\section{AI and the evolution of hedonic valuation}
\label{sec:ai}

The screening identified \nIncludedMlPredictiveStageOne{} machine-learning or automated-valuation studies of residential prices at
stage one, and Figure~\ref{fig:ai_evolution} traces the rise of tree ensembles, neural networks, computer vision and text methods
in the included literature since 2000. Only a small fraction of these studies could be retrieved in open access, and the
dedicated synthesis rests on \nAiStudies{} studies extracted under a light protocol that records, for each paper, the methods
compared, the best model, the evaluation metric and its value, the hedonic benchmark on the same test set when one is printed,
the test-set design, the AI-derived features used, the incremental gain from those features when reported, and whether
interpretable marginal effects are reported (Table~\ref{tab:ai}). Three conclusions follow, each bounded by the small sample.
First, when a hedonic benchmark is evaluated on the same test set (\nAiBenchmark{} studies), tree boosting and random forests
usually predict better: among the \nAiMapeCmp{} studies reporting mean absolute percentage errors for both, the median error is
\aiMapeBestMedian{} percent for the best model against \aiMapeHedMedian{} percent for the hedonic benchmark, and among the
\nAiRsqCmp{} studies reporting $R^2$ the medians are \aiRsqBestMedian{} and \aiRsqHedMedian{}; a few studies find the hedonic
model competitive or better, usually with small samples. Only \nAiHoldout{} studies use a true holdout and \nAiNoTest{} report
in-sample fit only; one study shows that random cross-validation overstates out-of-time accuracy relative to temporal validation.
Second, the incremental information of AI-derived characteristics is documented in only \nAiIncremental{} studies, and it is
modest: text embeddings, street-view or floor-plan features add between a few thousandths and a few hundredths to $R^2$ or
reduce prediction errors by fractions of a percentage point, with the largest gains for floor plans and interior-condition
scores; one study finds that a natural-language sea-view indicator adds nothing to fit but identifies an eight percent premium
that a GIS viewshed misses, which illustrates that the value of AI-derived variables lies as much in interpretable new
characteristics as in accuracy. Third, \nAiShap{} studies report SHAP values and \nAiCoef{} report regression coefficients; the
paper does not pool SHAP values with coefficients, because a SHAP value is a model-specific attribution for a particular
observation set, not an implicit price, and the marginal effects extracted from these studies are kept in the AI-derived branch
of the taxonomy where they appear in Table~\ref{tab:taxonomy}. No comparison across studies with incompatible datasets is made.

\begin{figure}[H]
\centering
\includegraphics[width=0.9\textwidth]{fig13_ai_methodology_evolution.pdf}
\caption{Methodological evolution in the included literature: records mentioning each class of method in title or abstract
(three-year moving average).}
\label{fig:ai_evolution}
\end{figure}

\begin{table}[H]
\centering
\caption{AI-enhanced valuation evidence (light extraction; metrics as printed, never compared across studies)}
\label{tab:ai}
\scriptsize
\resizebox{\textwidth}{!}{\input{tables/tab13_ai_evidence}}
\end{table}