A new arXiv preprint (2607.09578v1), posted 13 July by researchers at the German Research Center for Artificial Intelligence, argues that AI in urban mining should be judged not by prediction accuracy but by whether its decisions can survive regulatory scrutiny [S1][S2]. The paper proposes four ways to fuse knowledge graphs with explainable AI to make that defensibility possible [S1][S2]. If the authors are right, the benchmark culture that dominates AI research is measuring the wrong thing — and regulated industries are paying the price.

The audit nobody optimises for

Urban mining means recovering materials from buildings before demolition. It hinges on what the authors call "pre-demolition assessment," a regulated audit where qualified professionals decide what is inside a structure, what is hazardous, and what can be reclaimed [S1][S2]. The paper frames this as fundamentally an information problem, not a prediction problem. An AI that guesses a fire door's composition with high accuracy is useless if it cannot show its working to the auditor who signs off, and who remains legally accountable for the call [S1][S2].

The authors, Jan Gronewald, Andreas Emrich, and Nijat Mehdiyev from DFKI in Saarbrücken [P3], argue that the relevant measure of value in this setting is "defensibility" — broken into four properties: legibility (can a human follow the reasoning?), plausibility (does the conclusion make sense?), sourcing (where did each piece of evidence come from?), and contestability (can an auditor push back and revise?) [S1][S2].

None of those properties show up in a standard accuracy benchmark.

Four modes, one fire door

The paper proposes four "integration modes" for combining knowledge graphs (structured databases of relationships between entities) with explainable AI techniques [S1][S2]:

  • Lifting: raising raw explainability outputs into a knowledge-graph structure so they can be queried and traced.
  • Constraining: using the knowledge graph to bound what the AI can claim, preventing unsupported assertions.
  • Typing: assigning formal types to AI explanations so they slot into known regulatory categories.
  • Revising: feeding auditor corrections back into the knowledge graph so the system learns from contested decisions.

Each mode, the authors write, is a specific typed operation over explainability artefacts and knowledge-graph structures, not a vague aspiration [S1][S2]. Each targets a different property of defensibility. Together, they aim to produce the kind of regulatory artefact that a pre-demolition audit actually demands [S1][S2].

The paper illustrates all four with a single example: a fire door in a building slated for demolition. Using the W3C Linked Building Data stack, a set of web standards for describing building components, plus valuation extensions, the authors walk through how each mode would handle the door's identification, material composition, safety classification, and recovery value [S1][S2].

What it means

The core argument flips a deep assumption in AI development. Most AI research optimises for prediction accuracy: getting the right answer as often as possible. This paper says that in regulated settings, the right answer is not enough. What matters is whether a qualified human can understand, trust, source, and challenge the reasoning behind it.

This connects to a broader shift. Knowledge graph construction is an active research front: tools like AutoSchemaKG [P4] and agent frameworks like UrbanKGent [P5] are automating the building of structured knowledge representations. On the explainability side, projects like Microsoft's Debug-XAI [P6] are turning interpretability methods into practical debugging tools. This paper sits at the intersection, arguing that neither knowledge graphs nor explainable AI works alone in regulated contexts. They need to be fused in specific, typed ways.

The paper grounds its argument in complementarity theory from the information-systems resource-based tradition: the idea that two resources can produce more value together than either could alone, but only when combined in the right way [S1][S2]. The four modes are the authors' attempt to specify what "the right way" looks like.

What it means for business

For a demolition firm or environmental consultancy, the practical question is whether AI can reduce the cost of pre-demolition audits without increasing legal exposure. Today, these audits are labour-intensive. Qualified assessors physically inspect buildings, sample materials, and produce reports that regulators and contractors rely on.

If the four modes described in this paper were implemented, a two-person consultancy could use AI to draft initial material assessments, with the knowledge graph constraining the AI to only make claims backed by sourced data. The auditor would review, challenge, and revise. Those revisions would feed back into the system. The AI would not replace the auditor. It would produce work the auditor could actually defend.

That is the theory. The paper offers no performance metrics, no cost analysis, and no field trial. A firm considering this approach would need to build the knowledge-graph infrastructure first, no small task, and integrate it with whatever explainability tools their AI stack already supports.

What we don't know yet

This is an unpeer-reviewed preprint with no empirical validation [S1][S2]. The four modes are conceptual, illustrated with a single fire-door example rather than tested across real demolition projects. Several questions remain open:

  • Do the four modes cover all the defensibility requirements of real pre-demolition audits, or are there gaps a working auditor would immediately spot?
  • How much effort does building the knowledge-graph substrate require, and does that cost pay off for small firms?
  • How do the modes perform when the underlying AI model is wrong? Does "Constraining" actually catch hallucinations, or just the obvious ones?
  • The paper draws on the W3C Linked Building Data stack, but building codes and audit requirements vary by jurisdiction. The framework's portability across regulatory regimes is untested.

The next concrete signal to watch: whether the authors follow this conceptual paper with an empirical study, and whether any demolition or environmental consultancy partners with DFKI to pilot the framework on a real building.

Sources: arXiv cs.AI official RSS feed [S1]; arXiv preprint 2607.09578v1 [S2]; arXiv HTML full text [P3]; AutoSchemaKG repository [P4]; UrbanKGent preprint [P5]; Microsoft Debug-XAI repository [P6].

If this made you think differently about what AI should be measured on, subscribe for more stories that decode the research behind the headlines.

Sources

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.