本文探讨 AI 估值工具如何在价格常被隐藏的市场中用于艺术品定价，并以 Stephen Smith 为例：这位 56-year-old UK 收藏家在 2024 以 £7,500 ($10,300) 买下 Bridget Riley 的 Echo，在此前曾出价过高后，如今改为依赖平台估值。透明度是核心摩擦点，因为画廊与经销商交易很少公开；2025 Deloitte Private and ArtTactic Art & Finance Report 显示，受访收藏家中超过 50% 认为不透明定价是主要障碍。对更清晰基准的需求正推动资料导向工具的采用，尤其是在较新的买家与较低价格区间。
各家公司正围绕估值建立不同的 AI 技术栈与商业模式。Artscapy 表示其模型已训练 3.5 years，资料来自拍卖结果、平台销售、可验证的私下交易、收藏家资料与专有演算法；它提供免费估值，以及一份 £59.99 的评估报告，内容包含价值、流动性与波动性，并把输出用于借贷与保险产品。MyArtBroker 避免纯生成式输出，改用受监督机器学习并由人工稽核，结合 400 家拍卖行、私下交易、平台需求资料，以及如色彩、签名、纸张类型等另外 40 个变数；使用者会得到公平市场合理区间，而专家复核再补上品况、稀有性与来源。产业层面上，整并正在扩大资料范围：2025 年 Beowolff Capital 接管 Artnet 并收购 Artsy 的控股权，计划把拍卖、一级市场与行为资料合并为新的分析能力。
最强的限制是市场资料缺失：超过 60% 的艺术品销售是透过画廊与经销商私下完成，因此即使是先进模型也仍需在不完整可见性上外推。AI 可透过影像辨识与更广泛摄取结构化、非结构化输入（如展览史、引用、新闻、社会趋势与搜寻行为）来提升讯号品质，但受访者一致将这些工具定位为第一步参考，而非取代顾问与鉴价师。实务执行风险仍高，因为艺术品流动性低，报价多为理论值，无法立即以明确 bid-offer spread 成交；因此，信心提升对较低价或较标准化作品最可靠，而独特且高价作品仍依赖人类判断、关系、信任与情境。
The article examines how AI valuation tools are emerging to price art in a market where prices are often hidden, using the case of Stephen Smith, a 56-year-old UK collector who bought Bridget Riley’s Echo in 2024 for £7,500 ($10,300) and now relies on platform estimates after previously overpaying. Transparency is a core friction point because gallery and dealer transactions are rarely public, and the 2025 Deloitte Private and ArtTactic Art & Finance Report found that more than 50% of surveyed collectors saw opaque pricing as a major barrier. This demand for clearer benchmarks is driving adoption of data-led tools, especially among newer buyers and at lower price points.
Companies are building different AI stacks and business models around valuation. Artscapy says it has trained its model for 3.5 years using auction results, platform sales, verifiable private sales, collector data, and proprietary algorithms; it offers free estimates plus a £59.99 appraisal report including value, liquidity, and volatility, and also uses outputs for lending and insurance products. MyArtBroker avoids pure generative outputs and uses supervised machine learning with human auditing, combining data from 400 auction houses, private transactions, platform demand, and 40 additional variables such as color, signature, and paper type; users get a fair-market range, while specialist reviews add condition, rarity, and provenance. At an industry level, consolidation is expanding data scope: in 2025 Beowolff Capital took over Artnet and bought a controlling stake in Artsy, with plans to merge auction, primary-market, and behavioral data into new analytics.
The strongest constraint is missing market data: more than 60% of art sales occur privately through galleries and dealers, so even advanced models still extrapolate from incomplete visibility. AI can improve signal quality through image recognition and broader ingestion of structured and unstructured inputs like exhibition history, citations, news, social trends, and search behavior, but interviewees consistently frame these tools as a first-pass reference rather than a replacement for advisers and appraisers. Practical execution risk remains high because art is illiquid, so quoted values are often theoretical and not instantly tradable at a firm bid-offer spread, which is why confidence gains are most reliable for lower-value or more standardized works, while unique and high-value pieces still depend on human judgment, relationships, trust, and context.