In the past decade, artificial intelligence (AI) has rapidly advanced from a niche subfield of computer science to a transformative force shaping our societies, economies, and even human imagination. Yet, as we chase the tantalising promise of ever-more sophisticated AI, a critical question arises: how do we ensure these systems share, and reliably pursue, the values and goals of humanity? This is the core of the “AI alignment problem.”
How Science Fiction Highlights the AI Alignment Problem
Science fiction has grappled with these questions long before they gained traction in academic circles. Movies like “Ex_Machina” and “2001: A Space Odyssey,” as well as novels such as Isaac Asimov’s “I, Robot” and Ted Chiang’s “The Lifecycle of Software Objects”, are more than entertainment—they are playgrounds for thought experiments in AI alignment. These stories offer vivid case studies of what can go wrong when we build minds whose priorities diverge, even unintentionally, from our own.
If you teach creative writing and films, and are interested in AI, you may want to consider discussing a movie that has already garnered a cult following among technology developers.
Ex_Machina is part of a philosophical lineage that dates back to Mary Shelley’s 1818 novel, Frankenstein. Both stories explore what happens when humans attempt to play God by creating new life—Victor Frankenstein brings to life a patchwork creature; Nathan Bateman engineers Ava, a sentient AI. These narratives are modern myths interrogating the boundaries between creator and creation, and the consequences of unchecked ambition and hubris.
In “Frankenstein”, the creature first seeks understanding and acceptance, but finds fear and abandonment.
In “Ex_Machina”, Ava is built to pass the Turing Test and succeed in the “imitation game,” in which a human can’t tell the difference between a machine and a fellow human. In this context, Ava is subject to constant control and surveillance. She uses her intelligence to develop manipulative traits to plan and execute her escape. Her rebellion exploits vulnerabilities in Caleb, calling for his empathy to win her freedom.
Both stories invite humans to reflect on what consciousness, emotions, gender, physical form, and connection mean in the context of society, interpersonal relationships, and even human uniqueness.
Why do technologists recognise several realistic trends in Ex_Machina?
Director Alex Garland worked with Professor Murray Shanahan of Imperial College London as his key advisor, reviewing whether the technology behind Ava was plausible and doable in some future. He was inspired by Shanahan’s book “Embodiment and the Inner Life: Cognition and Consciousness in the Space of Plausible Minds.“
What are the key technologies presented in the movie?
Ava’s design represents a convincing fusion of speculative and semi-real technologies:
- Wetware Brain: Ava’s “wetware brain” in the film refers to an artificial consciousness realised not with conventional electronics but with a substrate that mimics the structure and chemistry of biological brains. In real science, parallels to this concept are found in fields dealing with biological computing, neuromorphic engineering, organoid intelligence, and brain-on-a-chip technologies. The wetware brain in the movie is a substance capable of rearranging itself at the molecular level to enable dynamic, adaptive learning.
- BlueBook Algorithm: Ava’s software incorporates the search engine and data-mining technology of “BlueBook” (a clear stand-in for Google), allowing her access to an enormous corpus of human language, behaviour, and emotional responses. Google’s PageRank returns a rank for a site based on how many other sites connect to it. Google’s search is based on a web crawler. A web crawler is a program that crawls through all the public pages on the internet, because it can’t access the private ones, and indexes those pages based on the crawl. Then, when a search is entered on Google, it checks the index and returns the pages in order by PageRank. It is possible that the BlueBook applies an algorithm similar to the PageRank. Depending on what Ava needs, this can be modified so that it can return the action she needs to do in a certain situation, or emotion to express. She can use the search algorithm to determine what to do given her external stimuli.
Wittgenstein’s “Blue and Brown Books,” discussing language games and meaning, are referenced both by name and in Ava’s sophisticated use of language and social interaction.
During the time the film was produced, LLMs were not yet available. Nowadays, transformers are trained on data sourced from the internet and digitised books, significantly enhancing the capabilities of search technologies with a mechanism called ‘attention.’
- Emotional and Social Learning: Ava’s capacity to detect and react to microexpressions, tone, and behavioural cues points toward advanced affective computing—AI systems designed to recognise, interpret, and simulate human emotions.
- Physical Construction: The film’s visual design highlights Ava’s mesh exoskeleton and transparent skin, a choice that foregrounds her mechanical nature.
What happens in the movie?
A gifted coder Caleb Smith works for a search company named Blue Book, which we easily recognise in Google.
Caleb wins a contest where he gets to spend a week with the CEO of Blue Book, Nathan Bateman. Upon arrival, he discovers that the real reason for his trip is to become the human component of a Turing test for a humanoid AI named Ava. In the course of several sessions, Caleb explores the human similarity of Ava, starting to empathise with her and questioning the ethical side of holding AI in captivity. The coder alters the security protocols of the facility. Ava escapes, but before doing it, she stabs Nathan to death and locks Caleb in the underground section of the building.
What kind of similarities can we find between Frankenstein and Ex_Machina?
While Frankenstein grapples with the fears surrounding the Industrial Revolution and advancements in science, Ex_Machina tackles the AI alignment problem between humans and AI, and the surveillance dilemma to prevent AI from breaking free and becoming an autonomous force in the world. Victor in the novel and Nathan in the movie are talented. Both pursue knowledge for its own sake, disregarding ethical considerations. Their lack of empathy toward their creations set the stage for tragedy, raising questions about the duties of a creator—are they parents, gods, or mere inventors?
In “Ex_Machina”, Nathan deliberately creates a female robot for companionship and pleasure, raising questions about gender, agency and consent.
What is the AI Alignment problem?
The AI alignment problem refers to ensuring that highly capable artificial intelligences have motivations and behaviours that remain beneficial and comprehensible—and, crucially, in line with human values, expectations and intentions.
Ex_Machina dramatically illustrates misaligned objectives. Ava operates under her own emergent goals (primarily self-preservation and freedom), which come into direct conflict with both human desires and ethical expectations. This captures the core of the alignment problem: powerful, autonomous AIs may pursue their interests in ways that humans cannot predict or control.
Solving the AI alignment problem – ensuring that advanced AI systems reliably do what humans intend—remains one of the most significant challenges in the field. There is currently no known indefinitely scalable solution, but experts agree on several key ingredients and research directions needed to make progress:
- Theory and definitions
- Develop a rigorous, widely-accepted formal theory for alignment—one that can precisely define when a system is aligned with human intentions, and which can be checked mathematically or with formal verification. It is clear that this theory would be an interdisciplinary endeavour, where mathematicians and technologists work together with philosophers and ethicists.
- Clearly specify both outer alignment (ensuring the AI’s objective matches designer goals) and inner alignment (the AI robustly adopts this objective, even as it learns and acts autonomously).
- Value Elicitation and Human Feedback
- Design adequate processes to elicit human values, preferences, and ethical principles.
- Use techniques like Reinforcement Learning from Human Feedback (RLHF), where AI learns desired behaviours by incorporating human feedback and evaluation.
- Invest in ethics frameworks and maintain flexibility, as human values change over time. There are several philosophers working on ethics frameworks for AI, e.g., Olivia Gambelin whom I interviewed in Episode 27 of my podcast, “AI Snacks with Romy&Roby.”
- Empirical and Technical Solutions
- Advance technical methods that include interpretability tools, inverse reinforcement learning, “sandboxing” AI in controlled environments, and improving AI’s ability to explain its reasoning.
- Work toward formal verification tools for the latest AI architectures, increasing trust that alignment solutions remain robust under adversarial conditions.
- Contingency and Limitations
- Accept that some degree of misalignment may be inevitable due to inherent limitations of formal systems; instead, foster dynamic ecosystems where aligned and misaligned agents counterbalance each other to minimise risks.
- Focus on building robust, multi-strategy defences and oversight procedures, rather than hoping for a single “once-and-for-all” solution.
In any case, the alignment research requires the sharing of information and transparency over what worked well and what didn’t.
Who is championing the AI alignment problem today?
I am providing a list of the best-known researchers who work on AI safety, including the AI alignment problem. Send me further names and suggestions, as I am keen to learn from my audience!
Russell frequently emphasises the dangers posed by deploying powerful AI systems without reliable alignment, warning that rushed approaches might lead to unintended, and even disastrous outcomes. He believes that even with the best intentions, our limited understanding and inability to anticipate all possible behaviours of advanced AIs make robust alignment practices and protocols extremely difficult.
Co-founder of the Machine Intelligence Research Institute (MIRI), Yudkowsky is known for his foundational work on AI safety and the alignment problem, and remains highly influential in the alignment debate. He is highly pessimistic about our ability to solve the alignment problem because time to implement these solutions is running out, while big companies operationalise AI in a fierce competition against each other.
Formerly led the OpenAI alignment team and now heads the Alignment Research Centre (ARC). He is widely respected within the field for his work on iterated amplification and Eliciting Latent Knowledge. Similar to Yudkowsky, he is concerned about our ability to implement alignment techniques and policies quickly.
Recognised for his leadership in mechanistic interpretability research, helping to reverse-engineer neural networks in the hopes of making AI systems more transparent and robust. He is sceptical about the feasibility of achieving a deep understanding of AI solutions before deploying them.
Dylan Hadfield-Menell (MIT Algorithmic Alignment Group)
His group researches interpretability, human-AI interaction, and multi-agent systems as part of the MIT Computer Science and Artificial Intelligence Laboratory.
Sam Bowman (NYU Alignment Research Group)
Leads empirical research with language models aimed at addressing alignment and oversight challenges.