Mihaela van der SchaarNEW White Paper
Modern machine learning is powerful once a problem has been formulated. Supervised learning assumes targets and labels, while reinforcement learning assumes rewards and environments. Even agentic systems, despite their broader capabilities, generally inherit a predetermined framing: they operate within a given problem space, optimise towards specified objectives, and treat their task definition as fixed. Yet consequential problems often emerge differently, in science, healthcare, business, and other real-world environments.
An agent may be pursuing an existing task, or simply observing its environment, when it encounters a phenomenon that was neither sought nor represented in the current problem formulation. The world is full of such phenomena, and most do not warrant further attention. A phenomenon does not arrive already marked as a clue, anomaly, failure, opportunity, or problem; it merely occurs. The crucial judgement is whether one phenomenon among many may be consequential before its significance is known, and whether to disregard, retain, or investigate it, allow it to motivate a new problem, or use it to reshape an existing one.
This paper introduces open-beginningness: the capacity of an AI system to make this selective judgement during ongoing activity. The challenge is that there may be no immediate evidence that the phenomenon matters, no agreed interpretation of what it means, and no ground truth against which the decision can be evaluated. Its significance may depend on an unidentified context, emerge through evidence accumulated over time, or become visible only when the current question or representation is reconsidered. Open-beginningness is therefore distinct from anomaly detection, curiosity-driven exploration, hypothesis generation, and open-ended learning. These approaches identify novelty, surprise, information gain, or candidate explanations within a supplied representation, environment, objective, domain, or question. Open-beginningness concerns the prior judgement of whether a phenomenon should enter inquiry at all, and, if so, whether it should motivate a new problem or reshape an existing one.
The opportunity is substantial. Progress in science, healthcare, and business often begins when a phenomenon that others overlook or dismiss is recognised as consequential. Bringing this capacity into AI would move machine learning beyond answering questions already posed towards helping identify the phenomena around which new problems and questions should be formed, opening new directions for discovery, decision-making, and action.
If you’re interested in reading the full white paper, please see the file below:
Marika Niihori
