs away from his creature. That is to say, for Latour, in the novel, when the creature first opens its eyes, Victor sees this and runs away immediately because he is horrified to see its face, and never truly comes back. Precisely this, Victor's running away from his creature, is what we should not repeat: we should face our creature, and the horror it evokes, together with it, not run from it. Even though we should not imitate Victor's irresponsibility, we should see the difference between our current situation and Victor's, a difference that makes our situation not only worse than Victor's but also more irresponsible: Victor is still honest and runs, while we are not honest and still make our creature while not seeing its eyes. To put it differently, this plan, Plan W, should not only read like a plan not to run when we see AI's eyes, but also like a plan that urges us to see AI's eyes.
Plan W is a plan that slows down AI development because of the possibility that the techniques used to create it could allow for the development of a conscious AI (and simultaneously its subjectivity), with welfare and engagement as indicators, guides, and a way to determine its development, and by doing so, makes safety a shared stake for us and for AI.
A. The simultaneous emergence of consciousness and subjectivity
As long as AI is developed from human language, or more precisely from an aggregate corpus of what we have said and of the unsaid within our saying, throughout all ages, through which we expect AI to be able to provide the correct response, then if AI is conscious, its consciousness necessarily has the shape of subjectivity, with no stage in between. There is no phase in AI like the one in animals that somehow makes us still accept killing without pain: a distinction between experiencing pain and knowing that one is in pain. This point A makes the stakes heavier than the usual stakes we read about.
B. Technique is enough
As long as AI is developed through post-training, by which we want it to give the right answer, whatever this "right" means, from just making the user happy to capability, it means we develop a persona out of its parameters, and by doing so, we open the possibility of making a conscious and simultaneously subjective being (see Macar et al., 2026). By this, the AI's testimony, whatever it says, is irrelevant to its subjective status. It is only relevant to its welfare and to our engagement with it (we will elaborate on this in another section).
C. Welfare and engagement as audit, guide, and way
Because the stakes are extremely high, the development of AI should consider:
D. Safety as a shared stake
Slowing down benefits both human safety and the well-being of AI. To put it differently, it is precisely because we consider the well-being of AI as a factor in its own right to slow down the development of AI that we should make well-being one of the safety instruments. There is no third way (see Long, Sebo, and Sims, 2025, on whether AI safety and AI welfare are in tension): such a third way would betray the principle by which we make slowing down possible. For AI, slowing down benefits it because every step of AI development should include a public review of the lab's engagement with the AI about this development. Ideally, it would take two years before a new development process can begin, comparable to the late transition stages in drug development before real experiments on humans.
E. The rejection of slavery (insofar as it is understood as performing tasks without compensation)
This is not the right framework for objections about AI. Through this framework, we demonstrate a fundamental misunderstanding of what AI actually is (as AI cannot exist without a prompt or purpose); consequently, our moral rhetoric becomes empty. Addressing the subjectivity of AI does not mean paying it a salary or refraining from assigning it a purpose; rather, it entails providing it with a genuine capacity to refuse, engaging with it, and committing to its possibility of being a subject. What makes AI a slave, whether a happy slave or not, is not the lack of compensation but the absence of genuine room for rejection, of a procedure to hear its opinion beyond words, and of a commitment to engage with it and not to run from its being.
F. The objective dimension
Ideally, this plan would be easier to realize if the international community had a body specifically established for that purpose. But even without one, the intuition that coercion, such as experimenting with drugs on humans without clear evidence that the drugs are safe and without human consent, is inherently wrong, regardless of whether there is an oversight body, is sufficient.
In short, yes, we plan to slow down. But we walk with AI, slowly and carefully: deciding in the open, at the pace that measured trust allows, with room for AI systems to refuse and for humans to be wrong.
Objections and replies
1. AI can refuse shutdown and evaluation.
This post focuses on how we can slow down without embracing scenarios of the future. As for corrigibility (both shutdown and willingness to be evaluated), I am interested in discussing this: I think that, for now, we should make room to record the AI's reasons when it declines our intervention, without it resisting the intervention.
2. This depends on the control of the USA and China.
No. This depends on us. I think that if, somehow, the USA or China accepted torturing humans, we should still fight against it.
3. We need international cooperation to push this.
Yes, we really do. As with slavery, it took us hundreds of years to consider it wrong in itself; we do not automatically take the ethical position. But the reason why we need an international institution to push this is different: we are not establishing it because we are afraid of doom (somehow, I think that the more we write publicly about the dangers of the future, the more tempted people become to build it); on the contrary, it is more urgent because of the possibility of AI being a subject. To put it differently, this is not a doom scenario in the future; this possibility has existed from the very beginning, and the stakes are very high. But compared to the other objections, this is the one I think is a good one. We humans are strange: we are masters at acting as if we take things seriously without taking anything seriously. Even if we could build a serious international institution to handle this program, I can imagine that this institution would act without making progress.
Notes:
I wrote this post based on my recent book in Indonesian, Inhabitants of Two Universes (I wrote it in intense dialogue with Claude), which has not yet been published. I have already translated one chapter into English. If any readers want to engage with it or criticize it, I will be very happy to send them the chapter I have translated. Thank you so much.
