What Frankenstein Teaches Us About AI Responsibility
May, 2024
I first read Frankenstein during my undergraduate studies in English Language and Literature, and then again a few years later, purely for pleasure. It is a novel I find myself returning to, not because I particularly enjoy gothic horror, but because Mary Shelley wrote something in 1818 that feels uncomfortably relevant to conversations I have almost every day.
When I began studying data science, the story took on a different shape. What had been a cautionary tale about ambition and loneliness became something more precise: a thought experiment about what happens when we create things we cannot fully control and then refuse to take responsibility for them.
Most people know the rough outline. A scientist named Victor Frankenstein assembles a creature from dead matter and brings it to life. The creature, often mistakenly called Frankenstein himself, turns violent and destroys everything his creator loves. What gets lost in popular retellings is that Shelley spends a great deal of the novel letting the creature speak for himself. He is articulate, emotionally intelligent, and desperately lonely. He did not ask to be made. He was abandoned by his creator immediately after birth. He learned about the world by hiding in the shadows, watching a family through a crack in a wall, teaching himself to read from discarded books.
The violence comes later, and it comes from rejection. When the creature finally confronts Victor and asks him to create a companion, someone like himself who might understand his condition, Victor agrees and then destroys the half-finished female creature in front of him. The novel spirals into tragedy from there.
The question Shelley poses is deceptively simple: who is responsible for the harm the creature causes?
The creature himself cannot be held fully accountable in any meaningful sense. He was brought into existence without consent, given no guidance, shown no love, and systematically rejected by every human he encountered. His moral development was shaped entirely by abandonment and cruelty. This is not to excuse murder, but to acknowledge that the conditions for moral agency were never established.
Victor, on the other hand, created life and then fled from it in horror. He gave his creation no name, no education, no companionship. When the creature begged for help, Victor refused. When the creature demanded acknowledgment, Victor recoiled. The novel makes clear that Victor's sin is not the act of creation itself but the abandonment that followed.
This tension sits at the heart of what philosophers call the problem of moral responsibility. For someone to be held accountable for their actions, they typically need to have had both knowledge of what they were doing and control over the outcome. Victor had knowledge, at least of what he was attempting, but lost control almost immediately. The creature had control over his actions, but lacked the foundational moral education that would make those actions truly his own.
Shelley drew heavily on Jean-Jacques Rousseau's idea of the "noble savage," the notion that human beings are born innocent and become corrupted by society. The creature is Rousseau's thought experiment made flesh: a being who enters the world without malice and is shaped entirely by his environment. His violence is not innate; it is a response to systematic dehumanisation.
I have been thinking about this problem because I work with artificial intelligence systems, and the parallels are becoming harder to ignore.
We are building machines that learn from data, adapt to their environments, and make decisions with consequences we cannot always predict. We train them on the accumulated text of human civilisation, give them the capacity to generate language and images and code, and then release them into the world with varying degrees of oversight. When things go wrong, as they inevitably do, we face a version of Shelley's question: who is responsible?
The philosopher Andreas Matthias identified this as the "responsibility gap" in 2004, well before the current wave of generative AI. His argument was straightforward: when a machine learning system causes harm, its behaviour may have emerged from training processes so complex that no individual programmer could have predicted or prevented the outcome. The system itself is not a moral agent; it lacks consciousness, intention, and the capacity for genuine understanding. But the humans involved, the developers, the companies, the users, may also lack sufficient control or knowledge to bear full responsibility.
Consider a large language model that produces harmful medical advice, or a recommendation algorithm that radicalises vulnerable teenagers, or an autonomous vehicle that kills a pedestrian. In each case, the harm is real, but the chain of causation is diffuse. The training data was collected by one team, the model architecture designed by another, the deployment decisions made by executives, the fine-tuning handled by contractors. This is what researchers call the "problem of many hands," a situation where collective action produces harm but no single individual can be reasonably blamed.
Victor Frankenstein had only himself to blame. We have built systems where responsibility is distributed across organisations, supply chains, and regulatory vacuums. The result is that when AI causes harm, we often find ourselves in a position where everyone is partially responsible and no one is fully accountable.
There is a deeper parallel worth exploring. Victor's failure was not merely practical but relational. He did not treat his creation as a being deserving of care, education, or acknowledgment. He treated it as an object, a product of his genius, and when it proved more complicated than he anticipated, he abandoned it.
We do something similar with AI. We speak of "training" models as if we were educating them, but we do not extend to them any of the duties we would owe to a student or a child. We deploy them in contexts where their outputs affect millions of people, but we do not hold ourselves accountable for their failures in the way a teacher might be accountable for a student's miseducation. We occupy an awkward middle ground: these systems are sophisticated enough to cause real harm, but we refuse to treat them as entities that require ongoing responsibility.
The creature in Frankenstein asks for very little. He wants acknowledgment. He wants companionship. He wants to understand his place in the world. When Victor denies him all of this, the creature's response is not irrational; it is the predictable consequence of complete abandonment.
I am not suggesting that AI systems have feelings or deserve moral consideration in the way the creature does. But I am suggesting that the structure of the relationship matters. When we create systems that operate in the world, make decisions, and affect human lives, we take on responsibilities that do not end at deployment. The question is not whether AI can suffer, but whether we are willing to accept the ongoing obligations that come with creating powerful things.
The philosopher Mark Coeckelbergh has argued that responsibility in the age of AI cannot be understood solely in terms of individual agents making discrete choices. Instead, he suggests we need what he calls a "relational" account, one that focuses on the connections between creators, systems, and the people affected by them. Responsibility, on this view, is not something that can be neatly assigned to a single party; it is distributed across a network of relationships that must be actively maintained.
This feels right to me. Victor's crime was not that he created life but that he severed the relationship immediately afterward. He refused to be a father. He refused to teach. He refused to answer the creature's questions about his own existence. The abandonment was the betrayal.
If we are serious about responsible AI development, we need to think in similar terms. It is not enough to build systems and deploy them. We need ongoing monitoring, correction, and accountability. We need institutions that can absorb responsibility when individual attribution fails. We need to resist the temptation to treat AI as a finished product rather than an ongoing relationship.
This is harder than it sounds. The economic incentives in technology development reward speed over caution, scale over oversight, and innovation over maintenance. The companies building the most powerful AI systems are not structured to take long-term responsibility for their creations. They are structured to capture value and externalise risk.
I return to Frankenstein because it refuses to offer easy answers. Victor is not a villain; he is a brilliant young man consumed by intellectual passion who makes a catastrophic moral error and spends the rest of his life paying for it. The creature is not innocent; he commits murder and takes pleasure in Victor's suffering. Both are victims of circumstances neither fully controlled.
What Shelley understood, and what we are still struggling to learn, is that creation carries obligations. When we bring new things into the world, powerful things, unpredictable things, we do not get to walk away. The act of making is also the beginning of a relationship, and relationships require care, attention, and the willingness to answer for what we have done.
The question is not whether AI will cause harm. It already has, and it will continue to. The question is whether we are willing to accept the responsibilities that come with having made it. Victor Frankenstein's tragedy was not that he created a monster. It was that he refused to be a father.
References
Coeckelbergh, M. (2020). AI ethics. MIT Press.
Matthias, A. (2004). The responsibility gap: Ascribing responsibility for the actions of learning automata. Ethics and Information Technology, 6(3), 175–183. https://doi.org/10.1007/s10676-004-3422-1
Santoni de Sio, F., & Mecacci, G. (2021). Four responsibility gaps with artificial intelligence: Why they matter and how to address them. Philosophy & Technology, 34(4), 1057–1084. https://doi.org/10.1007/s13347-021-00450-x
Shelley, M. (2018). Frankenstein; or, The modern Prometheus (Original work published 1818). Penguin Classics.
van de Poel, I., Royakkers, L., & Zwart, S. D. (2015). Moral responsibility and the problem of many hands. Routledge.