Assembling Ikea furniture is widely recognized as a trial of patience, spatial reasoning, and emotional endurance. Popular cultural folklore even suggests that the process of putting together a flat-pack dresser, a complex bed frame, or any other item from the iconic Swedish home furnishing brand carries enough inherent tension to strain or even destroy a relationship. The combination of vague instructions, cryptic diagrams, countless wooden dowels, and mysterious leftover hardware has become a universal punchline for household frustration.
However, beyond being an undeniable pain to build, it turns out that the intricate problem-solving and advanced spatial reasoning involved in constructing Malm dressers, Billy bookshelves, or other modular furniture pieces serve a much loftier purpose. According to recent research, these domestic assembly tasks provide a remarkably accurate and effective benchmark for assessing just how intelligent modern artificial intelligence models actually are.
Aiden Ament and Greg Burnham, researchers at Epoch AI—an artificial intelligence research nonprofit dedicated to forecasting and analyzing the development of machine intelligence—recently released a comprehensive report detailing the Furniture Assembly Benchmark. This innovative metric was specifically developed to test the true practical intelligence of artificial intelligence models by measuring how effectively a system can analyze physical reality, understand spatial relationships, and identify errors in partially assembled Ikea furniture.
In an era where technology companies constantly brag about the growing capabilities, parameters, and intelligence of their latest models, evaluating how smart each system truly is has grown increasingly complicated. This is especially true when attempting to draw direct comparisons between competing models developed by different labs. Asking advanced language models like OpenAI’s ChatGPT or Anthropic’s Claude to spot structural errors in Ikea furniture assembly puts these artificial intelligence systems to the test in complex, real-world applications that require physical reasoning rather than mere text prediction. While no single model tested managed to get the answers right one hundred percent of the time, the ongoing evaluation reveals that these systems are rapidly improving their comprehension of physical spaces and spatial logic.
The Ikea Test
The origin of this unusual testing method dates back to a corporate retreat held by the Epoch AI team. During the gathering, researchers were brainstorming and figuring out how to design a rigorous, real-world experiment to test the physical and spatial capabilities of modern artificial intelligence models outside of standard digital benchmarks.
Initially, the team attempted a very different approach that involved trying to let the artificial intelligence take the lead in a hands-on construction scenario. They experimented with acting as what they described as the artificial intelligence’s robot, asking the model what to do next and attempting to follow its directions somewhat literally in real time. However, the results of this initial trial were far from smooth.
According to Greg Burnham, who spoke with Fast Company about the experimental process, those early attempts were often messy and difficult to execute. Translating abstract instructions from a text-based model into physical movements in the real world proved to be a chaotic endeavor, highlighting the immense gap that still exists between digital reasoning and physical execution.
Fortunately, the team’s perspective shifted thanks to a refinement proposed by Burnham’s colleague, Aiden Ament. Instead of trying to have the artificial intelligence guide the entire construction process from scratch, the researchers pivoted to asking the model to spot mistakes in an already established setup.
As Burnham explained to Fast Company, this alternative concept proved to be much more effective and reflective of realistic human workflows. Asking an AI to identify where an assembly went wrong mirrors the reality that it is entirely acceptable to make mistakes mid-assembly, provided that the builder—or the intelligent system—possesses the critical reasoning skills required to catch and correct those errors before the project is completed.
