Management and Future Technologies | Volume 2, Issue 4 | № 008

RadLex, SNOMED CT), or automated report parsing—have partially mitigated these issues. More recent datasets, including PMC-CLIP and MedPix, aim to bridge these gaps by combining multiple modalities and linking images with domain-specific text from biomedical literature. The emergence of synthetic and weakly supervised datasets, created via large-scale caption generation or knowledge distillation from language models, also represents a promising approach to overcome data scarcity while preserving patient privacy. However, ensuring that synthetic data accurately reflect real-world clinical variability remains an open challenge. 4.2 Dataset Curation Challenges and Future Directions Building multimodal medical datasets involves complex ethical, technical, and operational considerations. The sensitive nature of medical data introduces stringent privacy regulations such as HIPAA, GDPR, and national health data protection laws, which restrict public sharing of clinical images and textual records. De-identification and anonymization techniques are therefore critical but often imperfect—text reports may still contain subtle identifiers like patient age, timestamps, or institution-specific phrasing. In addition, imaging devices and acquisition protocols vary across hospitals and manufacturers, causing domain shifts that degrade model performance when transferred between datasets. For instance, a model trained on MIMIC-CXR may not generalize well to CheXpert or private hospital data due to differences in scanner settings, population demographics, or report writing conventions. These issues underline the importance of domain adaptation and federated learning strategies, which allow VLMs to learn from distributed datasets without centralizing sensitive patient data. Another major challenge is the imbalance and bias inherent in clinical datasets. Disease prevalence in hospital records is skewed toward common conditions, leading to underrepresentation of rare or subtle pathologies. Similarly, demographic biases—across age, gender, or ethnicity—can cause systematic performance disparities, potentially amplifying existing healthcare inequities. Addressing these issues requires deliberate curation and bias auditing at the dataset level. Recent initiatives emphasize ethical dataset design through balanced sampling, metadata transparency, and clinician involvement in annotation processes. Furthermore, the integration of multimodal temporal data (e.g., longitudinal imaging and follow-up notes) offers a richer representation of disease progression but introduces new complexity in data synchronization and alignment. Advances in data governance, privacy-preserving computation, and synthetic data generation are expected to play a central role in expanding accessible, diverse, and ethically compliant datasets for training trustworthy medical VLMs. Future research trends are shifting toward comprehensive multimodal ecosystems that integrate not only image and text but also structured data such as laboratory tests, genomics, and clinical notes. Projects like MedPerf and MONAI exemplify collaborative frameworks that promote reproducibility and benchmarking across institutions. Synthetic data generation using diffusion models and text-toimage transformers further enhances dataset diversity, allowing the simulation of rare conditions or underrepresented demographics. Additionally, the rise of weakly supervised and self-supervised pretraining enables models to learn multimodal correlations without relying heavily on fully annotated data. Such strategies may revolutionize dataset creation, democratizing access to multimodal medical AI research while ensuring adherence to ethical and regulatory constraints 80

4.3 Evaluation Metrics for Report Generation and VQA Evaluating Vision-Language Models in medicine requires a careful balance between quantitative performance metrics and clinical relevance. Unlike traditional computer vision tasks, where accuracy or mean average precision suffice, multimodal tasks such as report generation and VQA demand multi-dimensional evaluation encompassing linguistic, semantic, and diagnostic consistency. For report generation, standard NLP metrics—such as BLEU, ROUGE, CIDEr, and METEOR—are commonly used to measure n-gram overlap between generated and reference reports. However, these metrics capture surface-level linguistic similarity and often fail to reflect the clinical correctness of generated text. To address this limitation, domain-specific metrics have been developed. For example, the Clinical Efficacy (CE) and RadGraph F1 metrics evaluate factual accuracy by comparing extracted clinical entities and relations between generated and reference reports. The CheXbert score, based on automated label extraction, measures agreement with groundtruth disease findings and provides a more meaningful proxy for diagnostic validity. For VQA in medicine, evaluation involves both exact match accuracy and semantic equivalence between predicted and ground-truth answers. Datasets like VQA-RAD and PathVQA employ accuracy and BLEU-based measures, while recent studies propose expert-based grading or semantic similarity metrics using medical language embeddings to capture nuanced correctness. Additionally, human evaluation by clinicians remains essential for assessing interpretability, justification quality, and clinical usefulness of model outputs. In research practice, hybrid evaluation protocols that combine automatic scoring with expert review have proven effective for benchmarking multimodal systems. Beyond task-specific metrics, broader assessments of model robustness, fairness, and uncertainty calibration are increasingly recognized as vital components of evaluation. For instance, calibration metrics such as Expected Calibration Error (ECE) can quantify how confidently models make predictions, which is crucial in safety-critical domains like healthcare. A comprehensive evaluation framework must therefore extend beyond text–image alignment or linguistic fidelity to encompass trustworthiness and clinical impact. Metrics assessing explainability, such as attention visualization overlap with radiologist-marked regions, are gaining traction as proxies for model transparency. Furthermore, emerging standards encourage multi-faceted evaluation protocols that integrate quantitative metrics, expert review, and uncertainty quantification to ensure that VLM outputs are both technically robust and clinically meaningful. As the field progresses, community-driven benchmarks—such as RadBench, MedVQA-Eval, and MedRGen— are expected to play a pivotal role in harmonizing evaluation practices, fostering reproducibility, and guiding the development of models that meet real-world medical standards. Ultimately, the effectiveness of any medical VLM depends not only on its architectural sophistication or dataset scale but also on rigorous, context-aware evaluation that reflects its potential to augment clinical reasoning and improve patient outcomes. 𝐵𝐿𝐸𝑈 = 𝐵𝑅 ∙ exp(∑𝑁 𝑛=1 𝑤𝑛 log 𝑝𝑛 ), 𝑅𝑂𝑈𝐺𝐸 − 𝐿 = 𝐹𝛽 = (1+𝛽 2 ) ∙ 𝑅𝐿𝐶𝑆 ∙ 𝑃𝐿𝐶𝑆 𝑅𝐿𝐶𝑆 + 𝛽2 ∙ 𝑃𝐿𝐶𝑆 , 5. COMPARATIVE ANALYSIS OF STATE-OF-THE-ART MEDICAL VLMS Over the past few years, a number of medical VLMs have pushed the boundaries of what is possible in tasks such as medical report generation, VQA, and diagnosis. Models such as RadFM, Med-Flamingo, LLaVA-Med, MedVInT, and more recent ones like LoGra-Med, Med3DVLM, and 81

LLaDA-MedV each contribute with different design choices in architecture, training data, and downstream task performance. In this section we compare these leading models along several key axes: visual and textual backbones, pretraining or instruction tuning strategies, performance on standard medical benchmarks, and trade-offs in generalization, interpretability, and computation. Models like RadFM furnish foundations for multimodal medical reasoning via large diverse datasets. For example, RadFM (Wu et al., 2023) is evaluated across nine public benchmarks including VQA-RAD, SLAKE, PMC-VQA, and others, for tasks spanning medical diagnosis, report generation, and VQA. In many settings, RadFM achieves strong zero-shot performance, leveraging both contrastive training and generative instruction tuning. It performs competitively or better than earlier VLMs such as MedVInT or Med-Flamingo. For instance, in the benchmarking study “Towards Generalist Foundation Model for Radiology”, RadFM outperforms certain baseline foundation models on multiple datasets in diagnosis accuracy and report generation. However, RadFM’s strengths are tempered by limitations: it can require large amounts of instruction data, and some of its generative outputs show occasional hallucinations or decreased factual correctness in rare pathologies. Moreover, its architectural complexity and parameter count make deployment in resource-constrained clinical settings more challenging. Med-Flamingo and LLaVA-Med exemplify models with instruction tuning and generative capabilities. Med-Flamingo extends Flamingo-style multimodal instruction following into the medical domain. It shows improved performance on medical VQA tasks via few-shot prompting. But its performance under zero-shot settings is generally lower than some of its successors. LLaVA-Med, in contrast, emphasizes both few-shot and zero-shot capability; it is more robust when adapting to unseen datasets or questions. In the comparative studies, LLaVA-Med typically outperforms MedFlamingo on averaged metrics across VQA-RAD, SLAKE, PathVQA, etc. These generative instruction-tuned models do well with natural language output and richer responses, but this richness sometimes comes at a cost in precision: false positives of findings, or verbosity that includes less relevant information when compared to concise clinical reports. Recent models attempt to address the trade-off between data efficiency, generalization, and alignment across modalities. LoGra-Med is notable for its multi-graph alignment approach: it enforces consistency between image modalities, conversation-style descriptions, and extended captions. LoGra-Med shows that with far fewer instruction examples (e.g. 10% of the full set), it can approach or exceed the performance of LLaVA-Med on certain VQA benchmarks, especially VQARAD. Also, Med3DVLM is an important advance in extending VLMs into 3D medical imaging: rather than being limited to 2D images, Med3DVLM handles volumetric data (e.g. 3D scans) with specialized encoder architectures (e.g., decomposed 3D convolutions) and maintains strong performance in retrieval, report generation, and VQA tasks. This broadens the applicability to radiology areas beyond chest X-rays. Meanwhile LLaDA-MedV, a diffusion-based instruction-tuned model, pushes forward in controlling response length and improving VQA accuracy across medical benchmarks (VQA-RAD, SLAKE, PathVQA) with new architectural or sampling innovations as shown in Table 1. When comparing these models, several patterns and trade-offs emerge. Models that rely heavily on LLaVA-Med, Med-Flamingo, etc. tend to achieve more fluent and contextually rich outputs but risk introducing errors or irrelevant content. Models emphasizing contrastive or alignment-based pretraining (RadFM, LoGra-Med) are better at ensuring consistency and accuracy, especially in zero82

or few-shot settings, but may produce less expressive text or fewer diagnostics details. 3D models like Med3DVLM show promise in fields where volumetric imaging is essential, yet bring computational and storage burdens. Additionally, there is a trade-off between model size / parameter count and practical deployment: large models are resource-intensive and difficult to integrate into clinic workflows or low-resource settings. Interpretability remains a common concern: while many models support attention visualization or grounding, few have rigorous evaluations of clinical explainability or user trust. Table 1: Summary of State-of-the-Art Medical Vision–Language Models Training Key Performan Architectu Main Main Model Data / Benchmar ce re Type Strengths Limitations Domain ks Highlights 9 multimod Achieves Strong al datasets state-of-the- multimoda High Encoder– (VQAReport art zero-shot l computationa Decoder RadFM RAD, generation, performanc reasoning; l cost; (CLIP + (2023) SLAKE, VQA e on scalable to occasional LLM) PMCmultiple unseen hallucination VQA, tasks data etc.) Large Strong Weak zeroDecoderradiology High fewgenerative VQAshot Medonly datasets shot reasoning; RAD, generalizatio Flamingo (Flamingo- (MIMICperformanc good PathVQA n; verbose (2023) based) CXR, e on VQA instruction outputs ROCO) following Robust Outperform MIMICmultimoda Encoder– VQAs MedProne to CXR, l Decoder RAD, Flamingo overgeneratio LLaVAPMCalignment; (LLaVA PathVQA, on average n; less factual Med (2024) VQA, versatile adaptation) SLAKE across precision SLAKE downstrea benchmarks m use DualEfficient Limited Encoder contrastive Solid generative (Vision Medical learning; ROCO, retrieval and ability; taskMedVInT Transforme image– interpretab VQA-RAD classificatio specific (2023) r + Text text pairs le n results tuning Transforme embedding needed r) s MultiVQAMatches or DataStill 2DGraphLoGrainstitution RAD, exceeds efficient limited; Med (2024) enhanced al SLAKE LLaVAalignment; modest 83

multimodal encoder Med3DVL M (2024) LLaDAMedV (2025) 3D Encoder– Decoder Diffusionbased Encoder– Decoder radiology datasets Volumetri c CT and MRI datasets Instructio n-tuned multimodal data Med using 10% data VQARAD-3D, LIDC-IDRI VQARAD, SLAKE, PathVQA First strong 3D-VLM results with volumetric understandi ng Improved accuracy and controlled response length interpretab generalizatio le crossn to unseen modal modalities mapping Extends VLMs to High training 3D cost; limited modalities; pretraining rich spatial resources awareness Enhanced Requires factual large precision; compute; balanced limited generation interpretabilit y validation 6. TECHNICAL, ETHICAL, AND PRACTICAL CHALLENGES Despite significant progress in the development of VLMs for medical applications, several technical challenges remain unresolved. A major issue lies in data scarcity and heterogeneity. Most publicly available medical datasets—such as MIMIC-CXR, SLAKE, or VQA-RAD—are limited to specific modalities, populations, or institutions, which restricts model generalization. Clinical imaging data are often protected by strict privacy regulations (e.g., HIPAA, GDPR), limiting access for large-scale pretraining. Moreover, the diversity of imaging devices, protocols, and annotation styles introduces substantial domain shift, making it difficult for models trained in one context to perform robustly in another. Another pressing challenge is the interpretability and explainability of VLMs. Unlike traditional rule-based systems, multimodal transformers are largely black-box architectures that make it difficult for clinicians to understand the reasoning behind a model’s prediction or generated report. Although attention maps and gradient-based saliency methods provide partial insights, these techniques often lack causal validity and are insufficient for clinical decisionmaking, where transparency is critical. Additionally, there are computational constraints in deploying these large models in hospitals and clinics. Many state-of-the-art models, such as LLaVA-Med and Med-Flamingo, require substantial GPU memory and processing power, which is often unavailable in real-time diagnostic environments. Techniques like knowledge distillation, quantization, and parameter-efficient fine-tuning (PEFT) can help mitigate these issues but require further refinement to maintain diagnostic accuracy and interpretability. Lastly, multimodal alignment itself remains technically challenging—ensuring that visual and textual features correspond meaningfully across diverse modalities is an open problem that affects both the accuracy and safety of model outputs. Beyond technical obstacles, the ethical challenges surrounding medical VLMs are increasingly critical as these systems approach clinical deployment. The first ethical concern pertains to data privacy and informed consent. Large-scale multimodal datasets frequently include sensitive personal information embedded in images, metadata, or textual notes. Improper anonymization or data leakage can lead to severe privacy violations. Ensuring that patients are informed about how their data will 84

be used for model training is often neglected in large-scale data collection. A second ethical challenge involves bias and fairness. VLMs trained on unbalanced datasets risk reproducing or amplifying existing health disparities—for instance, producing lower diagnostic accuracy for underrepresented populations such as certain age groups, ethnicities, or rare pathologies. These biases can have profound real-world consequences if the models are integrated into diagnostic systems or automated report generation pipelines. Another issue relates to hallucination and factual errors in generated reports. Even when models achieve high quantitative scores on benchmarks, they may produce confident yet clinically incorrect statements. This phenomenon, often termed “medical hallucination,” can mislead healthcare providers or patients if the generated outputs are trusted without human oversight. Ethical AI design in this context requires establishing frameworks for accountability, transparency, and human-in-the-loop verification, ensuring that clinicians remain the ultimate decision-makers. Additionally, intellectual property and authorship of AI-generated content raise new legal and ethical dilemmas—particularly when models trained on copyrighted or proprietary datasets generate outputs that resemble real medical records or reports. In addition to technical and ethical concerns, practical implementation challenges present another major barrier to the widespread adoption of medical VLMs. Integrating these models into existing healthcare systems requires careful consideration of regulatory compliance, workflow adaptation, and user training. Health institutions must adhere to medical device regulations such as those from the FDA (U.S.), EMA (Europe), or local regulatory bodies. However, current guidelines for the approval of AI-driven diagnostic tools rarely account for complex multimodal architectures that learn dynamically from new data. This creates uncertainty in how VLM-based systems can be certified as safe and effective. Furthermore, the integration of AI tools into clinical workflows must be seamless to avoid increasing the cognitive load of healthcare providers. Poorly designed user interfaces or non-intuitive feedback mechanisms can lead to underuse or misuse of the technology. There is also the issue of infrastructure inequality: many hospitals, especially in developing regions, lack the hardware or network capacity to deploy large-scale VLMs. Cloud-based solutions offer potential scalability but raise additional security and latency concerns. To promote trust and usability, systems must include robust mechanisms for model monitoring, continual learning, and performance auditing under real-world conditions. Another practical issue is the training of medical staff to effectively interpret AI outputs. Clinicians need not only technical literacy to operate AI systems but also critical understanding to question or override model predictions when necessary. Finally, the cost-benefit trade-off must be justified: training and maintaining large VLMs is expensive, and the measurable clinical benefit must outweigh these costs for healthcare institutions to adopt them sustainably. Therefore, addressing the interplay between technical, ethical, and practical challenges is crucial for the responsible evolution of medical Vision–Language Models. Overcoming these challenges requires a holistic approach that combines advancements in machine learning theory, ethical AI governance, and human-centered design. Future research should focus on developing dataefficient and interpretable VLMs that respect patient privacy while maintaining high diagnostic reliability. Establishing international standards for auditing medical AI models—similar to clinical trial protocols—could further ensure transparency and accountability. Collaboration between AI researchers, clinicians, ethicists, and policymakers is essential to build systems that are not only 85

accurate and powerful but also trustworthy, equitable, and sustainable in real-world healthcare environments. 7. FUTURE RESEARCH DIRECTIONS TOWARD RELIABLE AND INTERPRETABLE MEDICAL AI The next generation of VLMs for medicine must prioritize reliability, interpretability, and trustworthiness to enable safe clinical deployment. Current systems exhibit impressive multimodal reasoning but remain vulnerable to domain shift, factual inconsistencies, and unexplained errors. One promising direction is the development of domain-adaptive pretraining and self-supervised learning strategies tailored to medical data. Unlike general vision–language corpora, medical datasets are highly specialized and structured. Leveraging modality-specific pretraining objectives—such as anatomy-aware contrastive learning, pathology localization, or cross-report contextualization—could improve domain robustness and generalization. Another critical area is knowledge-grounded modeling, where structured medical ontologies (e.g., UMLS, SNOMED CT, RadLex) are incorporated into model architectures or training pipelines. This can anchor model predictions in verifiable medical knowledge and reduce hallucinations. In parallel, parameter-efficient fine-tuning and federated learning will be essential for scalability and data privacy. Federated frameworks allow hospitals to collaboratively train models on decentralized datasets without exchanging patient data, mitigating privacy risks while improving diversity. These methods, combined with continual learning techniques, can help maintain model relevance as new imaging modalities and clinical standards evolve. Research into multimodal uncertainty quantification is also gaining importance, enabling models to express confidence in their outputs—an indispensable feature for clinical decision support systems. Equally vital for the future is the emphasis on interpretability and human-centered design. The success of medical VLMs depends not only on algorithmic performance but also on how effectively they integrate into clinical practice. Future models must offer transparent reasoning pathways that clinicians can inspect and verify. Techniques such as attention-based grounding, counterfactual explanation, causal interpretability, and multimodal attention visualization can enhance user understanding of model decisions. Interdisciplinary research that bridges AI with cognitive science and radiology could lead to new interpretability frameworks reflecting how experts process visual and textual information. Another key direction involves human-in-the-loop learning, where medical experts iteratively correct or refine AI-generated reports, leading to co-adaptive systems that improve over time while maintaining accountability. Furthermore, standardized benchmarking protocols and explainability evaluation metrics are urgently needed to ensure consistent comparison across models. Current benchmarks (e.g., VQA-RAD, SLAKE) primarily measure task accuracy but fail to assess factual correctness, interpretability, and ethical alignment. Future initiatives should design holistic evaluation ecosystems that capture technical, clinical, and social dimensions of AI performance. Finally, building interdisciplinary collaborations among AI scientists, clinicians, ethicists, and policymakers will be pivotal to translating research into clinically validated and ethically compliant systems. As these research directions mature, they will pave the way for medical AI that is not only powerful and efficient but also transparent, accountable, and trustworthy, marking a transformative step toward responsible and interpretable healthcare innovation. 86

8. CONCLUSIONS This paper has provided a comprehensive review of recent advancements in medical VLMs, highlighting their architectures, datasets, evaluation metrics, and clinical applications in report generation and VQA. The integration of vision and language modalities has opened transformative opportunities for automated diagnostic reasoning, clinical documentation, and decision support. However, despite notable progress, current models still face significant limitations related to domain adaptation, interpretability, ethical compliance, and data diversity. Through comparative analysis, this study underscored the need for domain-specific training strategies, multimodal alignment mechanisms, and transparent evaluation protocols to ensure reliability and clinical trust. Looking ahead, the development of reliable and interpretable VLMs for healthcare will require a multidisciplinary effort that bridges AI research with clinical expertise and ethical governance. Future models must prioritize explainability, robust generalization across modalities, and dataefficient learning while adhering to patient privacy and regulatory frameworks. Standardized datasets and unified benchmarking platforms should be established to foster fair and reproducible evaluation. Furthermore, incorporating human-in-the-loop frameworks and knowledge-grounded reasoning can enhance both accuracy and clinical interpretability, ensuring that AI systems act as collaborative tools rather than opaque decision-makers. In conclusion, the convergence of vision and language in medical AI represents a paradigm shift in how machines understand and assist in healthcare diagnostics. Continued innovation, guided by ethical responsibility and clinical validation, will be essential to transform current VLMs from research prototypes into trusted, explainable, and patient-centered technologies. This work thus contributes a roadmap for advancing multimodal intelligence toward safer, transparent, and more effective medical applications. REFERENCES: [1] Zhou, D.-W., Zhang, Y., Ning, J., Ye, H.-J., Zhan, D.-C., and Liu, Z. 2023. Learning without forgetting for vision-language models. arXiv Preprint arXiv:2305.19270. doi: 10.48550/arXiv.2305.192705. [2] Waqas, A., Naveed, J., Shahnawaz, W., Asghar, S., Bui, M. M., and Rasool, G. 2024. Digital pathology and multimodal learning on oncology data. Artif. Intell. 1, 1–15. doi: 10.1093/bjrai/ubae014. [3] Chen, F., Zhang, D., Han, M., Chen, X., Shi, J., Xu, S., et al. 2023. VLP: a survey on vision-language pre-training. Machine Intell. Res. 20, 38–56. doi: 10.1007/s11633-022-1369-5 [4] Liu, H., Xue, W., Chen, Y., Chen, D., Zhao, X., Wang, K., et al. 2024. A survey on hallucination in large vision-language models. arXiv Preprint arXiv:2402.00253. doi: 10.48550/arXiv.2402.00253F [5] Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., et al. 2021. Learning transferable visual models from natural language supervision. arXiv Preprint arXiv:2103.00020. doi: 10.48550/arXiv.2103.00020 [6] Jia, C., Yang, Y., Xia, Y., Chen, Y., Parekh, Z., Pham, H., et al. (2021). “Scaling up visual and vision-language representation learning with noisy text supervision,” in International Conference on Machine Learning, Vol. 139, 4904–4916. [7] Moon, J. H., Lee, H., Shin, W., Kim, Y.-H., and Choi, E. 2022. Multimodal understanding and generation for medical images and text via vision-language pre-training. IEEE J. Biomed. Health 87

Informat. 26, 6070–6080. doi: 10.1109/JBHI.2022.3207502 [8] Lau, J. J., Gayen, S., Ben Abacha, A., and Demner-Fushman, D. 2018. A dataset of clinically generated visual questions and answers about radiology images. Sci. Data 5:180251. doi: 10.1038/sdata.2018.251 [9] Bannur, S., Hyland, S., Liu, Q., Perez-Garcia, F., Ilse, M., Castro, D. C., et al. 2023. Learning to exploit temporal structure for biomedical vision-anguage processing. arXiv Preprint arXiv:2301.04558. doi: 10.48550/arXiv.2301.04558 [10] Zhang, S., Xu, Y., Usuyama, N., Bagga, J., Tinn, R., Preston, S., et al. 2023. Largescale domain-specific pretraining for biomedical vision-language processing. arXiv Preprint arXiv:2303.00915. doi: 10.48550/arXiv.2303.00915 [11] Yuan, Z., Jin, Q., Tan, C., Zhao, Z., Yuan, H., Huang, F., et al. 2023 RAMM: retrievalaugmented biomedical visual question answering with multimodal pre-training. arXiv Preprint arXiv:2303.00534. doi: 10.48550/arXiv.2303. 00534 [12] Shu, C., Chen, B., Liu, F., Fu, Z., Shareghi, E., and Collier, N. 2023. Visual MEDALPACA: a Parameter-Efficient Biomedical LLM With Visual Capabilities. Available at: https://cambridgeltl.github.io/visual-med-alpaca/ (accessed February 20, 2024) [13] Ranjit, M., Ganapathy, G., Manuel, R., and Ganu, T. 2023. Retrieval augmented chest Xray report generation using openAI GPT models. arXiv Preprint arXiv:2305.03660. doi: 10.48550/arXiv.2305.03660 [14] Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., et al. 2023. LLaVA-Med: training a large language-and-vision assistant for biomedicine in one day. arXiv Preprint arXiv:2306.00890. doi: 10.48550/arXiv.230600890 [15] Pellegrini, C., Özsoy, E., Busam, B., Navab, N., and Keicher, M. 2023. Radialog: a large vision-language model for radiology report generation and conversational assistance. arXiv Preprint arXiv:2311.18681. doi: 10.48550/arXiv.2311.18681 [16] Johnson, A. E. W., Pollard, T. J., Greenbaum, N. R., Lungren, M. P., Ying Deng, C., Peng, Y., et al. 2019. MIMIC-CXR-JPG, a large publicly available database of labeled chest radiographs. arXiv Preprint arXiv:1901.07042. doi: 10.48550/arXiv.1901.07042 [17] Irvin, J. A., Rajpurkar, P., Ko, M., Yu, Y., Ciurea-Ilcus, S., Chute, C., et al. 2019. CheXpert: a large chest radiograph dataset with uncertainty labels and expert comparison. AAAI Conf. Artif. Intell. 33, 590–597. doi: 10.1609/aaai.v33i01.3301590 [18] Demner-Fushman, D., Kohli, M. D., Rosenman, M. B., Shooshan, S. E., Rodriguez, L. M., Antani, S. K., et al. 2015. Preparing a collection of radiology examinations for distribution and retrieval. J. Am. Med. Informat. Assoc. 23, 304–310. doi: 10.1093/jamia/ocv080 [19] He, X., Zhang, Y., Mou, L., Xing, E., and Xie, P. 2020. PathVQA: 30000+ questions for medical visual question answering. arXiv Preprint arXiv:2003.10286. doi: 10.48550/arXiv.2003.10286 [20] Li, P., Liu, G., He, J., Zhao, Z., and Zhong, S. 2023. “Masked vision and language pretraining with unimodal and multimodal contrastive losses for medical visual question answering,” in Medical Image Computing and Computer Assisted Intervention (MICCAI), 374–383. [21] Lu, M. Y., Chen, B., Williamson, D. F. K., Chen, R. J., Zhao, M., Chow, A. K., et al. 2024. A multimodal generative ai copilot for human pathology. Nature 24:3. doi: 10.1038/s41586024-07618-3 88

SUN’IY INTELLEKT ETIKASI BO‘YICHA XALQARO NAZARIY ME’YORLAR: TAMOYILLAR, MUAMMOLAR VA TA’LIM UCHUN TAVSIYALAR Kobuljon Ismanov1, Alisher Mamatov2, Said Maxanov3 O‘zbekiston respublikasi ta'limni rivojlantirish respublika ilmiy-metodik markazi ismanovqa@gmail.com Annotatsiya. Mazkur maqolada sun’iy intellekt etikasi bo‘yicha yetakchi xalqaro tashkilotlar tomonidan ishlab chiqilgan nazariy me’yorlar, jumladan UNESCO, OECD va Yevropa Ittifoqining “ishonchli sun’iy intellekt” konsepsiyalari tizimli ravishda tahlil qilingan. Tadqiqot davomida sun’iy intellekt texnologiyalarini joriy etish jarayonida yuzaga kelayotgan asosiy axloqiy muammolar — algoritmik diskriminatsiya, shaxsiy ma’lumotlar xavfsizligi, javobgarlik masalalari, shaffoflik va inson nazorati muammolari ilmiy asosda yoritilgan. Shuningdek, ta’lim tizimida sun’iy intellektdan foydalanishning ijobiy imkoniyatlari bilan bir qatorda uning etik xavflari ochib berilgan. Xalqaro tavsiyalar asosida ta’limda sun’iy intellektdan mas’uliyatli foydalanish uchun zarur bo‘lgan tamoyillar aniqlangan hamda O‘zbekiston ta’lim tizimi uchun moslashtirilgan konseptual yondashuv taklif etilgan. Mazkur tadqiqot natijalari ta’lim jarayoniga sun’iy intellektni joriy etishda etik, huquqiy va pedagogik muvozanatni ta’minlashga xizmat qiladi. Kalit so‘zlar: sun’iy intellekt, sun’iy intellekt etikasi, raqamli etika, xalqaro me’yorlar, mas’uliyatli sun’iy intellekt, algoritmik diskriminatsiya, ta’limda sun’iy intellekt, shaxsiy ma’lumotlar xavfsizligi. 1. Kirish. So‘nggi yillarda sun’iy intellekt (SI) texnologiyalarining jadal rivojlanishi jamiyat hayotining deyarli barcha sohalariga, jumladan ta’lim, sog‘liqni saqlash, iqtisodiyot, davlat boshqaruvi, xavfsizlik va ijtimoiy xizmatlar tizimiga chuqur kirib borishiga sabab bo‘lmoqda. Katta hajmdagi ma’lumotlarni avtomatik qayta ishlash, tahlil qilish, bashoratlash va mustaqil qarorlar qabul qilish imkoniyatiga ega bo‘lgan SI tizimlari inson faoliyatining samaradorligini oshirish bilan birga, tubdan yangi ijtimoiy, huquqiy va axloqiy muammolarni ham yuzaga chiqarmoqda. Ayniqsa, SI tizimlarining jamiyatga ko‘rsatadigan ta’siri ularning nafaqat texnik imkoniyatlari, balki axloqiy asoslanganligi bilan ham bevosita bog‘liq ekanligi tobora yaqqol namoyon bo‘lmoqda. Sun’iy intellektning joriy etilishi bilan bog‘liq asosiy axloqiy muammolar qatoriga algoritmik diskriminatsiya, shaxsiy ma’lumotlar maxfiyligining buzilishi, avtomatlashtirilgan qarorlar uchun javobgarlikning noaniqligi, inson nazoratining cheklanishi, shaffoflikning yetarli darajada ta’minlanmasligi hamda raqamli tengsizlikning kuchayishi kabi holatlar kiradi. Ushbu muammolar ayniqsa ta’lim tizimida keskin namoyon bo‘lib, o‘quvchilarning baholanishi, individual ta’lim yo‘llarining shakllanishi, pedagogik adolat, akademik halollik va shaxsiy ma’lumotlarning himoyasi bilan bevosita bog‘liqdir. Shu sababli bugungi kunda sun’iy intellektni ta’lim jarayoniga joriy etishda nafaqat texnologik samaradorlik, balki etik muvofiqlik ham muhim mezon sifatida qaralmoqda. Mazkur jarayonlar global miqyosda ham keng muhokamaga sabab bo‘lib, sun’iy intellekt etikasi bo‘yicha xalqaro normativ-huquqiy yondashuvlar shakllanmoqda. Jumladan, UNESCO tomonidan 2021-yilda qabul qilingan “Sun’iy intellekt etikasi bo‘yicha tavsiya” hujjati, OECDning 89

“ishonchli sun’iy intellekt” tamoyillari hamda Yevropa Ittifoqining “Trustworthy AI” konsepsiyasi sun’iy intellektni inson huquqlari, adolat, shaffoflik, xavfsizlik va javobgarlik asosida tartibga solishga qaratilgan yetakchi xalqaro nazariy me’yorlar sifatida e’tirof etilmoqda. Ushbu hujjatlar sun’iy intellektni rivojlantirishda “Inson manfaatlariga yo‘naltirilgan sun’iy intellekt” g‘oyasini asosiy ustuvor yo‘nalish sifatida belgilab bermoqda. O‘zbekiston Respublikasida ham raqamli transformatsiya jarayonlari, jumladan, ta’lim tizimini raqamlashtirish, sun’iy intellekt texnologiyalarini joriy etish bo‘yicha keng ko‘lamli islohotlar amalga oshirilmoqda. Oʻzbekiston Respublikasi Prezidenti Sh.Mirziyoyevning 2020 – yil 5– oktyabrdagi PF – 6079–son farmoni bilan “Raqamli Oʻzbekiston – 2030” strategiyasi tasdiqlangan ushbu strategiya doirasida 2024–yil 14–oktyabrda “Sunʼiy intellekt texnologiyalarini 2030 – yilga qadar rivojlantirish strategiyasini tasdiqlash toʻgʻrisida” PQ – 358-sonli qaror qabul qilingan, bunda, maktabgacha, umumiy o‘rta va oliy ta’lim tizimida raqamli platformalarning jadal joriy etilishi ta’lim jarayoniga sun’iy intellekt elementlarini bosqichma-bosqich tatbiq etishni hamda xavfsizlik nuqtai nazaridan sunʼiy intellekt texnologiyalari boʻyicha etika qoidalarini ishlab chiqishni taqozo etmoqda. Yuqoridagilarni inobatga olib sun’iy intellektdan mas’uliyatli foydalanish, etik xavflarni oldindan baholash, adolatni ta’minlash va shaxsiy ma’lumotlarni himoyalash masalalari yetarli darajada ilmiy asoslangan mexanizmlarni talab etmoqda. Shu nuqtai nazardan, sun’iy intellekt etikasi bo‘yicha xalqaro nazariy me’yorlarni tizimli tahlil qilish, ularning umumiy tamoyillarini aniqlash, amaliyotda yuzaga kelayotgan muammolarni ilmiy baholash hamda ta’lim tizimi uchun moslashtirilgan tavsiyalar ishlab chiqish dolzarb ilmiy vazifa hisoblanadi. Ayniqsa, xalqaro tajribaga tayangan holda O‘zbekiston ta’lim tizimida sun’iy intellektdan mas’uliyatli va xavfsiz foydalanish modelini shakllantirish ilmiy-amaliy jihatdan muhim ahamiyat kasb etadi. Mazkur maqolaning maqsadi sun’iy intellekt etikasi bo‘yicha yetakchi xalqaro tashkilotlar tomonidan ishlab chiqilgan nazariy me’yorlarni tahlil qilish, asosiy tamoyillar va muammolarni aniqlash hamda ta’lim tizimi uchun ilmiy asoslangan tavsiyalar ishlab chiqishdan iborat. Ushbu maqsadga erishish uchun quyidagi vazifalar belgilandi: - sun’iy intellekt etikasi tushunchasining nazariy mohiyatini ochib berish; - xalqaro me’yoriy hujjatlar asosida asosiy etik tamoyillarni tahlil qilish; - sun’iy intellekt bilan bog‘liq dolzarb axloqiy muammolarni tasniflash; - ta’lim tizimida sun’iy intellektdan foydalanishda yuzaga keladigan etik muammolarni aniqlash; - ta’lim uchun mas’uliyatli sun’iy intellektdan (responsible AI) foydalanishga oid tavsiyalar ishlab chiqish. Mazkur tadqiqot natijalari sun’iy intellektni ta’lim jarayoniga joriy etishda etik, huquqiy va pedagogik muvozanatni ta’minlash, shuningdek, milliy ta’lim tizimida sun’iy intellektdan barqaror va mas’uliyatli foydalanish bo‘yicha konseptual yondashuvlarni shakllantirishga xizmat qiladi. 1. Sun’iy intellekt etikasi tushunchasining nazariy asoslari Sun’iy intellekt (SI) tushunchasi zamonaviy ilm-fan va texnologiya rivojining eng muhim fenomenlaridan biri hisoblanadi. Ilmiy adabiyotlarda sun’iy intellekt odatda inson aqliy faoliyatiga xos bo‘lgan tahlil qilish, o‘rganish, mantiqiy xulosa chiqarish, bashoratlash va qaror qabul qilish funksiyalarini amalga oshira oladigan algoritmik va dasturiy tizimlar majmui sifatida talqin etiladi. 90

Biroq texnologik taraqqiyot natijasida sun’iy intellektning jamiyat hayotiga chuqur integratsiyalashuvi uni faqat texnik kategoriya sifatida emas, balki ijtimoiy-falsafiy va axloqiy hodisa sifatida ham ilmiy jihatdan tadqiq etishni taqozo etmoqda. Shu nuqtai nazardan, sun’iy intellekt etikasi tushunchasi raqamli etika va texnologik etika yo‘nalishlarining mantiqiy davomchisi sifatida shakllandi. Umumiy etika inson xulq-atvori, qadriyatlari va axloqiy me’yorlarini tartibga solsa, raqamli etika zamonaviy axborot texnologiyalari, internet muhiti va raqamli xizmatlarning jamiyatga ta’sirini o‘rganadi. Sun’iy intellekt etikasi esa aynan avtomatlashtirilgan aqlli tizimlar tomonidan qabul qilinayotgan qarorlarning inson hayoti, huquqlari, erkinliklari va jamiyat barqarorligiga ta’siri bilan bog‘liq axloqiy muammolarni ilmiy jihatdan tahlil etadi. Sun’iy intellekt etikasi nazariyasining markazida “inson manfaatlari” tamoyili turadi. Ushbu yondashuvga ko‘ra, har qanday sun’iy intellekt tizimi inson manfaatlariga xizmat qilishi, uning huquqlarini cheklamasligi, jamiyatning ijtimoiy adolat, tenglik va xavfsizlik tamoyillariga zid kelmasligi lozim. Bu g‘oya yetakchi xalqaro tashkilotlar, xususan UNESCO, OECD hamda Yevropa Ittifoqining rasmiy yondashuvlarida asosiy ustuvor yo‘nalish sifatida mustahkamlangan. Nazariy jihatdan sun’iy intellekt etikasi quyidagi asosiy kategoriyalar orqali tavsiflanadi: Birinchidan, inson huquqlari ustuvorligi. SI tizimlari shaxsning daxlsizligi, shaxsiy hayoti, fikrlash erkinligi va kamsitilmaslik huquqlarini buzmasligi kerak. Algoritmik qarorlar inson huquqlariga tahdid soluvchi omilga aylanmasligi zarur. Ikkinchidan, adolatlilik. Ma’lumotlarda mavjud bo‘lgan tarixiy yoki ijtimoiy noaniqliklar sun’iy intellekt algoritmlari orqali me’yordan oshmasligi zarur. Algoritmik diskriminatsiya zamonaviy SI etikasi nazariyasining eng muhim muammolaridan biri sifatida e’tirof etiladi. Uchinchidan, shaffoflik va tushuntiriluvchanlik. Sun’iy intellekt tomonidan qabul qilinayotgan qarorlarning mantiqiy asoslari foydalanuvchi uchun tushunarli bo‘lishi lozim. To‘rtinchidan, javobgarlik. Agar SI noto‘g‘ri qaror qabul qilsa, uning oqibatlari uchun mas’ul subyekt aniq belgilangan bo‘lishi kerak. Javobgarlikning noaniqligi sun’iy intellektni keng miqyosda tatbiq etishda asosiy to‘siqlardan biri hisoblanadi. Beshinchidan, xavfsizlik va barqarorlik. SI tizimlari texnik nosozlik, kiberxurujlar va noxolis foydalanishdan himoyalangan bo‘lishi zarur. Aks holda, bunday tizimlar ijtimoiy xavfsizlikka jiddiy tahdid tug‘dirishi mumkin. Shu bilan birga, sun’iy intellekt etikasi nazariyasi normativ va deskriptiv yo‘nalishlarni ham o‘z ichiga oladi. Normativ etika “qanday bo‘lishi kerak?” degan savolga javob izlab, sun’iy intellekt uchun axloqiy me’yorlarni belgilaydi. Deskriptiv yondashuv esa “hozirda qanday?” degan savol asosida jamiyatda real mavjud bo‘lgan sun’iy intellektdan foydalanuvchi inson xulq-atvorini tahlil qiladi. Bu ikki yo‘nalish uyg‘unlashgandagina sun’iy intellekt etikasi samarali nazariy poydevorga ega bo‘la oladi. Ta’kidlash joizki, sun’iy intellekt etikasi faqat texnologiya ishlab chiquvchilarga emas, balki davlat boshqaruvi, ta’lim, sog‘liqni saqlash, huquqni muhofaza qilish tizimlari, shuningdek keng jamoatchilik uchun ham muhim metodologik yo‘riqnoma vazifasini bajaradi. Ayniqsa, ta’lim tizimida sun’iy intellektdan foydalanishda etik me’yorlar o‘quvchilarning shaxsiy ma’lumotlari 91

himoyasi, baholashning adolatliligi, pedagogik halollik va teng imkoniyatlarni ta’minlash bilan uzviy bog‘liqdir. 2. Sun’iy intellekt etikasi bo‘yicha xalqaro nazariy me’yorlar. Sun’iy intellektning jadal rivojlanishi va global miqyosda keng joriy etilishi natijasida uning inson huquqlari, xavfsizlik, ijtimoiy adolat va barqaror rivojlanishga ta’siri masalalari alohida e’tibor markaziga chiqdi. Shu sababli so‘nggi yillarda yetakchi xalqaro tashkilotlar tomonidan sun’iy intellekt etikasi bo‘yicha umumiy konseptual yondashuvlar, universal tamoyillar va normativ tavsiyalar ishlab chiqildi. Ushbu bo‘limda mazkur yo‘nalishda asos bo‘lib xizmat qilayotgan UNESCO, OECD va Yevropa Ittifoqining “Trustworthy AI” konsepsiyalari hamda IEEE tashabbuslari tizimli tahlil qilinadi. 2.1. UNESCO yondashuvi: sun’iy intellekt etikasi bo‘yicha global universal standart 2021-yilda UNESCO tomonidan qabul qilingan “Recommendation on the Ethics of Artificial Intelligence” hujjati sun’iy intellekt etikasi bo‘yicha dunyodagi birinchi universal xalqaro normativ hujjat sifatida e’tirof etiladi. Ushbu hujjat 190 dan ortiq davlat tomonidan qo‘llab-quvvatlangan bo‘lib, unda sun’iy intellektni rivojlantirish va joriy etishda asosiy axloqiy, ijtimoiy va huquqiy yo‘nalishlar belgilab berilgan. UNESCOning sun’iy intellekt etikasi bo‘yicha konsepsiyasi o‘z markazida inson qadr-qimmati va manfaatlarini himoya qilish tamoyilini qo‘yadi. Ya’ni har qanday texnologik yechim yoki avtomatlashtirilgan qaror insonning huquqlari, erkinliklari va farovonligi ustuvor bo‘lgan holda qo‘llanilishi lozim. Hujjatda SI etikasi bir qator asosiy tamoyillar asosida belgilangan, bular inson huquqlarining ustuvorligi, inklyuzivlik va tenglikni ta’minlash, adolat va kamsitmaslik, shaffoflik va izoh berish imkoniyati, texnologik xavfsizlik va barqarorlik, ma’lumotlar maxfiyligi hamda javobgarlik va ijtimoiy mas’uliyat. Bu yondashuvning eng muhim jihati shundaki, sun’iy intellekt etikasi faqat texnik masala sifatida emas, balki jamiyat hayotining barcha qatlamlariga ya’ni madaniyat, ta’lim, ijtimoiy munosabatlar va siyosiy boshqaruvga bevosita ta’sir etuvchi omil sifatida talqin qilinadi. Ayniqsa ta’lim sohasida SI texnologiyalarini joriy etishda pedagogning hal qiluvchi roli, o‘quvchilarning manfaatlari, teng imkoniyatlar yaratish va shaxsiy ma’lumotlarning ishonchli himoyasi alohida ustuvor yo‘nalish sifatida ko‘rsatib o‘tiladi. Ilmiy va amaliy nuqtayi nazardan UNESCO konsepsiyasi global miqyosda yagona me’yoriy asos bo‘lib xizmat qiladi. Mazkur hujjat har bir davlat, jumladan O‘zbekiston uchun ham SI texnologiyalarini etik tamoyillar asosida joriy etish, ularni boshqarish va ta’lim tizimiga uyg‘un integratsiya qilishda muhim metodologik tayanch vazifasini bajarishi mumkin. 2.2 OECD yondashuvi: ishonchli sun’iy intellekt (Trustworthy AI) konsepsiyasi OECD tomonidan 2019-yilda qabul qilingan va 2024-yilda yangilangan “AI Principles” hujjati sun’iy intellekt etikasi masalasiga davlatlar va hukumat vakillari uchun mo‘ljallangan konsepsiya sifatida yondashadi. Ushbu hujjatda sun’iy intellektni “ishonchli sun’iy intellekt” tamoyili asosida rivojlantirish g‘oyasi ilgari suriladi. Bu tamoyillar ikki asosiy yo‘nalishga tayanadi: qiymatga asoslangan tamoyillar va siyosiy tavsiyalar. Qiymatga asoslangan tamoyillar qatoriga inson huquqlari va demokratik qadriyatlarni hurmat qilish, adolat, tenglik va inklyuzivlikni ta’minlash, shaffoflik va 92

tizimlarning tushuntiriluvchanligini oshirish, texnologik xavfsizlik va barqarorlikni mustahkamlash, shuningdek javobgarlikni aniq belgilash kiradi. OECD yondashuvining muhim jihati shundaki, u sun’iy intellektni texnik muammo sifatida emas, balki davlat boshqaruvi, iqtisodiy rivojlanish va ijtimoiy barqarorlik omili sifatida ko‘radi. Ushbu tamoyillar ko‘plab davlatlarda milliy sun’iy intellekt strategiyalarini ishlab chiqishda metodologik asos bo‘lib xizmat qilmoqda. 2.3. Yevropa Ittifoqi: “Trustworthy AI” yo‘riqnomasi Yevropa Ittifoqi tomonidan ishlab chiqilgan yo‘riqnoma sun’iy intellekt tizimlarining ishonchli, xavfsiz va inson huquqlariga mos ishlashini ta’minlashga qaratilgan kompleks yondashuvni taqdim etadi. Ushbu hujjatga ko‘ra, har qanday sun’iy intellekt tizimi uch asosiy mezonga javob berishi lozim: qonuniylik, axloqiylik va texnik ishonchlilik. Mazkur yondashuv asosida yetti asosiy talab ishlab chiqilgan bo‘lib, ular sun’iy intellektning inson nazorati ostida ishlashi, texnik xavfsizlikni ta’minlashi, ma’lumotlar maxfiyligini himoya qilishi, faoliyati shaffof bo‘lishi, diskriminatsiyaga yo‘l qo‘ymasligi, ijtimoiy va ekologik farovonlikka hissa qo‘shishi hamda aniq javobgarlik mexanizmlariga ega bo‘lishini talab qiladi. Yo‘riqnomada, shuningdek, sun’iy intellekt tizimlarini amaliy jihatdan baholash uchun etik mezonlar, risklarni baholash mexanizmlari va sertifikatlashtirish jarayonlari bo‘yicha metodik ko‘rsatmalar ham berilgan. Bu yondashuv ayniqsa ta’lim, sog‘liqni saqlash va davlat boshqaruvi kabi yuqori ijtimoiy mas’uliyat talab qiladigan sohalarda alohida ahamiyat kasb etadi, chunki ushbu sohalarda sun’iy intellektning ishonchliligi va adolatliligi bevosita inson hayoti va farovonligiga ta’sir ko‘rsatadi. 2.4. IEEE va ilmiy meta-tahlillar. IEEE tomonidan ishlab chiqilgan IEEE 7010-2020 standarti sun’iy intellekt va avtonom tizimlarning inson farovonligiga real ta’sirini baholash metodologiyasini taklif etadi. Bu yondashuv SI ni faqat texnologik emas, balki ijtimoiy indikatorlar orqali baholash imkonini beradi. Shuningdek, 200 dan ortiq xalqaro hujjatlarni qamrab olgan ilmiy meta-tahlillar shuni ko‘rsatadiki, sun’iy intellekt etikasi bo‘yicha 17 ta keng tarqalgan tamoyil mavjud, biroq ularning talqini va amaliy ustuvorligi hujjatdan hujjatga sezilarli darajada farqlanadi. “Adolat”, “shaffoflik”, “maxfiylik” tushunchalari turli kontekstlarda turlicha izohlanadi. 2.5. Xalqaro me’yorlarni qo‘llashdagi muammolar Xalqaro tamoyillar keng konsensus asosida shakllangan bo‘lsa-da, ularni amaliyotga tatbiq etish jarayonida bir qator murakkab muammolar yuzaga kelmoqda: - texnologiyalarning juda tez rivojlanishi sababli etik me’yorlarning tez eskirishi; - davlat, biznes va fuqarolik jamiyati manfaatlarining to‘qnashuvi; - milliy qonunchilik tizimlarining zaifligi; - maxfiylik va javobgarlik bo‘yicha huquqiy bo‘shliqlar. Ilmiy tadqiqotlar shuni ko‘rsatadiki, tamoyillar bo‘yicha umumiy kelishuv mavjud bo‘lsa-da, ularni “qanday” va “kim tomonidan” amalga oshirish masalalari ochiq qolib, natijada global 93

me’yorlardan milliy tartibga solish tizimlariga o‘tish jarayonida asosiy muammolar yuzaga kelmoqda. 2.6. Xalqaro yondashuvlarning taqqoslama tahlili Sun’iy intellekt etikasi bo‘yicha UNESCO, OECD va Yevropa Ittifoqi tomonidan ishlab chiqilgan konsepsiyalar mazmunan turli yo‘nalishlarga urg‘u bersa-da, ularni birlashtiruvchi umumiy yadroviy tamoyillar tizimi mavjud. Ushbu yondashuvlarni taqqoslash orqali tahlil qilish ularning o‘xshash va farqli jihatlarini aniq ilmiy asosda ochib berish imkonini beradi. UNESCO yondashuvi global va universal xarakterga ega bo‘lib, u davlat siyosati, ta’lim, sog‘liqni saqlash, madaniyat va ijtimoiy rivojlanishni qamrab oladi. OECD yondashuvi asosan davlat boshqaruvi, iqtisodiy, siyosiy va milliy strategiyalar uchun mo‘ljallangan. Yevropa Ittifoqi yo‘riqnomasi esa amaliy joriy etish, texnik baholash va sertifikatlash mexanizmlariga yo‘naltirilgan. Demak, UNESCO normativ-axloqiy me’yorlar, OECD siyosiy-strategik me’yorlar, Yevropa Ittifoqi yo‘riqnomasi esa amaliy-texnik me’yor vazifasini bajaradi. Mazkur farqlar ularni bir-birini inkor etuvchi emas, balki bir-birini to‘ldiruvchi global me’yoriy tizim sifatida qarash imkonini beradi. 3. Sun’iy intellekt etikasi bilan bog‘liq asosiy muammolar Sun’iy intellekt texnologiyalarining jadal rivojlanishi va turli sohalarga keng ko‘lamda joriy etilishi ularning samaradorligini oshirish bilan bir qatorda murakkab axloqiy, huquqiy va ijtimoiy muammolarni ham yuzaga keltirmoqda. Xalqaro me’yoriy hujjatlarda belgilangan tamoyillar nazariy jihatdan mukammal bo‘lsa-da, ularni amaliyotga tatbiq etish jarayonida 1-rasmda keltirilgan tizimli muammolar mavjud. 1-rasm. Sun’iy intellektni joriy etishda yuzaga kelayotgan etik va huquqiy muammolar Algoritmik diskriminatsiya bu sun’iy intellekt tizimlari tomonidan qabul qilinayotgan qarorlarning ayrim ijtimoiy guruhlarga nisbatan adolatsiz yoki tarafkashligi bilan bog‘liq muammo bo‘lib, u asosan mutaxassis tomonidan o‘qitilgan ma’lumotlardagi to‘liq bo‘lmagan axborot va model sozlamalaridagi xatolar bilan izohlanadi. Bunday holatlar avtomatlashtirilgan ishga qabul qilish 94

jarayonlari, kredit reytingini baholash, shuningdek, o‘quvchilarning bilim darajasini aniqlash kabi tizimlarda bilvosita xatoliklarga olib kelishi mumkin. Bundan tashqari, sun’iy intellekt tizimlarida shaxsiy ma’lumotlardan foydalanish jarayoni ham alohida e’tibor talab qiladi, chunki ularning samaradorligi katta hajmdagi maxfiy axborotlar bilan ishlashga tayanadi. Bu jarayon ta’lim, sog‘liqni saqlash, moliya hamda davlat xizmatlari kabi sohalarda yuqori darajada maxfiy axborotlar bilan ishlashni talab etadi. Bunday ma’lumotlarning yetarli darajada himoyalanmasligi natijasida shaxsga doir ma’lumotlar uchinchi tomonga chiqib ketishi natijasida shaxs erkinligining cheklanishi jiddiy xavflarni keltirib chiqaradi. Shu bois shaxsiy ma’lumotlar maxfiyligini ta’minlash masalasi bugungi kunda sun’iy intellekt texnologiyalarining eng muhim va eng zaif bo‘g‘inlaridan biri sifatida qaraladi. Keyingi tizimli muammolardan biri bu javobgarlikning noaniqligi bo‘lib, sun’iy intellektga asoslangan texnologiyalar noto‘g‘ri qaror qabul qilgan holatlarda javobgarlik kimning zimmasiga tushishi masalasi bugungi kunda eng murakkab huquqiy muammolardan biri bo‘lib qolmoqda. Amaldagi huquqiy tizimlar asosan inson faoliyatidan kelib chiqadigan javobgarlikni belgilaydi, biroq sun’iy intellekt ishtirokidagi qarorlar jarayonida soha mutaxassislari, texnologiyani joriy etgan tashkilotlar va tizim foydalanuvchilari kabi bir nechta subyektlar qatnashadi. Javobgarlikning ular o‘rtasida aniq belgilanmaganligi esa huquqiy bo‘shliqlarni yuzaga keltirib, sud jarayonlarida va normativ-huquqiy tartibga solishda ham jiddiy qiyinchiliklar tug‘dirmoqda. Sun’iy intellektning yana bir muhim jihati uning avtonomligi, ya’ni mustaqil ishlash qobiliyatidir. Biroq qarorlar to‘liq avtomatlashtirilgan holda, inson ishtirokisiz qabul qilinganda inson nazoratining kamayishi (loss of human oversight) muammosi yuzaga keladi. Bunday holatda noto‘g‘ri qarorlarning avtomatik ravishda ko‘payishi, texnik nosozliklarning kech aniqlanishi, foydalanuvchining tizimga ortiqcha ishonib ketishi (automation bias) hamda mas’uliyatni bevosita texnologiya zimmasiga yuklash tendensiyasi ortadi. Shu bois sun’iy intellektdan foydalanishda inson ishtirokini ta’minlash muhim etik talab sifatida qaralishi lozim. Shuningdek, texnologik tengsizlik va raqamli tafovut muammolariga e’tibor qaratsak, sun’iy intellekt texnologiyalari katta moliyaviy sarmoya va rivojlangan texnik infratuzilmani talab qilgani bois rivojlangan va rivojlanayotgan mamlakatlar, shuningdek, shahar va qishloq hududlarida texnik imkoniyatlar yetarli va cheklangan sharoitlarda raqamli tafovut (digital divide) yanada ortib bormoqda. Bu holat sun’iy intellekt texnologiyalaridan teng va adolatli foydalanish imkoniyatlarini cheklab, ijtimoiy adolat tamoyiliga zid bo‘lgan vaziyatlarning yuzaga kelish ehtimolini oshiradi. Tahlillar shuni ko‘rsatadiki, sun’iy intellekt etikasi bilan bog‘liq muammolar faqat texnik xarakterga ega emas, balki inson huquqlari, ijtimoiy adolat, huquqiy javobgarlik va mas’uliyat bilan uzviy bog‘liq bo‘lgan murakkab tizimli muammolar majmuini tashkil etadi. Algoritmik diskriminatsiya, shaxsiy ma’lumotlar maxfiyligi, javobgarlikning aniq belgilanmaganligi, inson nazoratining cheklanishi hamda texnologik tengsizlik kabi omillar sun’iy intellektdan foydalanishni qat’iy ilmiy asoslangan va normativ jihatdan mustahkam mexanizmlar orqali tartibga solish hozirgi kunning dolzarb muammolaridan hisoblanadi. 4. Ta’lim tizimida sun’iy intellekt etikasi muammolari Sun’iy intellekt texnologiyalarining ta’lim tizimiga jadal joriy etilishi ta’lim jarayonining samaradorligini oshirish, individual yondashuvni ta’minlash, monitoring va baholashni 95

avtomatlashtirish kabi muhim imkoniyatlarni yaratmoqda. Biroq mazkur jarayon bilan bir qatorda ushbu texnologiyalarni qo‘llash natijasida yuzaga keladigan pedagogik, axloqiy va huquqiy masalalarni inobatga olishni ham talab etadi. Mazkur masalalar o‘quvchilarning ta’lim jarayoni bilan bog‘liq huquqlari va manfaatlarini himoya qilish, pedagogning mas’uliyatini aniq belgilash va baholash jarayonining adolatli bo‘lishi kabi ta’limning asosiy tamoyillari bilan bevosita bog‘liqdir. Avtomatlashtirilgan baholash tizimlari va pedagogik adolat muammosi. Sun’iy intellekt asosida ishlovchi avtomatlashtirilgan baholash tizimlari test topshiriqlarini tekshirish, yozma ishlarni tahlil qilish va o‘quvchining rivojlanish dinamikasini muntazam monitoring qilish imkonini yaratmoqda. Biroq ushbu texnologiyalarning pedagogik adolatlilikni ta’minlash darajasi hali ham to‘liq kafolatlanmagan. Amaliyotda bir qator muammolar kuzatilmoqda: baholash mezonlarining shaffof emasligi, kreativ fikrlash va subyektiv yondashuvlarning avtomatik modellarda yetarli aks etmasligi, o‘quvchining psixologik holati va individual xususiyatlarining e’tibordan chetda qolishi hamda xatoliklarning tizimli takrorlanish xavfi. Natijada avtomatlashtirilgan baholash ko‘pincha o‘quvchining real bilim darajasini to‘liq aks ettirmasligi mumkin, bu esa pedagogik adolat tamoyiliga zid holatlarning yuzaga kelishiga olib keladi. O‘quvchilarning shaxsiy ma’lumotlari himoyasi muammosi. Ta’lim tizimida sun’iy intellektning faol qo‘llanilishi o‘quvchilarning akademik ko‘rsatkichlari, shaxsiy profillari, xulqatvori va shaxsga doir bo‘lgan ma’lumotlari yuqori darajada maxfiy axborotlar bazasining shakllanishiga olib kelmoqda. Ushbu ma’lumotlarning yetarli darajada himoyalanmasligi natijasida noqonuniy kuzatuv, ma’lumotlarning uchinchi shaxslarga tarqalishi, shaxs erkinligining cheklanishi, o‘quvchi sha’ni va qadr-qimmatiga putur yetkazilishi kabi xavflarga sabab bo‘lishi mumkin. Shu bois ta’limda sun’iy intellektdan foydalanishda “minimal ma’lumot yig‘ish” va “belgilangan maqsadlarda foydalanish” tamoyillari asosiy etik talab sifatida qaralishi zarur. Akademik halollik muammolari. Sun’iy intellektga asoslangan chatbotlar yordamida matn yaratish, tarjima qilish va masalalarni avtomatik yechish vositalari o‘quv jarayonida akademik halollik muammosini keskinlashtirmoqda. O‘quvchilar tomonidan mustaqil ishlarning sun’iy intellekt yordamida bajarilishi, ijodiy topshiriqlarni tayyorlash, manbalarni ko‘rsatmasdan foydalanish holatlari ko‘payishi natijasida bilimlarning real egallanish darajasi pasayishi mumkin. Bu esa ta’lim sifatiga, baholashga bo‘lgan ishonchga va ta’lim natijalarining haqiqiyligiga jiddiy xavf tug‘diradi. Pedagog rolining transformatsiyasi va kasbiy mas’uliyat muammosi. Sun’iy intellektning ta’lim jarayoniga jadal kirib kelishi pedagogning an’anaviy rolini tubdan o‘zgartirmoqda. Endilikda pedagoglar bilim beruvchi rolidan yo‘naltiruvchi, maslahatchi va o‘quv jarayonini boshqaruvchi subyektga aylanmoqda. Bu esa ta’lim sifatini oshirish uchun yangi imkoniyatlar yaratish bilan birga, ayrim xavflarni ham yuzaga chiqaradi. Jumladan, baholash, tahlil va qaror qabul qilish jarayonlarining ortiqcha darajada texnologiyaga topshirilishi pedagog mas’uliyatining pasayishiga olib kelishi mumkin. Shu sababli pedagoglar sun’iy intellektni yordamchi vosita sifatida qabul qilishi, o‘quv jarayoniga oid yakuniy qarorlar uchun mas’uliyatni esa o‘z zimmasida saqlab qolishi zarur. Raqamli tengsizlik va inklyuzivlik muammolari. Sun’iy intellekt texnologiyalarining ta’lim tizimiga joriy etilishi barcha ta’lim muassasalari uchun bir xil imkoniyatlarni ta’minlab bera olmayapti. Texnik infratuzilmaning yetarli emasligi, internet tarmog‘i sifatidagi farqlar hamda raqamli savodxonlik darajasidagi tafovutlar shahar va qishloq maktablari, iqtisodiy jihatdan 96

rivojlangan va chekka hududlar, shuningdek, ijtimoiy himoyaga muhtoj qatlamlar o‘rtasida yangi “raqamli tengsizlik” to‘lqinining yuzaga kelishiga sabab bo‘lmoqda. Bu holat esa ta’limda teng imkoniyatlar tamoyilini zaiflashtirib, ayrim guruhlarning zamonaviy ta’lim resurslari va sun’iy intellekt texnologiyalaridan foydalanish imkoniyatlarini cheklab qo‘yishi mumkin. Yuqoridagilarni inobatga olgan holda tahlillar shuni ko‘rsatadiki, ta’lim tizimida sun’iy intellektdan foydalanish pedagogik samaradorlikni oshirish bilan birga jiddiy etik xavflarni ham yuzaga keltirmoqda. Avtomatlashtirilgan baholashning adolatliligi, o‘quvchining shaxsiy ma’lumotlarining himoyasi, akademik halollik, pedagog mas’uliyati va raqamli tengsizlik muammolari ta’limda sun’iy intellektdan mas’uliyatli va ehtiyotkorona foydalanishni taqozo etadi. 5. Xalqaro tavsiyalar asosida ta’lim jarayonida qo’llaniladigan etik me’yorlar va O‘zbekiston ta’lim tizimi uchun moslashtirish yo‘nalishlari. Sun’iy intellekt texnologiyalarining ta’lim tizimiga joriy etilishi global miqyosda strategik ustuvor yo‘nalishlardan biri sifatida qaralmoqda. Biroq ushbu jarayonning muvaffaqiyati nafaqat texnologik infratuzilmaga, balki sun’iy intellektdan mas’uliyatli, adolatli va xavfsiz foydalanish uchun ishlab chiqilgan xalqaro etik tavsiyalarning ta’lim tizimiga to‘g‘ri integratsiya qilinishiga ham bevosita bog‘liqdir. Shu sababli ushbu bo‘limda yetakchi xalqaro tashkilotlar tomonidan ishlab chiqilgan ta’limga oid tavsiyalar umumlashtiriladi hamda O‘zbekiston ta’lim tizimi uchun moslashtirilgan konseptual yo‘nalishlar taklif etiladi. 5.1. Ta’lim jarayonida qo’llash tavsiya etiladigan xalqaro etik me’yorlar. UNESCO tomonidan sun’iy intellekt etikasi bo‘yicha ishlab chiqilgan tavsiyalarda ta’lim sohasi inson manfaatlariga yo‘naltirilgan sun’iy intellekt, teng imkoniyatlar va pedagogik mas’uliyatni ta’minlash nuqtai nazaridan ustuvor yo‘nalish sifatida belgilangan. UNESCO yondashuviga ko‘ra, ta’limda sun’iy intellektdan foydalanish jarayonida inson huquqlari to‘liq ta’minlanishi, o‘quvchilarning shaxsiy ma’lumotlari qat’iy himoyalanishi, raqamli tenglik va inklyuzivlik prinsiplariga amal qilinishi zarur. Shuningdek, pedagogning ta’limdagi yetakchi roli saqlanib qolishi hamda sun’iy intellekt asosida qabul qilinadigan qarorlar ustidan doimiy inson nazorati mavjud bo‘lishi asosiy talablar sifatida belgilanadi. Shuningdek, UNESCO ta’lim tizimida sun’iy intellektning etik ta’sirini baholash (ethical impact assessment) va tayyorgarlik holatini aniqlash (readiness assessment) mexanizmlarini joriy etishni tavsiya etadi. Bu ta’lim muassasalarida sun’iy intellekt texnologiyalarini joriy etishdan avval ularning ijtimoiy, axloqiy va pedagogik oqibatlarini bashorat qilish imkonini beradi. OECD yondashuvida ta’lim sohasida sun’iy intellekt bilan bog‘liq asosiy urg‘u raqamli kompetensiyalarni rivojlantirish, o‘quvchilarda sun’iy intellektga nisbatan tanqidiy fikrlashni shakllantirish, algoritmik xatoliklarni anglay olish ko‘nikmalarini hosil qilish hamda o‘qituvchilarning raqamli-pedagogik kompetensiyalarini oshirishga qaratiladi. Shuningdek, mehnat bozori ehtiyojlari bilan uyg‘unlikni ta’minlash va “umr bo‘yi ta’lim” konsepsiyasini qo‘llabquvvatlash ham ushbu yondashuvning muhim ustuvor yo‘nalishlari sifatida e’tirof etiladi. Yevropa Komissiyasi tomonidan ilgari surilgan “Trustworthy AI” konsepsiyasida esa ta’lim tizimi uchun sun’iy intellekt asosidagi avtomatlashtirilgan baholash tizimlarining shaffofligi, o‘quvchi ma’lumotlarini ishonchli texnik himoyalash, tavsiya beruvchi algoritmlarda adolatlilikni 97

ta’minlash hamda sun’iy intellekt tizimlarini majburiy sertifikatlash va auditdan o‘tkazish kabi talablar asosiy amaliy yo‘nalishlar sifatida belgilangan. 5.2. Ta’limda Mas’uliyatli sun’iy intellekt (Responsible AI) konsepsiyasi. Zamonaviy ilmiy adabiyotlar hamda xalqaro normativ hujjatlarda ta’lim tizimi uchun Mas’uliyatli sun’iy intellekt (Responsible AI) konsepsiyasi eng muhim ustuvor yo‘nalishlardan biri sifatida e’tirof etilmoqda. Ushbu yondashuv sun’iy intellekt tizimlarini ishlab chiqish va ularni amaliyotga joriy etishda faqat texnologik samaradorlikni emas, balki inson huquqlari, axloqiy me’yorlar, xavfsizlik va ijtimoiy mas’uliyatni ham bir butun tizim sifatida ta’minlashni nazarda tutadi. Mas’uliyatli sun’iy intellekt modeli quyidagi asosiy tarkibiy tamoyillarga tayangan holda shakllanadi: axloqiy me’yorlarga muvofiqlik (etika), amaldagi qonunchilik talablariga moslik (huquqiylik), qabul qilinayotgan qarorlarning tushuntiriluvchanligi va ochiqligi (shaffoflik), qarorlar ustidan inson ishtiroki va nazoratining mavjudligi (human-in-the-loop), tizimlarning xavfsizligi va barqaror ishlashi, shuningdek, javobgarlikning aniq belgilanishi. Ta’lim tizimi nuqtai nazaridan Mas’uliyatli sun’iy intellekt shuni talab etadiki, sun’iy intellekt pedagogning o‘rnini bosuvchi asosiy subyekt emas, balki uni qo‘llab-quvvatlovchi, o‘quv jarayonini samarali tashkil etishga xizmat qiluvchi yordamchi texnologiya sifatida faoliyat yuritishi lozim. Bunda barcha muhim va yakuniy pedagogik qarorlar doimo inson tomonidan qabul qilinishi, avtomatlashtirilgan tizimlardan esa asosan maslahat berish, tahlil qilish, monitoring va prognozlash vositasi sifatida foydalanilishi zarur. Aynan shu yondashuv ta’limda texnologiyalar bilan pedagog mas’uliyati o‘rtasida sog‘lom muvozanatni ta’minlaydi. 5.3. O‘qituvchilarning raqamli-etik kompetensiyalarini rivojlantirish. Ta’limda sun’iy intellektdan samarali, xavfsiz va mas’uliyatli foydalanish bevosita o‘qituvchilarning raqamli-etik kompetensiyalari rivojlanganlik darajasiga bog‘liqdir. Zamonaviy raqamli muhit sharoitida pedagog nafaqat texnologiyalardan foydalana olishi, balki ularning imkoniyatlari va cheklovlarini to‘g‘ri baholay olishi ham zarur. Xalqaro tavsiyalarga muvofiq, bugungi pedagog sun’iy intellekt asosida ishlovchi platformalardan ongli foydalanish, algoritmik qarorlarning imkoniyatlari bilan birga xavf-xatarlarini ham tushunish, o‘quvchilarning shaxsiy ma’lumotlari maxfiyligini ta’minlash, avtomatlashtirilgan baholash natijalarini tanqidiy tahlil qilish hamda akademik halollikni nazorat qilish kabi asosiy kompetensiyalarga ega bo‘lishi lozim. Mazkur kompetensiyalarni shakllantirish pedagogning sun’iy intellektga mutlaq tayanib qolishining oldini oladi, tanqidiy fikrlashni mustahkamlaydi va ta’lim jarayonida texnologiyalar bilan inson mas’uliyati o‘rtasidagi muvozanatni ta’minlaydi. Shu sababli o‘qituvchilarni qayta tayyorlash va malakasini oshirish tizimlariga sun’iy intellekt etikasi, raqamli xavfsizlik va algoritmik savodxonlikka oid maxsus modullarni kiritish muhim ilmiy-amaliy zarurat sifatida namoyon bo‘lmoqda. 5.4. O‘zbekiston ta’lim tizimi uchun moslashtirish yo‘nalishlari. O‘zbekiston Respublikasida ta’lim tizimini raqamlashtirish va sun’iy intellekt texnologiyalarini joriy etish bo‘yicha muhim islohotlar amalga oshirilayotgan bir sharoitda sun’iy intellekt etikasi 98

masalalarini milliy ta’lim siyosatiga integratsiya qilish alohida ahamiyat kasb etadi. Xalqaro tajriba va milliy sharoitdan kelib chiqib, quyidagi moslashtirish yo‘nalishlari taklif etiladi: Normativ-huquqiy moslashtirish: Ta’lim tizimida sun’iy intellektdan foydalanishni samarali va xavfsiz yo‘lga qo‘yish uchun uni tartibga soluvchi maxsus normativ-huquqiy hujjatlarni ishlab chiqish zarur. Mazkur hujjatlarda o‘quvchi va pedagoglarning shaxsiy ma’lumotlarini himoyalash mexanizmlari, avtomatlashtirilgan baholash tizimlarining shaffofligi, shuningdek, sun’iy intellekt asosida qabul qilinadigan qarorlar uchun javobgarlikning aniq chegaralari belgilab qo‘yilishi lozim. Bunday normativ-huquqiy asos ta’lim jarayonida sun’iy intellektdan foydalanishda qonuniylik, ochiqlik va mas’uliyat tamoyillarini ta’minlashga xizmat qiladi. Pedagog kadrlarni tayyorlash va qayta tayyorlash: Ta’lim tizimida sun’iy intellektdan mas’uliyatli va samarali foydalanishni ta’minlash uchun pedagog kadrlarni tayyorlash va qayta tayyorlash tizimini zamonaviy talablar asosida takomillashtirish zarur. Jumladan, pedagoglarni tayyorlash dasturlariga “Sun’iy intellekt asoslari”, “Sun’iy intellekt etikasi” hamda “Raqamli pedagogika” fanlarini bosqichma-bosqich joriy etish muhim ahamiyat kasb etadi. Bu fanlar orqali o‘qituvchilarda sun’iy intellekt texnologiyalarining ishlash tamoyillari, axloqiy jihatlari hamda raqamli ta’lim muhitida samarali ishlash ko‘nikmalari shakllantiriladi. Etik monitoring va ekspertiza tizimini yaratish: Ta’lim muassasalarida sun’iy intellekt asosida faoliyat yurituvchi platformalar va tizimlar uchun majburiy etik ekspertiza hamda doimiy monitoring mexanizmlarini joriy etish zarur. Bu jarayon sun’iy intellektdan foydalanishda axloqiy me’yorlarga rioya qilinishini, shaxsiy ma’lumotlar himoyasini va algoritmik qarorlarning adolatliligini doimiy nazorat ostida ushlab turishga xizmat qiladi. O‘quvchilar uchun sun’iy intellekt savodxonligini oshirish: O‘quvchilarda sun’iy intellekt savodxonligini shakllantirish ham muhim ustuvor yo‘nalishlardan biri hisoblanadi. Bunda ularda sun’iy intellektning imkoniyatlari va cheklovlarini tushunish, algoritmik xatolarga tanqidiy yondashish hamda raqamli xavfsizlik qoidalariga rioya qilish ko‘nikmalarini rivojlantirish lozim. Bu esa o‘quvchilarning texnologiyadan ongli va xavfsiz foydalanish madaniyatini mustahkamlaydi. Raqamli tenglik va inklyuziyani ta’minlash: Raqamli texnologiyalar, jumladan sun’iy intellektdan foydalanishda raqamli tenglik va inklyuzivlikni ta’minlash alohida ahamiyatga ega. Ayniqsa, qishloq va chekka hududlardagi ta’lim muassasalari uchun internet infratuzilmasini rivojlantirish, sun’iy intellekt platformalariga teng kirish imkoniyatlarini yaratish bo‘yicha davlat dasturlarini kengaytirish zarur. Bu esa barcha hududlardagi o‘quvchilar uchun teng ta’lim olish imkoniyatlarini yaratishga xizmat qiladi. 2-rasm. Ta’limda sun’iy intellekt etikasini joriy etish bo‘yicha amaliy tavsiyalar 99