There is a standard argument in favour of human annotation, and it is not wrong. Labelled data provides a clear training signal, reduces ambiguity about what the model should learn, and enables rapid iteration against a fixed evaluation target. For many commercially important tasks — named entity recognition, medical image classification, intent detection in dialogue — supervised learning on carefully curated annotations remains the most direct path to usable performance.
We are not arguing against labelling in those contexts. But we think the field has systematically misclassified the problem. Annotation is treated as a data pipeline issue: expensive, slow, requiring quality control, but fundamentally a procurement and logistics challenge. You have too little labelled data, you get more. You have noisy labels, you apply label-cleaning heuristics. The underlying assumption is that the annotation bottleneck is a cost curve, not a conceptual limit.
We think it is a conceptual limit, and the implications run deeper than they first appear.
What annotation actually encodes
When a human annotator labels a piece of data, they are encoding a specific kind of knowledge: knowledge that was articulable at the moment of labelling. A bounding box drawn around a pedestrian in a driving dataset encodes a human's understanding of "pedestrian" — an understanding that is itself the product of years of perceptual experience, but which is being expressed in a single discrete act of classification.
The annotation is a compression. It takes a rich, contextual judgment — one that might draw on posture, shadow, depth, motion, contextual plausibility — and reduces it to a label that a training objective can consume. Much of the information that produced the judgment is discarded. The model receives the conclusion without the reasoning.
This matters because the reasoning — the intermediate perceptual processing — is often what we actually want to learn. A pedestrian-detection model that has learned the label "pedestrian" as a surface pattern is brittle in different ways from one that has built representations of human anatomy, motion dynamics, and scene context. The latter kind of model requires representations that were not in the label, and may not be easily constructable from the label alone.
The structural dependency problem
A deeper issue is what annotation does to the structure of what can be learned.
Consider a domain like early drug discovery, where the relevant features of a candidate molecule — properties that will determine binding affinity, off-target effects, synthetic accessibility — are not fully known at the time of dataset construction. An annotator can label known molecules with known outcomes. But the annotation schema necessarily reflects current scientific understanding of what matters. Features that current science does not know to look for are not annotated. A model trained on this data will learn representations of what is currently understood, and will be blind to structure that lies outside that understanding.
This is not a failure of the annotators. It is a structural consequence of what annotation is. Annotation maps data to the categories that humans have already formed. It cannot, by construction, teach a model to form categories that humans have not yet discovered.
Self-supervised approaches sidestep this constraint. When a model learns to predict masked portions of a molecular graph from surrounding structure, or to capture which two views of a molecule should be represented similarly, it is not constrained by what the annotation schema can express. The learning objective is derived from the structure of the data itself — its internal regularities, its symmetries, its predictive relationships. Representations that form under these pressures may capture structure that no annotation would have pointed toward.
The active knowledge problem
There is also an epistemological issue that tends to get skipped in engineering-focused discussions of annotation. Some knowledge is tacit — not easily explicated even by experts who demonstrably possess it. A pathologist examining a tissue slide brings to bear perceptual skills that would be extremely difficult to fully decompose into annotatable features. They know something, and they can mark a region as suspicious, but the full content of that knowledge is not captured by the mark.
Annotation as a training signal works by assuming that what can be extracted from the annotator and placed in a label is sufficient to teach the target skill. But if a significant portion of expert knowledge is tacit — embodied in perception and pattern recognition that precedes and exceeds verbal articulation — then annotation systematically undertransmits what we want the model to learn.
This is not a niche concern. It is arguably the normal case in the most interesting domains. The situations where annotation is easiest are often the situations where the problem is already well-understood and formal enough that it could, in principle, be solved by explicit rules rather than learning.
Practical consequences at scale
None of the above is a criticism of any specific annotation effort. The groups building large labelled datasets are doing valuable work, and many of the representations learned on those datasets are genuinely useful. The concern is about what the field has been optimising toward, and where the long-run costs show up.
As model scale has grown, the annotation bottleneck has become more visible, not less. Training larger models requires more data. Generating more labelled data at the frontier is not linearly cheap — annotation tasks become harder as models improve and the remaining unlabelled problems are harder to describe. The cost curve for human labelling is not flat; it rises steeply as tasks move from straightforward classification toward nuanced judgment.
Meanwhile, the volume of unannotated data — text, sequence, sensor streams, imagery — grows without any corresponding need for human effort. The question of how to extract useful representations from raw, unannotated data is not just a research curiosity. It is becoming the binding constraint on what the field can do next.
What self-supervised learning offers — and does not
Self-supervised learning, broadly construed, is the attempt to derive learning signal from the structure of data itself rather than from human-assigned labels. Masked prediction, contrastive objectives, next-token prediction over unlabelled sequences — these are all ways of defining a training problem without asking a human to annotate anything.
The resulting representations have, in practice, proven to transfer well to downstream tasks. This is not accidental: representations formed by learning to predict or contrast natural data tend to capture structure that is useful across many applications, because they are responding to the structure of the data rather than to the arbitrary features of a specific labelling scheme.
We want to be careful not to overclaim here. Self-supervised learning does not eliminate the need for human judgment in AI development — evaluation still requires it, and task specification still requires it. What it removes is the requirement that human judgment be provided in the form of per-example labels before the model can begin learning. That is a meaningful architectural shift, not a wholesale solution to all of the problems annotation was trying to solve.
But the architectural shift matters. A system that forms representations from raw data, without an annotation-derived training signal, is not constrained to learn only what annotators knew to label. That freedom is, in our view, the most important property to preserve as AI systems are asked to operate in genuinely novel domains — the domains where no one has yet written the answer down.
