In 2025, machine-learning-based art valuation tools are moving from the laboratory into practical use. In the field of Chinese ceramics, the shift is especially visible: structured auction records, a relatively stable category system, and a large body of historical transaction data make the field an ideal training ground for AI valuation models.
How AI Valuation Works
Most current AI ceramics valuation systems are built on two types of methods:
Comparables-Based Estimation The system searches historical auction records for sale results most similar to the target object. It then derives an estimate from the hammer prices of those comparable lots, adjusted for time trends. The method is similar to the comparative approach used in real estate valuation; the central question is how accurately the system can measure "similarity."
Modern systems use two-stage retrieval:
- Text-feature matching: extracting structured features such as dynasty, category, decoration, and dimensions to filter an initial candidate set
- Visual-similarity reranking: using multimodal visual models such as CLIP to convert object images into high-dimensional feature vectors, then reranking the candidates by visual similarity
Machine-Learning Regression Models Using historical sale results as the training set, these models take object features such as dynasty, category, auction house, decorative technique, estimate, and dimensions as input variables and directly predict the hammer price. Common methods include gradient-boosted trees (XGBoost, LightGBM) and neural networks.
The two approaches have different strengths and weaknesses. High-accuracy systems usually combine them, with comparable-lot retrieval as the primary layer and machine-learning models as a secondary check when sufficient data are available.
Real-World Accuracy of AI Valuation
Typical performance for models trained on 140,000 auction records:
- Median absolute percentage error (MAPE): 65-85% -- meaning that for an object that actually sells for $10,000, there is a 50% probability that the model estimate will fall between $1,500 and $18,500
- P10/P90 interval coverage: roughly 75% of actual hammer prices fall within the model's 10th-90th percentile range
- Directional accuracy: judgments such as "above market estimate" or "below market estimate" are correct approximately 65% of the time
At first glance, this accuracy may not seem high. In the context of art valuation, however, it is already a meaningful reference point. Even internal estimates by major auction houses, made by human specialists, can vary widely for top-tier rarities, because the number of relevant comparables is often extremely small.
Core Limits of AI Valuation
Limit 1: Blind Spots in Quality Tiering The largest problem with current AI systems is their inability to distinguish quality levels within the same category. A "Kangxi blue-and-white landscape jar" may sell anywhere from $2,000 to $200,000 depending on painting quality, refinement of body and glaze, completeness of composition, and other factors that still require expert human judgment. AI can recognize that an object is a Kangxi blue-and-white jar; it struggles to decide whether it is a fine example within that category.
Limit 2: Insufficient Data for Rarities The rarer the object, the fewer historical comparables exist and the lower the accuracy of an AI estimate. A piece of Ru ware has only a handful of auction comparables in the historical record; under such data-scarce conditions, a statistical model can barely produce a meaningful valuation. AI performs best precisely with those "mid-market standard pieces" that are more numerous and more regular in pricing behavior.
Limit 3: Condition and Restoration Cannot Be Detected Reliably AI vision models cannot reliably identify restoration, tiny chips, or fine hairline cracks from photographs. Yet differences in condition can affect hammer prices by more than 50%. This means AI estimates usually assume sound condition by default; in practical use, a human condition adjustment must be added.
Limit 4: Market Sentiment Cannot Be Forecast During the 2021 pandemic period, the global art market was unusually heated and sale prices often exceeded normal levels by 20-40%. In 2022-2023 the market corrected, and prices for comparable objects declined. AI models trained on historical data cannot foresee this kind of macro sentiment shift, so estimates can become systematically biased around market turning points.
The Right Relationship Between AI and Human Specialists
AI valuation is not meant to replace human experts. It is meant to expand access.
Where AI is best used:
- Rapidly screening large numbers of lots to identify objects with stronger value potential
- Giving buyers a first-level market reference range
- Helping dealers and collectors make preliminary judgments when no specialist is present
- Tracking long-term price trends in specific categories
Where human specialists are still essential:
- Authentication, with AI as an aid but not the final authority
- Valuing top-tier rarities
- Assessing condition and identifying restoration
- Legally binding insurance valuations
Future Directions
Over the next three to five years, AI valuation systems are most likely to make progress in the following areas:
- Multimodal integration: combining text descriptions, front images, reverse images, reign-mark photographs, and detail shots to improve depth of visual understanding
- Quality-tier modeling: introducing expert scoring data to train models that can distinguish "ordinary" from "fine" examples
- Real-time market-data access: updating models continuously with near-real-time sale results to reduce time-lag bias
- Uncertainty quantification: providing not only a point estimate but a confidence interval, so users understand how reliable the valuation is
AI is still at an early stage in Chinese ceramics, but its potential to reduce information asymmetry and improve market transparency is real. For collectors, learning how to use these tools correctly will become an important skill in the decade ahead.