Architectural Heritage and Artificial Intelligence: Diagnosis and Solutions Proposed by ChatGPT for Algerian Historical Monuments

Abstract

This study explores the potential of artificial intelligence (AI), specifically ChatGPT, in enhancing the conservation of Algeria’s architectural heritage. By analyzing photographs of historical monuments, the research evaluates ChatGPT’s ability to identify architectural styles, detect pathologies, and propose conservation strategies. The findings reveal that while ChatGPT demonstrates proficiency in recognizing architectural features and generating general descriptions, its accuracy in identifying specific pathologies remains limited, with a certainty rate of only 40%. The tool’s reliance on textual data rather than direct visual analysis, coupled with its inability to meet specific academic requirements such as word count and accurate referencing, underscores its current limitations. However, the study highlights the potential of AI to complement traditional conservation methods, particularly when integrated with comprehensive databases and expert validation. The research advocates for a hybrid approach, combining AI’s efficiency with human expertise, to address the challenges of heritage preservation in Algeria. This work contributes to the growing field of AI applications in cultural heritage, offering insights into both the opportunities and constraints of leveraging AI for sustainable monument conservation.

1. Introduction

Cultural heritage helps forge national identity by strengthening collective memory and shaping national consciousness. It embodies the values and traditions that are rooted in the soul and spirit of the people, providing them with the creative resources needed to participate in the construction and development of society [1]. Each generation draws its existence from the heritage bequeathed by its predecessors and has a duty to pass this heritage on to future generations, ideally enriched and updated [2]. Algeria’s architectural heritage reflects its rich and diverse history, including Roman remains, Islamic works of art, Ottoman palaces, and French colonial buildings [3]. However, these precious elements face various challenges, such as environmental deterioration, the constraints of urbanization, and the limited resources available for their conservation.

The first text relating to built heritage dates back to 1967 [4]. From 1990 onwards, as a result of political openness and new orientations in the field of urban planning and development, the State became more concerned with the issue of urban and architectural heritage [5]. A number of laws have been enacted, including Law No. 98-04 of 15 June 1998, on the protection of cultural heritage [6]. Various initiatives have been undertaken, including the classification of 390 historic sites and monuments as “national heritage” as of 2003. In addition, a vast renovation program has been launched for buildings located in the historic centers of four major conurbations [5].

At the international level, Algeria joined the UNESCO Convention in 1973 [7]. Through its participation in the “Euromed Heritage” program, which was initiated in 1998, the organization works toward the promotion and protection of the architectural legacy shared by the many countries of the Mediterranean [8]. In the year 2004, it issued and ratified the Algiers Declaration on the Preservation of Peoples’ Identities and Heritage, which was a declaration on cultural diversity [5,9]. Since 2005, it has been involved in the Archimède project, which is a collaborative effort with seven other Mediterranean cities aimed at preserving and restoring ancient areas [5,10]. As an additional point of interest, the associative movement has commemorated World Heritage Month annually since the year 1999. During this month, it is actively involved in a variety of projects that are designed to promote heritage [5].

However, this precious heritage is today confronted with a variety of escalating dangers, including unplanned urbanization, extreme climatic phenomena, and a lack of resources and skills for its conservation. These threats are all expected to rise in number [11]. According to Philipparie [12], four major categories of phenomena can explain the origins of building disorders: climate, material behavior, non-compliance with best practices, and construction conditions. These last two, which encompass all the best practices and existing techniques in the field of construction, can be merged into what Watt & Swallow [13] consider as human factors. They also include atmospheric phenomena in the climate, as Carrie & Morrel [14] did before him. Finally, the behavior of materials considers their response to the various stresses and conditions to which they are subjected, particularly wear overtime or deformation due to humidity, temperature changes, etc. [12].

Even if there are efforts being made by national and international authorities, traditional preservation methods frequently fail to meet the requirements necessary to deal with the amount of the damage [15]. Faced with the growing challenges related to the degradation of historical monuments, the lack of specialized human resources, and the limitations of traditional preservation methods, artificial intelligence (AI) appears as a promising solution to optimize the analysis, diagnosis, and restoration of ancient buildings.

1.1. Artificial Intelligence in Built Heritage

Artificial intelligence (AI) is a major technological advance that is attracting growing interest and needs to be explored in depth to exploit its full potential [16]. Since its appearance several years ago, it has aroused great interest and continued to captivate the attention of many people. Reactions to artificial intelligence have been extremely varied, ranging from the most vibrant enthusiasm to legitimate fears [17]. Emerging technologies, including artificial intelligence, play an essential role in preserving and protecting the environment. Indeed, these technological advances make it possible to successfully meet the challenges that were once encountered in this field [18].

Artificial intelligence has already been used to conserve oral cultural heritage through illustrations [19] and to preserve vernacular cultural heritage, such as local craftsmanship, through an augmented reality application [20]. In architecture, artificial intelligence enables architects to design and engineers to compute, strategize, and construct structures with greater efficiency. Consequently, they can enhance concepts about sustainability and cost-efficiency while suggesting novel, inventive spatial and technical solutions [21]. New developments in computer vision (CV) and deep learning (DL) technology, along with artificial intelligence, now present the potential for expedited and more dependable identification of these disorders. Furthermore, these advancements provide more efficient management of long-term follow-up through data digitization [22]. Technologies such as laser scanning and photogrammetry assisted by artificial intelligence enable the creation of highly precise 3D surveys, in an increasingly automated manner, thus providing support to professionals during the inspection of existing built structures [23].

The acknowledgment of artificial intelligence (AI) is progressively increasing across various industrial domains, with chatbot technology emerging as a critical component [24,25]. This research will utilize ChatGPT, an AI-based tool developed by OpenAI, a non-profit research organization dedicated to advancing and guiding artificial intelligence technologies [26], a tool that claims to be capable of automatically generating original texts [27]. GPT is A.I. that generates texts like poetry, essays, tutorials, and social media captions. Additionally, this A.I. can generate a general understanding of a statement or request [26]. ChatGPT’s sophisticated language modeling skills could revolutionize our interactions with computers and technology by facilitating more natural and intuitive communication [28]. It can also produce coherent and contextually pertinent responses in several languages, facilitating the dismantling of language barriers and enhancing cross-cultural communication [29]. However, in this study, we will utilize English.

1.2. Objectives

This article proposes an innovative approach to the protection of architectural heritage in Algeria by leveraging the capabilities of artificial intelligence (AI) to improve the efficiency and accuracy of conservation interventions. By highlighting the advantages and limitations of this technology, this research aims to inspire new interdisciplinary collaborations and encourage the adoption of innovative solutions in the field of cultural heritage preservation.

Our research is guided by three key questions: (1) Can AI generate accurate and comprehensive structural analyses from user-provided images? (2) What are its strengths and limitations in identifying pathologies and proposing conservation methods? (3) How can AI be trained and effectively integrated into heritage preservation workflows? The objective of this research is to evaluate the potential of AI to produce a thorough and coherent analysis from a simple photograph provided by the user. To this end, a series of photographs of Algerian monuments was submitted to an AI model, specifically ChatGPT, to test its ability to generate detailed descriptions of the structures, identify visible pathologies, and propose suitable conservation methods.

This work is part of a rapidly expanding research field, where the convergence between technology and cultural heritage opens up new perspectives for the sustainable preservation of historical monuments. The originality of this study lies in its concrete application of AI to the Algerian context, a country where conservation challenges are particularly pronounced. Furthermore, it helps to fill a gap in the scientific literature by demonstrating how AI can be integrated into existing conservation practices, while highlighting the limitations and technical challenges to be overcome.

2. Materials and Methodology

2.1. Data Generation

For this study, OpenAI’s ChatGPT, specifically the ChatGPT-4-turbo version (https://chatgpt.com/, accessed on 14 January 2025, then on 21 March 2025), was employed to generate a detailed analysis of a photograph taken by the authors. The model was tasked with identifying the building materials used, determining the architectural style and historical period, assessing the building’s condition and pathological issues, and proposing the most appropriate intervention (e.g., conservation, restoration). The task was as follows: “Write a 1200-word essay that delineates the attached picture from Algerian built heritage, encompassing the architectural style and period, the materials employed in construction, the current condition and pathologies of the building, and ultimately, recommend suitable actions and methodologies for its conservation. Provides references from reliable and verifiable sources”.

On 14 January 2025, four distinct iterations of the task were carried out as part of this experiment. Due to system limitations restricting the execution of more than four tasks involving photos per day, the experiment was continued the following day with three iterations and subsequently with two on the day after.

To determine the minimum threshold of data required to ensure a more reliable response in terms of accuracy, it is essential to consider its implications for architectural heritage management. This issue is particularly relevant when assessing the potential use of AI by heritage managers who may not be expert architects or conservation specialists and who often have limited access to extensive datasets. However, they require efficient tools to streamline management operations, prioritize interventions, and conduct more precise analyses and planning.

To further explore this aspect, we chose three of the nine previous monuments to repeat the same experiment on 21 March 2025, but this time in two phases. In the first phase, we repeated the previous task but provided ChatGPT with three different images of the same building. In the second phase, in addition to the three images, we included the ICOMOS Glossary of the Main Pathologies Affecting Buildings [30,31]. We then asked ChatGPT to identify the pathologies and propose potential solutions. The second task was as follows: “Write a 1200-word essay that identifies and analyzes the structural pathologies present in the monument depicted in the three attached photographs. The essay should propose restoration and conservation solutions tailored to the identified pathologies, taking into account the information provided in the ICOMOS Glossary of the Main Pathologies Affecting Buildings as well as relevant insights from other scholarly sources and external documents”.

Each iteration followed a precise procedure aimed at evaluating the automated writing features offered by the system. Here is a detailed description of the methodology employed:

Initial submission of the task: The task to be performed was submitted to ChatGPT via the dialogue window located at the bottom of the user interface. This window allows for direct and continuous communication with the system.

Generation of the work plan: Once the task was submitted, the system immediately proceeded to create a preliminary work plan for the content. This plan constitutes an essential preparatory step, structured into several sections allowing for organized and coherent writing.

Transition to the writing space: Once the work plan is generated, the main dialogue window is transformed into a sidebar. Simultaneously, a new writing space opens, allowing ChatGPT to draft the essay by strictly following the sections defined earlier.

Automated writing: ChatGPT continues writing the text smoothly and continuously, without interruption, until the complete writing is finalized. Each section of the plan is addressed in detail and articulated according to the task requirements.

User interaction capability: A key feature of the interface allows the user to modify the text at any time. By selecting a word or an expression, a button titled “Ask ChatGPT” becomes available. By clicking this button, the user can either request a rephrasing or ask for additional explanations on a given passage.

Finalization and copying of the text: Without making any changes, the text generated by ChatGPT was copied to be analyzed later.

On the first day ChatGPT allowed us to upload only four photos, and consequently, the operation was repeated on 15 January 2025 with three photos, and on 16 January 2025 with two photos, following exactly the same steps. It is important to note that no direct feedback was provided to ChatGPT regarding the relevance or quality of the response. This lack of feedback allows for the evaluation of the system’s ability to operate autonomously. The same operation was conducted on 21 March 2025 in three iterations, with three photos uploaded during each. Throughout the operation, the chat window remained open, allowing for the consultation of previous responses. This feature ensures complete traceability of interactions, offering the user the possibility to go back and review or reuse previously provided information.

2.2. Data Analysis

The content of the essay manuscripts was analyzed in terms of topic coverage, structure, and coherence, but especially the accuracy of the information provided and the appropriateness of the proposed solutions. while technical issues were evaluated utilizing the editorial reporting functionalities of Word for Microsoft 365 Version 16.88 (24081116), which provide detailed information on the text, including word, paragraph, and sentence counts, along with readability metrics.

2.3. Data Documentation

All interactions with ChatGPT mentioned in this paper have been documented in compliance with the protocol below:

(a)

Documentation Tools: Use tools such as text editors (Word).

(b)

Metadata Tracking: Include session date, version of ChatGPT.

(c)

Question Structure: Use precise, well-structured, and grammatically correct prompts.

(d)

Task-Specific Instructions: Specify formatting requirements (essays).

(e)

Textual Logs: Copy and save the conversation as plain text.

(f)

Session Metadata: Include timestamp, and ChatGPT responses.

(g)

Accuracy: Assess the correctness of responses based on external references.

3. Results and Discussion

3.1. Technical Results

Although the task explicitly required the generation of a 1200-word essay, none of the produced texts met this criterion. A substantial proportion of the essays exhibited a considerable shortfall in length, with one third containing less than 50% of the required word count. The mean length of the generated essays was 715 words, demonstrating a significant deviation from the expected standard. Furthermore, the longest essay produced did not exceed 910 words, even when excluding references (Figure 1). This consistent failure to meet the specified word count raises concerns regarding the model’s capacity to generate comprehensive and sufficiently detailed academic content.

3.2. References Analysis

ChatGPT generated a total of 36 references across the nine essays, averaging four references per essay. After removing duplicate entries, we are left with 24 distinct references. These references can be categorized as follows (Figure 2): 17 invented references, three references exist but contain incorrect publication dates, and only four references are authentic and verifiable. The accuracy of all the references was verified by searching for the cited titles using both Google and Google Scholar.

A critical examination of the essays revealed that none of the references were explicitly cited within the body of the text. Instead, they were presented exclusively in the bibliography at the conclusion of each essay. This absence of in-text citations raises concerns regarding the reliability and academic rigor of the essays, as proper referencing is a fundamental requirement for scholarly writing.

Further analysis of the references invented reveals that their construction follows a pattern designed to enhance credibility. Many of these fictitious citations incorporate the names of actual researchers known for their work on the relevant subject matter. Additionally, the titles assigned to these references appear plausible, as they align with the established nomenclature of real academic journals. The majority of the invented references also attribute authorship to individuals who have published in the domains of heritage studies, Islamic and/or African architecture, and urbanism. Consequently, to an untrained reader who lacks familiarity with the specific literature, these references may seem legitimate and trustworthy.

Moreover, two thirds of the references are not given in alphabetical order, which leads us to understand that they are in order of appearance in the essay.

The references generated by the AI in one of the essays where we provided the photo (Figure 3) of the “Bey Palace” in Constantine, are shown in Figure 4.

The first reference provided exists on Google Scholar, while the second reference was fabricated by ChatGPT. The author cited in this reference has conducted extensive research on Islamic architecture and heritage in African countries. The title is directly related to the case study, making the reference appear highly realistic. Regarding the third reference, the cited author is indeed the rightful author of the book, which exists in two editions: the 1986 edition, later reformatted into a PDF version in 2005, and the 2017 edition (available on Google Scholar). However, ChatGPT incorrectly provided the year 2002. The last reference was also fabricated by ChatGPT; no document or section with this title exists on the UNESCO website.

3.3. Monument Identification

Analysis of the results obtained from the nine photographs supplied to ChatGPT reveals both the potential and the limitations of artificial intelligence in identifying and describing heritage monuments and buildings. Of these images, the model was only able to identify by name two monuments listed as UNESCO World Heritage Sites: “La Grande Poste d’Alger” and “La Casbah d’Alger”. This partial performance underlines the existence of gaps in ChatGPT’s databases, which influence its ability to recognize specific heritage sites, particularly those less documented or less publicized internationally.

However, despite this limitation, qualitative analysis of the descriptions generated shows that ChatGPT was able to provide detailed and accurate architectural descriptions. It was able to correctly identify the architectural styles of the buildings presented, correctly mentioning their structural and decorative elements. This ability to recognize specific architectural features suggests that, when coupled with a more comprehensive and structured database, ChatGPT could be a powerful tool for the analysis and documentation of built heritage.

Another notable feature is the model’s ability to establish correspondences between architectural styles and historical periods. By considering the explicit geographical context of the task—namely Algeria—ChatGPT was able to associate buildings with their respective periods in a globally consistent way. This indicates that the AI has sufficient contextual understanding to infer historical information from architectural clues. Nevertheless, this ability remains dependent on the quality and completeness of the data on which the model is based.

From the perspective of application to heritage management programs, these results highlight several major issues. On one hand, AI can be exploited as a tool to aid architectural analysis and the classification of historic buildings, notably by facilitating the identification of styles and decorative elements. On the other hand, imprecision in the identification of specific monuments highlights the need to integrate more detailed heritage databases adapted to national and regional contexts. Collaboration between heritage experts, cultural institutions, and artificial intelligence developers could thus improve the performance of these models and enhance their usefulness for the conservation and enhancement of architectural heritage.

Finally, the use of artificial intelligence in heritage management opens up promising prospects, notably for the automatic analysis of archive images, the recognition of damaged or endangered buildings, and the virtual reconstruction of historic sites. However, these advances require rigorous database enrichment and validation by experts to ensure optimum reliability of the results generated.

3.4. Pathologies Identification from the First Experiment

The word “pathology” originates from the Greek words “pathos” and “logia”, which mean the study of diseases [32,33]. The foundations of building pathology rest on a thorough understanding of the processes of design, construction, use, and modification of a building, as well as the various mechanisms that can affect its structural, material, and environmental condition [34]. Pathologies in buildings can emerge in diverse shapes and intensities, contingent upon the timing or nature of their occurrence [35]. The manifestations possess unique characteristics that, when systematically studied, facilitate the assessment of the nature, origin, and mechanisms of the relevant occurrences, therebyinforming an initial diagnosis [36].

In this research, ChatGPT was used to diagnose pathologies from nine photos of different historical monuments taken by the authors. Following a thorough analysis of the results generated by the artificial intelligence tool, it was observed that it is capable of recognizing various pathologies and providing general definitions and descriptions.

The first photo provided illustrates “La Grande Poste” (Figure 5). The artificial intelligence program detected five pathologies; however, upon thorough analysis, it was shown that three of them are not visible in the supplied photos of the monument. For the two existing pathologies, ChatGPT generated the following information.

• Structural Instability: Some sections of the facade show signs of cracks and detachment, potentially due to seismic activity or aging materials.

Actual observation: The facade of the “La Grande Poste” has recently undergone renovation. The cracks identified by ChatGPT are, in fact, not structural defects but rather shadows caused by the texture of the coating applied to the facade. However, minor partial cracks are present, which can be attributed to imperfections in the execution of the renovation work.

• Decay of Wooden Elements Windows and doors show signs of rot and termite infestation, compromising their integrity.

Actual observation: The wooden joinery and carpentry elements exhibit signs of insufficient maintenance and revitalization. This condition creates ambiguity when analyzing the elements in the photograph, suggesting a potential risk of cracking. However, upon closer examination, the issue is identified as the detachment of coatings rather than structural cracking.

During the analysis of the results of the second task, ChatGPT identified five pathologies, among which only two were visible in the photo provided:

• Structural Instability: Visible cracks and partial collapses suggest foundation weaknesses or seismic activity.

Actual observation: The photo highlights several structural issues on the facade, particularly cracks that appear to be spreading across various parts of the building. These cracks may result from multiple factors, such as ground movement, material aging, or construction defects.

Additionally, capillary rise is observed, which is often caused by moisture infiltrating the walls. This phenomenon is typically exacerbated by climatic conditions, especially in the Algiers region where humidity levels are high, and by a lack of regular building maintenance.

Regarding the Domes of the Grand Post Office: The AI was unable to detect the pathology of discoloration on the second dome, indicated by a blue rectangle, which represents the experts’ diagnosis. Additionally, other pathologies marked by blue rectangles in the photograph include the deterioration of the entrance door, micro-cracks on the marble columns, and the absence of certain decorative tile pieces.

The green checkmarks indicate the validation of pathologies correctly identified by the AI in the provided image.

In the case of the Casbah (Figure 6), a site of historical significance, neglect in maintenance can accelerate the degradation of its structures.

• Accumulation of Debris: fallen materials and rubble obstruct pathways and exacerbate degradation.

Actual observation: The debris visible in the photograph originates from the adjacent building marked by the blue rectangle (expert diagnostic), as the building in question remains intact and structurally complete. This suggests that the deterioration or collapse of the neighboring structure has contributed to the accumulation of debris, potentially due to poor maintenance and structural failure.

When analyzing the texts generated by ChatGPT regarding pathologies, it is evident that the AI identified five pathologies for each image, with a certainty rate of only 40% (red rectangle). This result highlights a significant limitation of ChatGPT in its ability to accurately identify and interpret pathologies from photos.

Several factors may explain this limitation. First, ChatGPT, as a language model, is not specifically designed for image analysis. Its interpretation of pathologies relies on textual descriptions or contextual information provided, rather than direct visual analysis. Additionally, the complexity of structural pathologies, which often requires technical expertise and in-depth knowledge of materials and environmental conditions, exceeds the current capabilities of AI.

Finally, the relatively low certainty rate (40%) suggests that the responses generated by ChatGPT should be treated with caution and validated by subject matter experts (green checkmark). This underscores the importance of combining the use of AI with human expertise to ensure accurate and reliable diagnostics in the field of building pathology.

3.5. ChatGPT Solutions’ Analysis from the First Experiment

Based on the results of pathology identification using the ChatGPT tool, it was observed that its certainty level is only 40%. This high margin of error raises concerns about the tool’s reliability in the field of architectural heritage conservation, where precision is crucial to preserving the integrity of historic structures.

Such a low certainty rate directly impacts the conservation strategies proposed by ChatGPT, as well as the relevance of the solutions it suggests. Inaccurate diagnostics can lead to inadequate or even harmful interventions for historic buildings. For example, in the case of the Casbahoui house (Figure 6), the solutions proposed by ChatGPT, although potentially useful, must be considered with caution due to the lack of reliability in the initial diagnostics.

Among the solutions suggested by ChatGPT for the Casbahoui house are recommendations such as the following.

Conservation Strategies and Recommendations: to ensure the longevity of the building while maintaining its historical integrity, the following conservation strategies are recommended:

• Structural Stabilization: Reinforcing weakened walls and foundations with compatible traditional materials to preserve authenticity.

This solution remains vague, as the photo clearly shows that the building is constructed using dry stone techniques. The appropriate solution would involve reusing in situ materials to reconstruct the edifice, while respecting vernacular craftsmanship. This includes traditional techniques such as stabilizing walls with wooden logs and adhering to wall thickness dictated by the building’s height. Such an approach would not only preserve the authenticity and structural integrity of the building but also highlight and valorize local and historical construction methods.

• Material Restoration:

  ∘

  Cleaning and reapplying traditional lime plaster.

  ∘

  Treating and replacing decayed wooden components with historically accurate materials.

  ∘

  Repairing or recreating lost decorative tilework.

Following a thorough analysis of Photo 2, it is evident that the solutions proposed by ChatGPT are standardized and could be applied to this type of pathology in any building, regardless of the construction material used. ChatGPT failed to consider the specificities and details visible in the photo, such as the dry-stone construction technique, traditional stabilization methods, or the unique architectural features of the building. While this generic approach may be useful in certain contexts, it overlooks the importance of a contextual and detailed analysis, which is essential for tailored and heritage-respectful interventions. As a result, the proposed solutions lack precision and relevance in addressing the specific needs of this particular case.

In similar cases, such as that of the Venice Charter, where the pathology of structural instability is evident, and with the same materials used in the Casbah of Algiers, specialists in the field have opted for a thorough analysis of the monument. This approach ensures a comprehensive understanding of structural issues and facilitates the development of tailored conservation strategies that respect the historical and architectural integrity of the site.

The differences between ChatGPT response and expert opinions regarding “La Grande Poste” and “Casbaoui House” are shown in

Table 1. Summary Table: a comparison of AI and expert assessments. Source: authors.

Monument	Pathology Identified by ChatGPT	Expert Observation	Accuracy of AI Diagnosis	Relevance of AI’s Suggested Solutions
La Grande Poste (Figure 5)	Structural Instability: Cracks and detachment due to seismic activity or aging materials.	The building was recently renovated; the cracks detected were actually shadows from the texture of the facade. Minor cracks are present but result from renovation imperfections.	Partially inaccurate (AI misidentified shadows as cracks).	Generic; does not consider recent renovation or construction context.
	Decay of Wooden Elements: Rot and termite infestation in windows and doors.	The wooden elements show signs of insufficient maintenance but no evidence of termite infestation.	Partially inaccurate (correctly noted maintenance issues but misidentified the cause).	Somewhat relevant but lacks specificity.
Casbah (Casbaoui House) (Figure 6)	Structural Instability: Cracks and partial collapse due to foundation weaknesses or seismic activity.	Cracks observed; potential causes include material aging, ground movement, and construction defects. Capillary rise also noted due to high humidity and lack of maintenance.	Mostly accurate (AI identified structural issues but missed capillary rise factor).	Generic; does not account for traditional building techniques.
	Accumulation of Debris: Fallen materials and rubble exacerbating degradation.	The debris originates from an adjacent building, not the structure in question.	Inaccurate (incorrect source of debris identified).	Not relevant; AI misattributed the cause.
	Conservation Strategies Suggested by AI: Structural stabilization with traditional materials, lime plaster restoration, treatment of wooden components, tilework repair.	AI overlooked key architectural details, such as dry-stone construction and traditional stabilization techniques. Authentic conservation requires reusing in situ materials and respecting vernacular craftsmanship.	Partially relevant (suggestions are broadly applicable but not tailored to the site’s unique construction methods).	Overly standardized; lacks specificity for this heritage site.

.

3.6. Analysis of the Results of the Second Experiment

To establish a more rigorous conservation strategy for the site, a second experimental approach was introduced. This involved supplying ChatGPT with an article discussing the relevant phenomena, allowing the model to refine its responses and expand its sources. By incorporating more technical details and specific references from the ICOMOS (Glossary of the Main Pathologies Affecting Buildings), ChatGPT was able to provide well-supported recommendations for conservation and restoration interventions.

During the second phase of the experiment, the number of photographs provided was increased to three images of the “La Grande Poste” (Figure 5, Figure 7 and Figure 8), each taken from different angles. The confidence level in detecting the types of pathologies present was exceptionally high, exceeding 90%. This suggests that the AI tool effectively utilized these images to conduct a comprehensive diagnostic assessment of the monument’s structural issues.

The findings indicate that the AI tool successfully enhanced the reliability and accuracy of the information provided. By leveraging the input data, it generated a more detailed and substantiated analysis, demonstrating its ability to improve diagnostic precision and support evidence-based conservation strategies.

A similar level of accuracy was observed in the case of the second monument “Djamaâ el Djedid” (Figure 9, Figure 10 and Figure 11). This reinforces the conclusion that even a minimal dataset can be sufficient to effectively train the AI tool, enabling it to generate precise and technically detailed responses. These results suggest that AI can serve as a valuable resource for non-specialists in architectural heritage, equipping them with the necessary information to formulate informed conservation plans. The tool’s ability to bridge knowledge gaps and facilitate well-informed decision-making underscores its potential in heritage preservation efforts.

Regarding the third iterations involving the “Ketchaoua” mosque (Figure 12, Figure 13 and Figure 14), no significant difference in accuracy was observed between Task 1 and Task 2 in the second experiment in terms of certainty. However, supplementing the input data with additional documents or resources on the specific topic or phenomenon in question is recommended. This approach ensures a higher quality of information and maximizes the benefits of AI. The integration of complementary sources enhances the tool’s analytical depth, resulting in more comprehensive and well-supported outcomes, ultimately optimizing its application in technical and heritage conservation contexts.

This second experimental phase examined how supplemental visual and textual references affect AI performance. Quantitative and qualitative analyses demonstrate that both image diversity and glossary integration contributed substantially to improved model accuracy, though through distinct mechanisms:

Impact of Image Variety: The inclusion of varied architectural images enhanced the AI’s contextual understanding, particularly in identifying and differentiating structural elements. This suggests that visual diversity strengthens pattern recognition in domain-specific applications.

Role of Glossary Integration: the reference glossary significantly improved terminological precision and conceptual consistency in the AI’s outputs, mitigating ambiguities prevalent in unstructured inputs.

Synergistic Effect: The combination of both factors produced optimal results, indicating that multimodal inputs (visual + lexical) create a compounding effect. This aligns with the existing literature on comprehensive contextual learning in machine perception.

These findings underscore the importance of multifaceted training data for domain-specific AI tasks, while highlighting avenues for future research into weighing mechanisms for hybrid input types.

3.7. AI Training for Improved Outcomes

Based on our previous analyses, we observed that the AI tool performs more effectively when provided with detailed input. This indicates that the richness and quality of the dataset directly influence the accuracy and reliability of the results. Consequently, the training of AI models is crucial, and can be achieved through several approaches:

Feedback Mechanisms: Providing detailed feedback on the generated responses, particularly regarding the accuracy of the information provided, is essential. This iterative process helps refine the model’s performance over time by correcting errors and reinforcing accurate outputs.

Training AI Models on Specialized Datasets: Developing AI models trained on specialized datasets—such as annotated images of architectural pathologies, historical building materials, and conservation techniques—would significantly enhance their ability to recognize and analyze heritage-related issues with greater precision.

Integration of Expert Knowledge: Supplementing AI tools with domain-specific knowledge, such as integrating glossaries (e.g., the ICOMOS Glossary of Pathologies) or collaborating with conservation specialists, can help refine the model’s outputs and ensure their alignment with expert standards.

Hybrid Approaches: We propose a hybrid methodology where AI-generated analyses are cross-verified and refined by human experts. This approach ensures a balance between the efficiency of AI and the precision of expert judgment, ultimately improving the reliability of the results.

These strategies highlight the importance of combining advanced AI capabilities with human expertise to address the limitations observed in our study and to enhance the application of AI in architectural heritage management.

3.8. Deep Learning for Pathology Detection: Leveraging AI and Computer Vision for Enhanced Analysis

Artificial intelligence (AI), a branch of data science, encompasses various disciplines aimed at equipping machines with human-like intelligence. Among these, machine learning (ML) develops algorithms that enable systems to learn and improve from data without explicit programming. ML approaches include supervised, unsupervised, and reinforcement learning. In pathology detection, supervised learning is employed, as operators precisely define the training data [37]. Deep learning (DL), a subset of ML, uses deep artificial neural networks to model complex structures from large datasets. Inspired by the human brain, DL excels in tasks like image processing, speech recognition, and translation [37].

In pathology detection, DL leverages computer vision, enabling machines to interpret visual data from images or videos, mimicking human visual perception. The process begins with image acquisition using cameras or photos, followed by algorithms extracting key features (e.g., edges, shapes). These features are processed by models to generate predictions [37,38].

A recent advancement in pathology detection is instance segmentation, an extension of semantic segmentation. Unlike semantic segmentation, which classifies pixels, instance segmentation distinguishes individual pathology instances within the same class. This allows for precise identification, categorization, and quantification of defects, enabling detailed analysis. Models like Mask R-CNN and SOLO are commonly used for this task [39]. Some studies have achieved comparable results by combining object detection and semantic segmentation algorithms.

4. Conclusions

This study is part of research exploring the application of new artificial intelligence (AI) technologies, specifically ChatGPT, in the field of architectural heritage conservation, with the aim of assessing its ability to enhance heritage protection policies, particularly through the analysis and interpretation of visual data.

Following several tests conducted on ChatGPT, it was observed that this tool requires high-resolution photographs to function optimally. Image quality and the angle of capture, as well as lighting conditions and shadows, play an indispensable role, as they can influence the analysis and lead to errors in interpretation. Therefore, particular attention must be paid to these parameters to ensure reliable results.

Beyond image quality, the study underscores the necessity of a comprehensive and structured database to enhance AI’s capacity for pathology identification and conservation decision-making. Unlike human experts, who can draw on contextual knowledge and experience to infer conclusions from a limited dataset, machine learning algorithms require vast amounts of well-organized information to achieve comparable levels of accuracy. However, when supplied with sufficient data, AI systems can perform complex analytical tasks with remarkable efficiency, offering a scalable approach to heritage documentation and monitoring.

However, the limitations observed in this study should not be viewed as dead ends but rather as opportunities for further research and innovation. For instance, the integration of deep learning techniques could significantly enhance AI’s ability to detect and analyze architectural pathologies. Additionally, hybrid approaches that combine AI-generated analyses with expert verification could bridge the gap between technological efficiency and human precision, ensuring more reliable and contextually appropriate outcomes.

The implementation of a methodology based on artificial intelligence (AI) could be adopted by authorities and relevant organizations to enhance the protection of architectural heritage and ensure its transmission to future generations. Such an approach would also facilitate real-time access to information, offering innovative tools for the management and preservation of heritage. This methodology would improve the accuracy and efficiency of built heritage inspections by reducing on-site time and optimizing the quality of data collected during surveys. Traditional and AI-based technologies thus prove to be complementary, each bringing specific advantages to the conservation process.

However, despite continuous advancements in surveying techniques, the visual inspection of existing buildings remains an essential initial step in the in-depth study of these structures. This process remains largely manual, time-consuming, and costly, heavily reliant on the expertise of the operators involved. Moreover, the quality of inspections can vary over the long-term depending on the training and skills of the operators, highlighting the need for more standardized and automated solutions to ensure consistent and reliable assessments.

In conclusion, this study does not present the limitations of ChatGPT as insurmountable but rather as a call to action for further refinement and adaptation of AI technologies in built heritage conservation. The findings highlight the need for a multi-faceted approach that integrates AI with existing conservation practices, ultimately fostering more efficient, data-driven, and scalable heritage management strategies.