Beyond Task Intelligence: Artificial Intelligence, Embodied Robotics, and the Epistemic Limits of Machine Cultural Understanding
This paper argues that despite AI's superior task performance, it lacks the embodied, situated, and reflexive nature of human knowledge, necessitating a shift in evaluation frameworks toward epistemic competence and positioning the humanities as the essential intellectual environment for defining and testing true general intelligence.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a world where machines can run faster than the fastest human sprinter, yet still struggle to pick up a simple cable without dropping it. This is not a story about robots failing, but about a deeper puzzle in how we measure intelligence. Today, artificial intelligence systems can write fluent essays, translate languages, and retrieve facts from vast libraries of text with startling speed. They are being tested on specific tasks, like solving math problems or identifying objects in photos, and they are often winning. But winning a specific test does not mean a machine truly understands the world it is operating in. The question facing scientists and scholars is no longer whether a machine can do a task, but whether it can understand the context, history, and human relationships that give that task its meaning. This is the difference between following a set of instructions perfectly and knowing when those instructions no longer make sense.
A new systematic review by Prudhvi Kumar Kanyadari of Loughborough University explores this gap between doing and understanding. The paper argues that while machines are becoming incredibly good at specific, measurable jobs, they often lack a deeper kind of knowledge that humans take for granted. This deeper knowledge is not just about having information; it is about knowing how that information fits into a specific time, place, and culture. The author suggests that we need to stop judging artificial intelligence only by how well it scores on standard tests and start asking if it can handle the messy, uncertain, and often contradictory nature of human life.
The paper begins by looking at a striking example from the 2026 World Humanoid Robot Games. In these events, robots like the Tiangong Ultra ran the 100-meter dash in 8.64 seconds, beating the human world record of 9.58 seconds held by Usain Bolt. These machines are incredibly fast and efficient at running. However, the same robots often struggle with much simpler, everyday tasks, such as plugging in a cable or handling a package in a changing environment. The author points out that running is a closed problem: the track is flat, the distance is fixed, and the goal is clear. Plugging in a cable, by contrast, is an open problem. The robot must feel the shape of the plug, sense the resistance of the socket, and adjust its grip if the cable moves. It must adapt to a world that is not perfectly designed for it. The robot can win the race, but it does not yet know how to navigate the world.
This same pattern appears in the world of language and culture. Large language models, the systems behind many modern chatbots, can retrieve literary quotes or summarize historical events faster than any human researcher. Yet, when asked to interpret a cultural ritual or explain a social custom, these systems often make mistakes that reveal a lack of true understanding. They might produce a beautifully written explanation of a tradition that is factually wrong, or they might repeat harmful stereotypes about a group of people because those stereotypes are common in the data they were trained on. The paper notes that these systems can retrieve information, but they often fail to understand the "thick description" of culture—the idea that the same action can mean different things depending on who is doing it, where they are, and what history they carry.
The author introduces three distinct ways to measure intelligence to clarify this problem. The first is task competence, which asks if a machine can perform a specific operation, like running a race or finding a word. The second is contextual competence, which asks if the machine can still perform well when the situation changes, such as when the language is different, the history is complex, or the physical environment is unpredictable. The third and most difficult is epistemic competence. This asks if the machine knows when it does not have enough information to give an answer, or if it can recognize when a question itself is flawed. Current artificial intelligence is very good at the first type of competence but often fails at the second and third. It can give a confident answer even when the evidence is missing, or when the topic is too complex for a single, simple explanation.
One of the most serious issues highlighted in the paper is the problem of "hallucinated" citations. In academic research, a citation is a reference to a source that proves a point. However, these systems sometimes invent sources that look real but do not exist, or they cite real books but claim those books say things they never actually said. This is dangerous because it breaks the chain of evidence. In the humanities, knowledge is built on a chain of documents, archives, and interpretations. If a machine inserts a fake link into that chain, the whole structure can collapse. The paper argues that this is not just a bug to be fixed with better data; it is a fundamental difference between how machines process information and how humans build knowledge. Humans know when to stop and say, "I do not know," or "This source is unreliable." Machines, driven to complete a task, often push forward with a confident but wrong answer.
The review also examines how these systems handle culture and representation. When asked to generate images of people from specific cultures, these systems often rely on stereotypes. For example, when generating images of people from India, the system might consistently show upper-caste individuals in modern settings while depicting lower-caste individuals only in scenes of poverty or protest. This happens because the system learns from patterns in existing data, and those patterns often reflect historical biases. The result is an image that looks plausible but is actually a distortion of reality. The paper suggests that a culturally convincing image can be an epistemic failure, meaning it fails to represent the truth of the culture it claims to show.
The author proposes that we need a new way to test artificial intelligence, one that goes beyond simple scores. Instead of just asking if a machine can answer a question correctly, we should ask if it can handle changes in context. Can it still understand a story if the historical period changes? Can it recognize when a cultural term has no direct translation? Can it admit when a question is based on a false assumption? The paper suggests that future benchmarks should include "perturbation" tests, where researchers deliberately change the conditions of a task to see if the system breaks down. They should also test for "epistemic abstention," rewarding the system for knowing when to say it does not have enough information to answer.
This new approach requires a partnership between computer scientists and scholars from the humanities, such as historians, anthropologists, and literary critics. These fields have spent centuries studying how knowledge is created, how categories are formed, and how power influences what we consider true. The paper argues that these scholars should not just be critics of technology but active participants in designing it. They can help define what it means for a machine to be "culturally competent" and can help build tests that reveal the hidden biases and limitations of current systems.
The paper concludes that the future of artificial intelligence is not about building machines that are faster or more powerful, but about building systems that are more accountable. A truly intelligent system should not just be able to solve a problem; it should be able to understand the problem itself. It should know when a question is unfair, when a category is outdated, or when a representation is harmful. It should be able to recognize the limits of its own knowledge. The author suggests that the humanities are the perfect place to test these capabilities, because they deal with the very things that machines struggle with: ambiguity, history, and the complex web of human relationships.
Ultimately, the goal is not to make machines act exactly like humans, but to create systems that can work responsibly alongside us. This means developing machines that can distinguish between a fact and an opinion, between a stereotype and a reality, and between a confident guess and a well-supported truth. The paper suggests that if we can achieve this, we will have moved beyond simple task performance to a form of intelligence that is truly general and truly useful. The journey ahead is not just about engineering better robots or smarter software; it is about rethinking what it means to know something in a world that is constantly changing. The most important step is to ensure that as these systems grow more capable, they also grow more humble, recognizing that some questions require more than just an answer.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.