Can AI agents safely assess a historic minaret?
A Turkish case study combined vibration sensors, inspection images and retrieved engineering records to prioritise parts of one sixteenth-century minaret for inspection. It did not verify damage, test safety decisions or establish accuracy on other structures.
The short version
3 key points
- 1.The proof of concept used four vibration channels from one 23.2-metre minaret, with Measurements 1–12 defining the reference baseline and Measurements 13–20 used for anomaly-oriented evaluation.
- 2.A deterministic signal pipeline produced anomaly and zone scores; language and vision agents interpreted those scores and images, while retrieved project records constrained the final explanation.
- 3.There were no independent ground-truth damage labels, no accuracy benchmark and no second structure, so the outputs are inspection priorities rather than verified diagnoses or safety findings.
The answer: useful for inspection triage, not a structural verdict
The study shows that several kinds of evidence can be organised into a traceable inspection brief for a historic building. It does not show that an AI agent can certify the building as safe, identify hidden damage accurately or replace a conservation engineer. The system flagged changes in vibration relationships, examined linked images and consulted a small collection of project records before assigning zone-level priorities. Every final result could be traced back to those separate evidence streams.
That distinction matters because an anomaly is not automatically damage. Temperature, wind, traffic, sensor drift, construction nearby and changes in operating conditions can alter a structure’s measured response. A vision-language model can also confuse stains, mortar texture, restoration traces or lighting with cracking and spalling. The authors therefore present the output as a way to decide where engineers should look next, not as an independently confirmed description of the minaret’s physical condition.[1]
What was measured on the İskenderpaşa Mosque minaret
The case study concerns the masonry minaret of the İskenderpaşa Mosque in Trabzon, Türkiye, a structure associated with a mosque commissioned in 1529 or 1530. The project profile describes a combined cut-stone and brick minaret about 23.2 metres high. The analysis mapped the structure into six zones — base, lower shaft, transition, balcony, upper shaft and cap — so that sensor and image findings could be reported in terms an inspection team could use.
Four vibration channels supplied twenty measurement records. The first twelve were treated as healthy reference measurements for model learning and baseline estimation; Measurements 13–20 formed the evaluation period. Missing values were interpolated where possible and unresolved invalid rows removed. Power spectral density was used to estimate dominant frequencies, after which the software selected windows long enough to preserve several vibration cycles. Each window was detrended, robustly scaled and compressed to reduce the influence of extreme amplitudes.[1]
The numerical pipeline and the AI agents had different jobs
A one-dimensional convolutional autoencoder learned to reconstruct multichannel windows from the reference period. Reconstruction error then supplied the global anomaly score for both reference and evaluation windows. Separate autoencoders modelled relations between sensor pairs, allowing the workflow to ask not only whether the whole signal looked unusual but also whether the coupling between particular parts of the minaret had changed. Predefined sensor-to-zone mappings converted those relations into inspection-oriented zone scores.
The language model did not calculate reconstruction errors or set the numerical thresholds. One agent interpreted the signal summaries; a vision-language agent reviewed sensor-linked inspection images and assigned visual severity and confidence; a third agent fused their structured outputs. This separation is a strength because the deterministic calculations can be replayed without asking a language model to invent numbers. The AI layer’s role was explanation, conflict handling and report generation — activities that still require review because plausible wording is not engineering validation.[1]
Retrieval made the interpretation more conservative, not proven more correct
The retrieval component used a curated project memory containing geometry, material information, previous interventions, baseline dynamic behaviour, finite-element notes and an inspection playbook. Passages were selected through lightweight lexical matching for zones, sensors and evidence terms; the study did not use a vector database or retrain the models. Retrieved material constrained interpretation after the numerical scores had already been produced, helping agents recognise known texture or structural context instead of treating every visual irregularity as new damage.
Compared with a run using only current evidence, the retrieval-grounded run produced a narrower and more conservative risk assignment by reducing the influence of ambiguous images. That is evidence that context changed the output, not evidence that it improved accuracy. Because no independent inspection labels existed for the evaluation period, the study could not calculate accuracy, precision, recall or false-alarm rates. A safer conclusion is that retrieval may make the reasoning easier to audit while its effect on correct engineering decisions remains untested.[1]
What this could mean for conservation teams and the public
For conservation teams, the immediate opportunity is prioritisation. Historic towers are difficult to inspect repeatedly, interventions must be minimally invasive and records are often fragmented across surveys, drawings and past repair notes. A system that keeps vibration evidence, photographs and structural memory separate but linked could help an engineer explain why a particular balcony, transition or shaft zone deserves closer examination. Stored JSON records, image findings and risk maps also make it easier to reconstruct how a recommendation was produced.
For worshippers, neighbours and visitors, the benefit would depend on whether earlier inspection actually catches consequential deterioration without producing disruptive false alarms. Neither outcome was measured. A high-priority AI output should trigger a qualified inspection, not evacuation or repair by itself; a low-priority output should never overrule visible damage, new cracks or a professional safety concern. Accountability must remain with the engineers and public authorities who can inspect materials, foundations and connections the model cannot see.[1]
Evidence limits and the test that comes next
This is a peer-reviewed, open-access paper with a clearly described data split and a deliberate separation between deterministic scoring and model-generated interpretation. Yet it is still one structure with one geometry, one intervention history and uncertain support conditions. The evaluation compares two configurations inside the same deployment. The dataset is available from the corresponding author on reasonable request rather than as an immediately downloadable benchmark, which limits independent replay at publication time.
The next decisive test is prospective and independent. Researchers should lock the pipeline, collect repeated measurements across weather and operating conditions, obtain time-matched inspections from multiple structural specialists and test several monuments with different materials and forms. They should report false alarms, missed damage, stability across prompts and model versions, and whether using the system changes inspection time, cost or safety outcomes. Until that happens, this is a credible workflow demonstration — not evidence that AI has learned to diagnose cultural heritage.[1]
What this means for people
- Conservation engineers could receive more traceable inspection priorities, but they remain responsible for confirming physical damage and deciding action.
- Visitors and local communities should not interpret an AI risk map as a safety certificate or an evacuation order; neither decision was tested.
- Public heritage bodies would need model-version records, sensor maintenance, repeatable inspections and a clear appeals or escalation route before operational use.
Global context
Historic towers, churches, mosques, temples and civic monuments face similar access and monitoring constraints, but masonry, climate, seismic exposure, maintenance history and construction practice vary substantially. A single Turkish minaret cannot establish performance elsewhere. The study is valuable precisely because it makes a regional heritage case visible while keeping the claimed scope narrow.
What the evidence does not yet show
- One historic minaret was studied, with no external structure-level validation.
- No independent, time-synchronised ground-truth damage labels existed for Measurements 13–20.
- The with-RAG comparison measured a change in risk assignment, not an improvement in diagnostic accuracy.
- Vision findings remain vulnerable to surface texture, lighting, restoration traces and other appearance confounders.
- The data are available on request rather than as a public benchmark at publication time.
What to watch next
- Locked prospective validation on several monuments and construction types
- Multiple-engineer inspection labels and measured disagreement between specialists
- False-alarm and missed-damage rates across seasons, weather and sensor drift
- Prompt, model-version and retrieval-memory stability tests
- Evidence that the workflow improves inspection decisions, time or public safety
Living evidence record
Impact record IAI-1BXDIRH
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
15 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Scientific Reports published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 15 October 2026
Research question, topics and regions
Whether retrieval-grounded AI agents can combine vibration, image and structural-memory evidence into auditable inspection priorities for historic masonry
Topics: Cultural heritage · Structural monitoring · AI agents · Vision-language models · Retrieval-augmented generation · Human inspection
Regions: Türkiye · Middle East · Europe · Global cultural heritage
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Watch the background, then check the evidence
Google DeepMind
YouTube loads only when you choose. Press play in the player to start. If playback is unavailable, use “Watch on YouTube” below.
AI & scientific discovery · Background explainer
Published 2020-11-30 · 1 min 52 sec
Protein folding explained
A short introduction to why a protein’s shape matters. It provides background for understanding what AI structure prediction contributes to biological research.
Evidence check: This is the developer’s explanation. Predicting a structure does not by itself establish a treatment’s safety or benefit for patients.
Source checked 14 October 2026. Check available captions on the publisher’s video page.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
Science & Research
Can an AI choose your next experiment before you spend the compute?
A new preprint tests training intuition. We compare its narrow task with engineering, replication and architecture-search benchmarks to explain what research teams should measure next.
6 min · 4 sources
Science & Research
Do AI agents follow their plans? What a benchmark found
An unreviewed preprint found gaps between AI agents’ stated plans and their actions. Structured execution helped on selected benchmarks; real-world reliability remains unproven.
6 min · 2 sources
Science & Research
Can AI keep living cells under the nanoscope longer?
Two reconstruction strategies cut light dose tenfold or increased frame rate fourfold in RESOLFT experiments. They extend what researchers can observe, but computational restoration still needs experiment-specific validation.
7 min · 2 sources
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.