Artificial intelligence in food safety Authors: Floor van Meer, Masami Takeuchi, Phillis E. Ochieng, Raffaella Tavelli, Arjen Gerssen, Bas H. M. van der Velden Published: 2026 ABSTRACT Artificial Intelligence (AI) is revolutionizing many fields including food safety. In food safety, the available data generated within the agrifood systems can potentially feed into AI applications. This review presents an overview of the use of AI in different fields of food safety. The study introduces a framework of AI in food safety, focusing on the research domain, application context, data collection method, and AI technique. The AI tool ASReview, which uses active learning to select top-priority papers, was used to conduct the study in a systematic way. Over 150 peer-reviewed journal publications were reviewed and described according to this framework. The review concludes with a summary of the challenges for AI in food safety and an outlook of future opportunities for AI in food safety. Introduction Food safety is a critical component of food security and healthy diets. Food safety has significant public health and economic impacts and is everyone’s responsibility in the agrifood systems1. Consequently, researchers in academia and industry have devoted significant attention to this global issue2. The term food safety often covers all practices aimed at ensuring the safety of our food, including proper production, processing, handling, storage, transportation, retail, and preparation to prevent various food safety hazards to become a health risk3. Therefore, research in food safety needs to have a holistic approach to capture all relevant elements in the whole agrifood system. This includes national food safety management, including legislative matters and regulatory frameworks, in addition to systematic investigation of relevant food safety risks, both chemical and microbiological, as well as issues around food fraud. Artificial intelligence (AI) is transforming processes that assist decision-making across many fields4. AI is a field in computer science that focuses on methods and systems that enable machines to perceive, learn, and take intelligent actions to achieve specific goals5. In food safety research, AI typically refers to machine learning and deep learning. In machine learning, researchers or engineers extract features from data and use algorithms to analyze data, learn patterns, and make predictions6,7. In deep learning, which is a subset of machine learning, raw data is directly processed by algorithms without requiring explicit feature extraction8. The advances in machine learning and deep learning are expected to enable new applications of AI across various domains of food safety research. Global developments in AI for food safety reflect a rapidly expanding research field, with publications increasing from just one paper in 2012 to 46 in 2023. This growth is expected to continue as advances in computational power, data availability, and AI methods increasingly enable real-time applications across the food chain, from hazard detection to supply chain monitoring. Therefore, the goal of this review is to synthesize the rapidly growing evidence base, identify in which food safety research domains AI is most actively being applied, and highlight remaining gaps, key challenges, and opportunities for future research. For this review, we focused on peer-reviewed scientific literature where AI was used in the field of food safety. In this systematic review, we present a comprehensive overview of the use of AI in the different fields of food safety. The scoping review method using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework was employed. We included peer-reviewed journal publications and peer-reviewed conference proceedings. To prioritize relevance, we employed the Active learning for Systematic Reviews toolbox9, which streamlines systematic reviews by using machine learning to assist in screening large volumes of literature. Through active learning, the tool ranks papers based on predicted relevance, refining its recommendations every time the user provides input, thus enabling more efficient identification of relevant studies in extensive datasets9. This review is organized as follows. “Methods” presents the results of the systematic search, structured by research domain and implementation context, and summarizes the AI techniques used. “AI in food safety - Food safety research domain” discusses key challenges for applying AI in food safety. “Discussion” concludes with a summary and future directions. Methods The search strategy for paper inclusion was as follows: we conducted a query in SCOPUS using ("artificial intelligence" OR "machine learning") AND "food safety". We selected SCOPUS due to its broad coverage across relevant disciplines and the ability to export full query results in a format compatible with ASReview. Our initial search yielded 783 publications. Using ASReview, we screened 434 records (title/abstract). Screening was stopped when no additional relevant papers were identified in the subsequent 20 records, after which the remaining 349 unscreened records were considered unlikely to be relevant. This reduced the manual screening workload by 349 records compared to screening the full set without ASReview. Of the 434 records that were screened by a researcher at title/abstract level, 188 were marked as potentially relevant, and 161 (21% of all retrieved records) remained eligible after full-text screening (Fig. 1a).Fig. 1Overview of the review workflow and schematic classification framework.Top left: Overview of the paper inclusion. Top right: Identification of the relevant papers with AS Review (black line) versus traditional random reading of papers (grey dashed line). It can be seen that the curve plateaus around 370 reviewed papers. Bottom: Schematic framework in which this review grouped the included studies. Overview of the review workflow and schematic classification framework. Top left: Overview of the paper inclusion. Top right: Identification of the relevant papers with AS Review (black line) versus traditional random reading of papers (grey dashed line). It can be seen that the curve plateaus around 370 reviewed papers. Bottom: Schematic framework in which this review grouped the included studies. We included primary research articles that applied AI methods to address food safety–related research questions. We considered AI in the scope of this review to comprise machine learning approaches, excluding conventional statistical regression methods. We excluded review articles, studies that did not use AI, and studies that were not directly focused on food safety. For example, Logan et al.10 (“Hyperspectral imaging and machine learning for monitoring produce ripeness”) was excluded because, although it applies machine learning, it focuses on ripeness assessment rather than food safety. Guo et al.11 (“Quantitative assessment of zearalenone in maize using multivariate algorithms coupled to Raman spectroscopy”) was excluded because, although it addresses a food safety hazard, it relies on partial least squares regression, which we classified as a conventional multivariate modelling technique rather than an AI method within the scope of this review. For multiple publications by the same authors on similar topics, we selected the journal publication when available or the first conference proceeding. This review includes papers published until April 1st, 2024, a starting date was not specified since it is a relatively new field. Figure 1 (top) shows how using the ASReview toolbox enabled to increase the ratio of relevant papers over the total of reviewed papers compared to traditional search methods. To structure our findings, we grouped the studies into several categories, see Fig. 1 (bottom) for a schematic overview. We first categorized all studies into five research domains based on the most common topics covered by AI in food safety studies: (1) microbiological hazards, (2) chemical hazards, (3) food fraud/authenticity, (4) foodborne disease surveillance/investigation, (5) other food safety issues. The category “other food safety issues” includes research looking into multiple food safety hazards, shelf-life and food spoilage issues, regulatory compliance, traceability and food recalls. We then further categorized the studies based on the implementation context (i.e., the specific purpose) of AI. These are categorized as aid for laboratory testing and analysis, risk analysis, risk forecasting, priority setting for inspection and border testing, and other efficiency for the activities of food safety authorities (regulatory efficiency). For each study, we recorded the data collection method and the AI technique employed. An overview of the proportions of food safety research domains and the respective implementation contexts used in the reviewed articles is depicted in Fig. 2. The following sections will describe each food safety research domain in more detail. For each food safety research domain, an accompanying table is included in which all papers are systematically analyzed.Fig. 2Food safety research domains, implementation contexts, and AI method trends.Top: Overview of the food safety research domains (middle ring of the donut plot), implementation contexts per research domain (outer ring of the donut plot), and the use of deep learning per research domain in percentages (inner ring of the donut plot). FBD = foodborne disease. Bottom: There is an increasing use of AI in food safety. The pink bars represent classical machine learning, the gold bars represent deep learning. The transparent bars represent an extrapolation for the year 2024. Food safety research domains, implementation contexts, and AI method trends. Top: Overview of the food safety research domains (middle ring of the donut plot), implementation contexts per research domain (outer ring of the donut plot), and the use of deep learning per research domain in percentages (inner ring of the donut plot). FBD = foodborne disease. Bottom: There is an increasing use of AI in food safety. The pink bars represent classical machine learning, the gold bars represent deep learning. The transparent bars represent an extrapolation for the year 2024. AI in food safety - Food safety research domain In the following paragraphs we will describe the overall trends and findings per food safety research domain. For clarity of reading, we only discuss typical example papers for each sub application in detail, but we also provide a meticulous overview of all identified papers, categorized on implementation context, hazard, data collection method, and representative AI technique (Tables 1–5).Table 1Papers using AI in the domain of microbiological hazardsImplementation contextPaperTarget organismData collection methodAI TechniqueLaboratory testingAhmed et al.47VibrioOptical spectroscopy (Scatterometer)SVMChung et al.48Pathogenic bacteriaGenomics (16S)RFDu et al.49Pathogenic bacteriaOptical spectroscopy (Raman)GAN SVMFeng et al.50Pathogenic bacteriaMicroscopyComputer visionFong et al.51Pathogenic bacteriaHyperspectral microscope imagingResNet CNNGonçalves et al.12SalmonellaE-noseLDAHussain et al.52Pathogenic bacteriaOptical spectroscopy (Scatterometer)SVMJia et al.53Pathogenic bacteriaChromogenic assayDNNJia et al.54Pathogenic bacteriaChromogenic assayDNNKang et al.55E. coliHyperspectral microscope imagingAutoencoderKang et al.56Pathogenic bacteriaHyperspectral microscope imagingCNNLi et al.57ListeriaMass Spectrometry-Based Analysis (maldi-ToF-MS)SVM CNN Denoising AutoencoderMa et al.58E. coliMicroscopyCNNMaeda et al.59FungiMass Spectrometry-Based Analysis (maldi-tof ms)SVM RFMangmee et al.60SalmonellaMass Spectrometry-Based Analysis (maldi-ToF-MS)CNNMarcoux et al.61Pathogenic bacteriaOptical spectroscopy (scatter)BNMorgan et al.62E. coliOptical spectroscopy (IR)RFPark et al.63Pathogenic bacteriaHyperspectral microscope imagingCNNQin et al.64Pathogenic bacteriaOptical spectroscopy (Raman)SVMRahi et al.65E. coliOptical spectroscopy (Vis/NIR)GATsakanikas et al.66Pathogenic bacteriaOptical spectroscopy multispectral), Colony countsSVMVakilian67Pathogenic bacteriaOptical spectroscopy (Raman)GA ANNXu et al.68CampylobacterColony countsRFYadav et al.69E. coliOptical spectroscopy (fluorescence)CNN SVMYamamoto et al.70Pathogenic bacteriaOptical spectroscopy (Raman)SVMYan et al.71E. coliOptical spectroscopy (surface-enhanced Raman)SVM XGBOOSTYan et al.72Pathogenic bacteriaOptical spectroscopy (Raman)DTYang et al.73Pathogenic bacteriaChromogenic assayANNYi et al.74Pathogenic bacteriaMicroscopyCNNZeng et al.75Pathogenic bacteriaOptical spectroscopy (Raman)RNNZeng et al.76Pathogenic bacteriaOptical spectroscopy (Raman)XGBOOSTZhang et al.77Pathogenic bacteriaOptical spectroscopy (Raman)Active LearningZhang et al.78Pathogenic bacteriaOptical spectroscopy (Raman)RF SVM XGBOOST KNN ANNForecastingAl et al.79E. coliColony countsANN RF SVMBerhilevych et al.80enterobacteriaColony countsANNBolinger et al.81SalmonellaGenomics (whole genome sequencing)RFGmeiner et al.82ListeriaGenomics (whole genome sequencing)RF SVM NNKaranth et al.83SalmonellaGenomics (whole genome sequencing)RF SVMLin et al.84SalmonellaMicroscopyCNNMunck et al.85SalmonellaGenomics (whole genome sequencing)RFStocker et al.86E. coliColony countsXGBOOST KNN SVM RFTanui et al.13SalmonellaGenomics (whole genome sequencing)RFWang et al.87Pathogenic microorganismsMetabolomicsANNWeller et al.88E. coli SalmonellaChromogenic assaySVM KNN XGBOOSTRisk factorsCox89SalmonellaPCR-based molecular identificationRF BNGolden et al.90ListeriaColony counts, PCRRFIm et al.91E. coliGenomics (whole genome sequencing)RFKaranth et al.92SalmonellaGenomics (whole genome sequencing)Elastic net regressionMurphy et al.93Pathogenic bacteriaColony countsRFPillai et al.94SalmonellaColony countsGAN XGBOOST RFRodríguez et al.95SalmonellaPCR-based molecular identificationRFStanosheck et al.96E. coliColony countsSVM RFToro et al.97SalmonellaPCR-based molecular identificationRFVangay et al.98ListeriaCAMP test, Gram stain, EcoRI ribotypingRF SVMWeller et al.14E. coliChromogenic assayNN RF SVMWeller et al.99ListeriaPCR-based molecular identificationRF SVMANN artificial neural network, BN Bayesian network, CNN convolutional neural network, DNN deep neural network, DT decision tree, ELM extreme learning machine, GA genetic algorithm, GAN generative adversarial network, KNN K-nearest neighbor, LDA linear discriminant analysis, NB Naïve Bayes, ResNet residual network, RNN recurrent neural network, RF random forest, SVM support vector machine, XGBOOST extreme gradient boosting.Table 2Papers using AI in the domain of chemical hazardsImplementation contextPaperTarget chemicalData collection methodAI techniqueLaboratory testingAtaş et al.100MycotoxinsHyperspectral microscope imagingSVMChen et al.101MultipleMass Spectrometry-Based Analysis (LC-HRMS)ANN RF SVMCheng et al.102CadmiumHyperspectral imagingELMDarwish et al.103Foreign objectsNear-field microwave sensingSVMHe et al.15PesticidesOptical spectroscopy (fluorescence imaging)SVM ANNHu et al.104PesticidesHyperspectral imagingCNNJin et al.105Preservatives and heavy metalsOptical spectroscopy (Raman)DF CNNKalkan et al.106MycotoxinsOptical spectroscopy (multispectral)SVM ANNKim et al.107MycotoxinsOptical spectroscopy (Raman)SVMLu et al.108Veterinary drugsOptical spectroscopy (Raman)BP-NNMa et al.109NitrateOptical spectroscopy (ATR-FTIR)ELMMa et al.110MycotoxinsBioluminescenceSVM RF NNSmeesters et al.111AcrylamideOptical spectroscopyELM LDASmeesters et al.112AcrylamideOptical spectroscopyLDASun et al.113Sudan redElectrochemical sensorCNN ResNetTan et al.114HistamineOptical spectroscopy (Raman)SVMTang et al.115PesticidesElectrochemical sensorSVM XGBOOSTWang et al.116Sulfur dioxideImage analysisSVM BP-NN RFWang et al.117PesticidesOptical spectroscopy (surface enhanced Raman)SVMXu et al.118Mycotoxins, heavy metalsCombined dataLSTM XGBOOSTZhao et al.119PropachlorElectrochemical sensorXGBOOSTForecastingBranstad-Spates et al.120MycotoxinsELISAGradient BoostingHuang et al.16CadmiumSeveralSVM RF BP-NNLiu et al.121MycotoxinsMass Spectrometry-Based Analysis (LC-MS/MS)BNMa et al.122CadmiumSeveralRF ANNMarvin et al.123MultipleCombined dataBNMarzec-Schmidt et al.124MycotoxinsMass Spectrometry-Based Analysis (LC-MS/MS)SVM RFMateo et al.125MycotoxinsMass Spectrometry-Based Analysis (LC-MS/MS)XGBOOST RF NNMi et al.126CadmiumSeveralBP-NNPortz et al.127Veterinary drugsMass Spectrometry-Based Analysis (LC-MS/MS)DT NNWang et al.128MycotoxinsMass Spectrometry-Based Analysis (LC-MS/MS)RFXie et al.129MycotoxinsMetabolomicsRF XGBOOSTRisk factorsBouzembrak and Marvin18MultipleCombined dataBNCamardo Leggieri et al.17MycotoxinsMass Spectrometry-Based Analysis (LC-MS/MS)DNNHu et al.130Heavy metalsInductively Coupled Plasma Optical Emission SpectrometryRFPetrea et al.131Heavy metalsSeveralRFTang et al.132CadmiumGenomics (SNP)XGBOOST SVM KNN RFThao et al.133PesticidesVideo analysisCNNWang et al.128MycotoxinsMass Spectrometry-Based Analysis (LC-MS/MS)SVM XGBOOSTXiang et al.134PFASSeveralRF SVMANN artificial neural network, BN Bayesian network, CNN convolutional neural network, DNN deep neural network, DT decision tree, ELM extreme learning machine, GA genetic algorithm, GAN generative adversarial network, KNN K-nearest neighbor, LDA linear discriminant analysis, NB Naïve Bayes, ResNet residual network, RNN recurrent neural network, RF random forest, SVM support vector machine, XGBOOST extreme gradient boosting. Combined data refers to studies that used more than one data collection method.Table 3Papers using AI in the domain of food fraudImplementation contextPaperData collection methodAI techniqueLaboratory testingAnantha Krishna et al.135Mass Spectrometry-Based Analysis (LC-MS/MS)SVM KNN XGBOOST ANNAsharindavida et al.136Optical spectroscopy (IR)SVM KNN RFChauvin et al.137Hyperspectral imagingGAChen et al.138Optical spectroscopy (FT-NIR)ResNet CNNChung et al.139Optical spectroscopy (IR)XGBOOSTde Oliveira et al.140Mass Spectrometry-Based Analysis (Direct-infusion MS)DTFeng et al.141Optical spectroscopy (Raman)SVM RF XGBOOSTGao et al.142Mass Spectrometry-Based Analysis (ICP-MS)RF SVMGundavarapu et al.143pH sensor, methane sensor etc.RFJo et al.20Optical spectroscopy (diffuse reflectance)AlexNet CNNKim et al.144Genomics (SNP)SVM kNNLi et al.145Combined dataRFLiu et al.19FT-NIRResNet CNNLyu et al.146Mass Spectrometry-Based Analysis (LC-MS/MS)SVM RFMohammadian et al.147E-noseSVM ANNNallan Chakravartula et al.148Optical spectroscopy (FT-NIR)CNNNanou et al.149Optical spectroscopy (laser breakdown)SVMPradana-López et al.150Image analysisResNetSaha et al.151Optical spectroscopy (NIR)CNNShin et al.152Optical spectroscopy (laser breakdown)SVM ANNWu et al.153Optical spectroscopy (fluorescence)KNN RF SVM CNNYang et al.154E-tongueCNNRisk factorsBuyuktepe et al.155Combined dataLIME SHAPLim et al.156GC-FIDDNNBorder inspectionKusuma et al.157Genomics (whole genome sequencing)SVMRegulatory efficiencyChang et al.21Combined dataRFRortais et al.28Combined dataLDAANN artificial neural network, BN Bayesian network, CNN convolutional neural network, DNN deep neural network, DT decision tree, ELM extreme learning machine, GA genetic algorithm, GAN generative adversarial network, KNN K-nearest neighbor, LDA linear discriminant analysis, NB Naïve Bayes, ResNet residual network, RNN recurrent neural network, RF random forest, SVM support vector machine, XGBOOST extreme gradient boosting. Combined data refers to studies that used more than one data collection method.Table 4Papers using AI in the domain of food-borne illnessImplementation contextPaperData collection methodAI techniqueForecastingTeyhouee et al.158Combined dataSVMRisk factorsGoldberg et al.159Combined dataNLPMaharana et al.160Combined dataSVM BERTTao et al.161Combined dataBERTZhang et al.22Combined dataXGBOOSTZheng et al.162Combined dataRFRegulatory efficiencySadilek et al.163Combined dataSVMSadilek et al.164Combined datalog-linear maximum entropySadilek et al.23Combined datalog-linear maximum entropyBERT bidirectional encoder representations from transformers, NLP natural language processing, RF random forest, SVM support vector machine, XGBOOST extreme gradient boosting. Combined data refers to studies that used more than one data collection method.Table 5Papers using AI in food safety in a broader senseImplementation contextPaperData collection methodAI techniqueLaboratory testingAzman et al.165E-noseSVM KNNChen et al.166Not availableNot availableMadan et al.167Image analysisResnet CNN AlexnetXiao et al.168Hyperspectral microscope imagingActive learningZhong et al.24Optical spectroscopy (Raman)LSTMForecastingGeng et al.169Combined dataELMMakridis et al.170Combined dataRNNMakridis et al.171Text analysis (food recalls)DNN ResNetNogales et al.172SeveralCNN RFOldroyd et al.25Combined dataRF SVMZhang et al.173Combined dataELMZuo et al.174SeveralDenoising AutoencoderZuo et al.175Combined dataGNNRisk factorsLiu et al.176Combined dataBNSetiawan et al.177Image analysisDCNN ResNetZhu et al.178Combined dataBP-NN SVMBorder inspectionBisgin et al.179MicroscopyANN SVMWu et al.180Combined dataRF BNWu et al.26Combined dataXGBOOSTRegulatory efficiencyBarros et al.181Combined dataGAChen and Zhang182Text analysis (social media)SVM BERTKate et al.183Text analysis (news)SVMSilva et al.184Combined dataANN KNN DT NBSong et al.185Combined dataNB SVM XGBOOST CNN LSTM BERTSong et al.186Combined dataBERT LSTM SVMTalari et al.187Combined dataRF KNN SVMTonda et al.27Combined dataRFWang et al.188Combined dataELMZhang et al.189Combined dataLSTM SVMANN artificial neural network, BN Bayesian network, CNN convolutional neural network, DNN deep neural network, DT decision tree, ELM extreme learning machine, GA genetic algorithm, GAN generative adversarial network, KNN K-nearest neighbor, LDA linear discriminant analysis, NB Naïve Bayes, ResNet residual network, RNN recurrent neural network, RF random forest, SVM support vector machine, XGBOOST extreme gradient boosting. Combined data refers to studies that used more than one data collection method. Table 1Papers using AI in the domain of microbiological hazardsImplementation contextPaperTarget organismData collection methodAI TechniqueLaboratory testingAhmed et al.47VibrioOptical spectroscopy (Scatterometer)SVMChung et al.48Pathogenic bacteriaGenomics (16S)RFDu et al.49Pathogenic bacteriaOptical spectroscopy (Raman)GAN SVMFeng et al.50Pathogenic bacteriaMicroscopyComputer visionFong et al.51Pathogenic bacteriaHyperspectral microscope imagingResNet CNNGonçalves et al.12SalmonellaE-noseLDAHussain et al.52Pathogenic bacteriaOptical spectroscopy (Scatterometer)SVMJia et al.53Pathogenic bacteriaChromogenic assayDNNJia et al.54Pathogenic bacteriaChromogenic assayDNNKang et al.55E. coliHyperspectral microscope imagingAutoencoderKang et al.56Pathogenic bacteriaHyperspectral microscope imagingCNNLi et al.57ListeriaMass Spectrometry-Based Analysis (maldi-ToF-MS)SVM CNN Denoising AutoencoderMa et al.58E. coliMicroscopyCNNMaeda et al.59FungiMass Spectrometry-Based Analysis (maldi-tof ms)SVM RFMangmee et al.60SalmonellaMass Spectrometry-Based Analysis (maldi-ToF-MS)CNNMarcoux et al.61Pathogenic bacteriaOptical spectroscopy (scatter)BNMorgan et al.62E. coliOptical spectroscopy (IR)RFPark et al.63Pathogenic bacteriaHyperspectral microscope imagingCNNQin et al.64Pathogenic bacteriaOptical spectroscopy (Raman)SVMRahi et al.65E. coliOptical spectroscopy (Vis/NIR)GATsakanikas et al.66Pathogenic bacteriaOptical spectroscopy multispectral), Colony countsSVMVakilian67Pathogenic bacteriaOptical spectroscopy (Raman)GA ANNXu et al.68CampylobacterColony countsRFYadav et al.69E. coliOptical spectroscopy (fluorescence)CNN SVMYamamoto et al.70Pathogenic bacteriaOptical spectroscopy (Raman)SVMYan et al.71E. coliOptical spectroscopy (surface-enhanced Raman)SVM XGBOOSTYan et al.72Pathogenic bacteriaOptical spectroscopy (Raman)DTYang et al.73Pathogenic bacteriaChromogenic assayANNYi et al.74Pathogenic bacteriaMicroscopyCNNZeng et al.75Pathogenic bacteriaOptical spectroscopy (Raman)RNNZeng et al.76Pathogenic bacteriaOptical spectroscopy (Raman)XGBOOSTZhang et al.77Pathogenic bacteriaOptical spectroscopy (Raman)Active LearningZhang et al.78Pathogenic bacteriaOptical spectroscopy (Raman)RF SVM XGBOOST KNN ANNForecastingAl et al.79E. coliColony countsANN RF SVMBerhilevych et al.80enterobacteriaColony countsANNBolinger et al.81SalmonellaGenomics (whole genome sequencing)RFGmeiner et al.82ListeriaGenomics (whole genome sequencing)RF SVM NNKaranth et al.83SalmonellaGenomics (whole genome sequencing)RF SVMLin et al.84SalmonellaMicroscopyCNNMunck et al.85SalmonellaGenomics (whole genome sequencing)RFStocker et al.86E. coliColony countsXGBOOST KNN SVM RFTanui et al.13SalmonellaGenomics (whole genome sequencing)RFWang et al.87Pathogenic microorganismsMetabolomicsANNWeller et al.88E. coli SalmonellaChromogenic assaySVM KNN XGBOOSTRisk factorsCox89SalmonellaPCR-based molecular identificationRF BNGolden et al.90ListeriaColony counts, PCRRFIm et al.91E. coliGenomics (whole genome sequencing)RFKaranth et al.92SalmonellaGenomics (whole genome sequencing)Elastic net regressionMurphy et al.93Pathogenic bacteriaColony countsRFPillai et al.94SalmonellaColony countsGAN XGBOOST RFRodríguez et al.95SalmonellaPCR-based molecular identificationRFStanosheck et al.96E. coliColony countsSVM RFToro et al.97SalmonellaPCR-based molecular identificationRFVangay et al.98ListeriaCAMP test, Gram stain, EcoRI ribotypingRF SVMWeller et al.14E. coliChromogenic assayNN RF SVMWeller et al.99ListeriaPCR-based molecular identificationRF SVMANN artificial neural network, BN Bayesian network, CNN convolutional neural network, DNN deep neural network, DT decision tree, ELM extreme learning machine, GA genetic algorithm, GAN generative adversarial network, KNN K-nearest neighbor, LDA linear discriminant analysis, NB Naïve Bayes, ResNet residual network, RNN recurrent neural network, RF random forest, SVM support vector machine, XGBOOST extreme gradient boosting. Papers using AI in the domain of microbiological hazards ANN artificial neural network, BN Bayesian network, CNN convolutional neural network, DNN deep neural network, DT decision tree, ELM extreme learning machine, GA genetic algorithm, GAN generative adversarial network, KNN K-nearest neighbor, LDA linear discriminant analysis, NB Naïve Bayes, ResNet residual network, RNN recurrent neural network, RF random forest, SVM support vector machine, XGBOOST extreme gradient boosting. Table 2Papers using AI in the domain of chemical hazardsImplementation contextPaperTarget chemicalData collection methodAI techniqueLaboratory testingAtaş et al.100MycotoxinsHyperspectral microscope imagingSVMChen et al.101MultipleMass Spectrometry-Based Analysis (LC-HRMS)ANN RF SVMCheng et al.102CadmiumHyperspectral imagingELMDarwish et al.103Foreign objectsNear-field microwave sensingSVMHe et al.15PesticidesOptical spectroscopy (fluorescence imaging)SVM ANNHu et al.104PesticidesHyperspectral imagingCNNJin et al.105Preservatives and heavy metalsOptical spectroscopy (Raman)DF CNNKalkan et al.106MycotoxinsOptical spectroscopy (multispectral)SVM ANNKim et al.107MycotoxinsOptical spectroscopy (Raman)SVMLu et al.108Veterinary drugsOptical spectroscopy (Raman)BP-NNMa et al.109NitrateOptical spectroscopy (ATR-FTIR)ELMMa et al.110MycotoxinsBioluminescenceSVM RF NNSmeesters et al.111AcrylamideOptical spectroscopyELM LDASmeesters et al.112AcrylamideOptical spectroscopyLDASun et al.113Sudan redElectrochemical sensorCNN ResNetTan et al.114HistamineOptical spectroscopy (Raman)SVMTang et al.115PesticidesElectrochemical sensorSVM XGBOOSTWang et al.116Sulfur dioxideImage analysisSVM BP-NN RFWang et al.117PesticidesOptical spectroscopy (surface enhanced Raman)SVMXu et al.118Mycotoxins, heavy metalsCombined dataLSTM XGBOOSTZhao et al.119PropachlorElectrochemical sensorXGBOOSTForecastingBranstad-Spates et al.120MycotoxinsELISAGradient BoostingHuang et al.16CadmiumSeveralSVM RF BP-NNLiu et al.121MycotoxinsMass Spectrometry-Based Analysis (LC-MS/MS)BNMa et al.122CadmiumSeveralRF ANNMarvin et al.123MultipleCombined dataBNMarzec-Schmidt et al.124MycotoxinsMass Spectrometry-Based Analysis (LC-MS/MS)SVM RFMateo et al.125MycotoxinsMass Spectrometry-Based Analysis (LC-MS/MS)XGBOOST RF NNMi et al.126CadmiumSeveralBP-NNPortz et al.127Veterinary drugsMass Spectrometry-Based Analysis (LC-MS/MS)DT NNWang et al.128MycotoxinsMass Spectrometry-Based Analysis (LC-MS/MS)RFXie et al.129MycotoxinsMetabolomicsRF XGBOOSTRisk factorsBouzembrak and Marvin18MultipleCombined dataBNCamardo Leggieri et al.17MycotoxinsMass Spectrometry-Based Analysis (LC-MS/MS)DNNHu et al.130Heavy metalsInductively Coupled Plasma Optical Emission SpectrometryRFPetrea et al.131Heavy metalsSeveralRFTang et al.132CadmiumGenomics (SNP)XGBOOST SVM KNN RFThao et al.133PesticidesVideo analysisCNNWang et al.128MycotoxinsMass Spectrometry-Based Analysis (LC-MS/MS)SVM XGBOOSTXiang et al.134PFASSeveralRF SVMANN artificial neural network, BN Bayesian network, CNN convolutional neural network, DNN deep neural network, DT decision tree, ELM extreme learning machine, GA genetic algorithm, GAN generative adversarial network, KNN K-nearest neighbor, LDA linear discriminant analysis, NB Naïve Bayes, ResNet residual network, RNN recurrent neural network, RF random forest, SVM support vector machine, XGBOOST extreme gradient boosting. Combined data refers to studies that used more than one data collection method. Papers using AI in the domain of chemical hazards ANN artificial neural network, BN Bayesian network, CNN convolutional neural network, DNN deep neural network, DT decision tree, ELM extreme learning machine, GA genetic algorithm, GAN generative adversarial network, KNN K-nearest neighbor, LDA linear discriminant analysis, NB Naïve Bayes, ResNet residual network, RNN recurrent neural network, RF random forest, SVM support vector machine, XGBOOST extreme gradient boosting. Combined data refers to studies that used more than one data collection method. Table 3Papers using AI in the domain of food fraudImplementation contextPaperData collection methodAI techniqueLaboratory testingAnantha Krishna et al.135Mass Spectrometry-Based Analysis (LC-MS/MS)SVM KNN XGBOOST ANNAsharindavida et al.136Optical spectroscopy (IR)SVM KNN RFChauvin et al.137Hyperspectral imagingGAChen et al.138Optical spectroscopy (FT-NIR)ResNet CNNChung et al.139Optical spectroscopy (IR)XGBOOSTde Oliveira et al.140Mass Spectrometry-Based Analysis (Direct-infusion MS)DTFeng et al.141Optical spectroscopy (Raman)SVM RF XGBOOSTGao et al.142Mass Spectrometry-Based Analysis (ICP-MS)RF SVMGundavarapu et al.143pH sensor, methane sensor etc.RFJo et al.20Optical spectroscopy (diffuse reflectance)AlexNet CNNKim et al.144Genomics (SNP)SVM kNNLi et al.145Combined dataRFLiu et al.19FT-NIRResNet CNNLyu et al.146Mass Spectrometry-Based Analysis (LC-MS/MS)SVM RFMohammadian et al.147E-noseSVM ANNNallan Chakravartula et al.148Optical spectroscopy (FT-NIR)CNNNanou et al.149Optical spectroscopy (laser breakdown)SVMPradana-López et al.150Image analysisResNetSaha et al.151Optical spectroscopy (NIR)CNNShin et al.152Optical spectroscopy (laser breakdown)SVM ANNWu et al.153Optical spectroscopy (fluorescence)KNN RF SVM CNNYang et al.154E-tongueCNNRisk factorsBuyuktepe et al.155Combined dataLIME SHAPLim et al.156GC-FIDDNNBorder inspectionKusuma et al.157Genomics (whole genome sequencing)SVMRegulatory efficiencyChang et al.21Combined dataRFRortais et al.28Combined dataLDAANN artificial neural network, BN Bayesian network, CNN convolutional neural network, DNN deep neural network, DT decision tree, ELM extreme learning machine, GA genetic algorithm, GAN generative adversarial network, KNN K-nearest neighbor, LDA linear discriminant analysis, NB Naïve Bayes, ResNet residual network, RNN recurrent neural network, RF random forest, SVM support vector machine, XGBOOST extreme gradient boosting. Combined data refers to studies that used more than one data collection method. Papers using AI in the domain of food fraud ANN artificial neural network, BN Bayesian network, CNN convolutional neural network, DNN deep neural network, DT decision tree, ELM extreme learning machine, GA genetic algorithm, GAN generative adversarial network, KNN K-nearest neighbor, LDA linear discriminant analysis, NB Naïve Bayes, ResNet residual network, RNN recurrent neural network, RF random forest, SVM support vector machine, XGBOOST extreme gradient boosting. Combined data refers to studies that used more than one data collection method. Table 4Papers using AI in the domain of food-borne illnessImplementation contextPaperData collection methodAI techniqueForecastingTeyhouee et al.158Combined dataSVMRisk factorsGoldberg et al.159Combined dataNLPMaharana et al.160Combined dataSVM BERTTao et al.161Combined dataBERTZhang et al.22Combined dataXGBOOSTZheng et al.162Combined dataRFRegulatory efficiencySadilek et al.163Combined dataSVMSadilek et al.164Combined datalog-linear maximum entropySadilek et al.23Combined datalog-linear maximum entropyBERT bidirectional encoder representations from transformers, NLP natural language processing, RF random forest, SVM support vector machine, XGBOOST extreme gradient boosting. Combined data refers to studies that used more than one data collection method. Papers using AI in the domain of food-borne illness BERT bidirectional encoder representations from transformers, NLP natural language processing, RF random forest, SVM support vector machine, XGBOOST extreme gradient boosting. Combined data refers to studies that used more than one data collection method. Table 5Papers using AI in food safety in a broader senseImplementation contextPaperData collection methodAI techniqueLaboratory testingAzman et al.165E-noseSVM KNNChen et al.166Not availableNot availableMadan et al.167Image analysisResnet CNN AlexnetXiao et al.168Hyperspectral microscope imagingActive learningZhong et al.24Optical spectroscopy (Raman)LSTMForecastingGeng et al.169Combined dataELMMakridis et al.170Combined dataRNNMakridis et al.171Text analysis (food recalls)DNN ResNetNogales et al.172SeveralCNN RFOldroyd et al.25Combined dataRF SVMZhang et al.173Combined dataELMZuo et al.174SeveralDenoising AutoencoderZuo et al.175Combined dataGNNRisk factorsLiu et al.176Combined dataBNSetiawan et al.177Image analysisDCNN ResNetZhu et al.178Combined dataBP-NN SVMBorder inspectionBisgin et al.179MicroscopyANN SVMWu et al.180Combined dataRF BNWu et al.26Combined dataXGBOOSTRegulatory efficiencyBarros et al.181Combined dataGAChen and Zhang182Text analysis (social media)SVM BERTKate et al.183Text analysis (news)SVMSilva et al.184Combined dataANN KNN DT NBSong et al.185Combined dataNB SVM XGBOOST CNN LSTM BERTSong et al.186Combined dataBERT LSTM SVMTalari et al.187Combined dataRF KNN SVMTonda et al.27Combined dataRFWang et al.188Combined dataELMZhang et al.189Combined dataLSTM SVMANN artificial neural network, BN Bayesian network, CNN convolutional neural network, DNN deep neural network, DT decision tree, ELM extreme learning machine, GA genetic algorithm, GAN generative adversarial network, KNN K-nearest neighbor, LDA linear discriminant analysis, NB Naïve Bayes, ResNet residual network, RNN recurrent neural network, RF random forest, SVM support vector machine, XGBOOST extreme gradient boosting. Combined data refers to studies that used more than one data collection method. Papers using AI in food safety in a broader sense ANN artificial neural network, BN Bayesian network, CNN convolutional neural network, DNN deep neural network, DT decision tree, ELM extreme learning machine, GA genetic algorithm, GAN generative adversarial network, KNN K-nearest neighbor, LDA linear discriminant analysis, NB Naïve Bayes, ResNet residual network, RNN recurrent neural network, RF random forest, SVM support vector machine, XGBOOST extreme gradient boosting. Combined data refers to studies that used more than one data collection method. Microbiological hazards Fifty-six out of the 161 reviewed papers (35%) used AI to examine microbiological hazards, including both studies on food samples and on microbiological isolates (Table 1). Out of these papers, 33 (59%) used AI to aid in the accuracy, speed, cost, or efficiency of laboratory testing. Data on which the AI aided for decision-making were typically obtained through various laboratory analyses, including but not limited to: genomics such as whole genome sequencing, microscopy, colony counts, PCR-based molecular identification, chromogenic assays, and optical spectroscopy such as Raman spectroscopy. AI aided in detection or quantification of bacteria such as Salmonella, E. Coli, and Listeria. For example, Gonçalves et al.12 explored the use of an “electronic nose” based on ionogel composites, paired with principal component analysis and various classifier algorithms, to distinguish Salmonella from other microorganisms. They reported classification accuracies ranging from 85% to 100%, depending on the incubation media used for the microorganisms. The study implied that this electronic nose method could be used as a cost-effective and straightforward screening or prioritization methods for Salmonella Spp. detection in food that can be followed by typing or confirmatory testing. They emphasized its potential to complement existing diagnostic techniques by offering faster results, lowering costs, and minimizing manual handling steps. How AI can be used to forecast future microbiological risk was examined in 11 out of 56 papers (20%). For instance, Tanui et al.13 used machine learning to predict disease outcomes from untagged Salmonella genome sequences isolated from ground chicken. A random forest model, trained on data from an extensive literature review, could forecast 87% of disease endpoints. The model identified genes linked to virulence and specific disease outcomes, which could be used as features in future predictive models. The authors state that the approach could be used to identify pathogenicity in specific genes, predict antibiotic resistance, and eventually contribute to the formal risk assessment of the targeted pathogens in food. Additionally, 12 out of 56 papers (21%) explored the use of AI in investigating risk factors and the etiology of microbiological hazards. Weller et al.14 compared different machine learning algorithms and data sources to predict E. coli levels in agricultural water. They found that physicochemical water quality and weather features combined with random forest algorithms achieved the highest accuracy. The authors state that deploying these models in the field could reduce the reliance on costly and time-consuming E. coli testing. Most of the papers (42 out of 56, 75%) used classical machine learning to study microbiological hazards in food safety. Deep learning was used in 14 papers (25%). Chemical hazards A total of 40 out of 161 (25%) papers examined the use of AI to detect or quantify chemical hazards (Table 2). Of these 40 papers, 20 (50%) focused on AI for improving the accuracy, speed, cost-effectiveness, or overall efficiency of laboratory testing for chemical hazards. The data were acquired with methods such as liquid chromatography coupled with tandem mass spectrometry (LC-MS/MS), Raman spectroscopy, hyperspectral imaging, and fluorescence imaging. Residues of pesticides, veterinary drugs, and mycotoxins such as aflatoxins were often the target hazards, as was the presence of heavy metals, in for example, rice and fish. A growing trend was the development of on-site non-destructive, cost-effective instrumentation paired with machine learning techniques to match the results of more expensive, labor intensive lab-based analytical chemical methods. For instance, He et al.15 developed a sensing method to identify residues of two types of pesticides (thiabendazole and fuberidazole) in red wine using fluorescence spectroscopy combined with a support vector machine and an artificial neural network. The authors reported that they were able to quantify pesticide levels with high accuracy without the need for the usual complex pretreatment procedures. AI was used to forecast risk of chemical contamination of food in 11 out of 40 papers (28%). For example, Huang et al.16 used multiple machine learning methods (support vector machine, random forest, neural network) to forecast the concentration of cadmium in rice. They found that the support vector machine showed the best performance to predict cadmium concentration in rice grain on a regional level. Additionally, Camardo Leggieri et al.17 used a deep neural network to examine how weather conditions and cropping systems affect the growth of mycotoxin-producing fungi and whether the maximum limit set for particular mycotoxin contamination would be exceeded based on a thirteen-year dataset acquired in Northern Italy. Eight out of 40 papers (20%) used AI to study potential factors and sources of chemical hazards in food, particularly in the context of international trade. For instance, Bouzembrak and Marvin18 used a Bayesian Network with data from the EU Rapid Alert System for Food and Feed (RASFF) for fruits and vegetables combined with climate, economic, and agricultural data. They identified several factors that contributed the most to the chemical hazard prediction model such as product category and country of origin. Similar to the papers on microbiological hazards, most of the papers (30 out of 40, 75%) used classical machine learning to study chemical contaminants in food safety. Deep learning was used in 10 papers (25%). Food fraud/authenticity Twenty-seven of 161 reviewed papers (17%) investigated food fraud or food authenticity using AI. A large majority of these papers (22 out of 27, 81%) applied analytical techniques to obtain data, including Near-infrared spectroscopy, fluorescence spectroscopy, but also e-noses and e-tongues. For instance, Liu et al.19 used Fourier transform near-infrared (FT-NIR) spectroscopy to analyze spectral differences between the stipe and cap of various porcini species. They used support vector machines, partial least-squares discriminant analysis, and residual neural networks to identify porcini species, of which the latter showed the highest accuracy. The authors note that FT-NIR spectroscopy combined with machine learning is an accurate, fast, and non-destructive method, which makes it a promising tool in food safety applications. Furthermore, Jo et al.20 showed that beef alteration could be identified using deep learning-aided spectroscopy. Two out of 27 papers (7.4%) examined the use of AI to help identify relevant factors that potentially indicate fraud; and one paper (3.7%) used AI to aid to detect possible food fraud at border inspection. Two papers (7.4%) examined how to use AI to screen to detect food fraud at a regulatory level. For example, Chang et al.21 developed an automated alarm system for edible oil manufacturing using nearly 100 million electronic invoices. They used k-nearest neighbor (KNN), support vector machine (SVM), neural network, logistic regression, and random forest (RF), of which the latter had to highest accuracy to identify relevant invoices and flag problematic manufacturers. The authors pointed out the potential of electronic invoice data for traceability and highlighted its value in proactively preventing food safety crises by identifying suspicious activities before they escalate. In the research domain of food fraud and authentication, deep learning was used in 11 out of 27 papers (41%), while most of the papers (16 out of 27, 59%) again used classical machine learning. Foodborne disease outbreak investigations and surveillance Nine out of 161 papers (6%) used AI to assess the occurrence of foodborne disease outbreaks. All studies used AI with combined data from different sources. The studies focused on determining risk factors for the outbreaks of foodborne diseases (5 out of 9 papers, 56%), assessing regulatory efficiency like inspection prioritization (3 out of 9 papers, 33%), or forecasting future outbreaks (1 paper, 11%). For example, Zhang et al.22 used retrospective foodborne disease outbreak monitoring data that contained both confirmed and suspected outbreaks. Several machine learning algorithms (support vector machine, logistic regression, decision tree, random forest, gradient boosting decision tree, adaptive boosting, and extreme gradient boosting) were trained on the confirmed outbreaks and used to classify the expected outbreaks into real or false. In parallel, medical staff confirmed or excluded the suspected outbreaks, and the results were compared. The authors concluded that the extreme gradient boosting model had the best performance. Shapley values were used as explainable AI technique to determine which predictor variables contributed most to the likelihood of individuals with the same exposures becoming ill. As an example in the regulatory context, Sadilek et al.23 used anonymized search and location data from smartphones to identify potential sources of foodborne diseases or outbreaks in real-time. Their method was to use internet searches related to various indications for possible foodborne diseases, such as visits to pages like the CDC website or Wikipedia articles on the topic, using a log-linear maximum entropy model. They then analyzed anonymized location histories to determine which restaurants had been visited by users who later made these queries. For each restaurant, the model estimated the proportion of visitors who exhibited symptoms of illness based on their search behavior. This approach enhanced the identification of possibly problematic venues by over three times compared to traditional methods, providing health departments with a tool to more quickly identify and investigate potential outbreak locations. Most of the studies (7 out of 9, 88%) used classical machine learning to study foodborne disease outbreak investigations and surveillance. Deep learning was used in 2 out of 9 studies (12%). General food safety Twenty-nine out of 161 papers (18%) examined the use of AI on broader food safety issues. For example, they were studies examining early warning signals for both microbiological and chemical contaminants, food spoilage or pests, or focusing on inspection reports or food recall. Five of these 29 papers (17%) used AI to aid in the identification of food spoilage and the detection of multiple hazards simultaneously in laboratory testing. As an example, Zhong et al.24 compared classical machine learning techniques such as support vector machine and k-nearest neighbor with a long short-term memory (LSTM) to predict salmon freshness based on Raman spectroscopy performed at different storage temperatures. They found that Raman spectroscopy combined with the LSTM was able to accurately predict salmon freshness in a non-destructive way. Eight of the 29 papers (28%) used AI to forecast various food safety issues. For instance, Oldroyd et al.25 used a synthetic minority oversampling technique with a random forest model to forecast compliance in new and uninspected food businesses, using neighborhood characteristics such as socio-demographics, business types, and neighborhood information. Given the scarcity of local authority resources, the authors state that this may be a promising method to prioritize inspections and make monitoring more efficient. The use of AI in identifying factors that contribute to raising the risks of various food safety hazards was studied in three out of 29 papers (10%). The potential of AI to aid border inspections for food and feed was also studied in three out of 29 papers (10%). An example of the latter is Wu et al.26, who developed a machine learning-based system to decide whether quality control sampling is necessary for imported food at border checkpoints. They introduced an ensemble learning prediction model and compared its performance to an older model and random sampling methods. The ensemble model demonstrated significantly higher efficiency, flagging cases for inspection where non-compliance rates were up to three times greater than those identified through random sampling. Wu et al. state that this enables swift adaptation to global changes, enhancing the effectiveness of border management in food safety. Lastly, 10 out of 29 papers (34%) examined the use of AI for regulatory efficiency of food safety. For instance, Tonda et al.27 used questionnaires collected across seven countries to examine food safety concerns. Using a random forest, they found that consumers with higher awareness of environmental impact and caloric contents of food were also the most concerned about food safety, hinting at the role of food education as an explanatory factor. Most of the studies (17 out of 29, 59%) used classical machine learning. Deep learning was used in 12 out of 29 papers (41%). Discussion Key trends in AI in food safety research Our systematic review showed that AI has been actively used in food safety across various domains in different implementation contexts, using a wide variety of data types. The use of AI in food safety is growing rapidly, increasing from a single publication in 2012 to 46 in 2023 (Fig. 2b). Although most studies identified in our review still used classical machine learning methods, the use of deep learning is steadily increasing from 2 out of 9 papers in 2019 (22%) to 20 out of 46 papers in 2023 (43%) (Fig. 2b). The ratio of deep learning papers differed between the research domains: 14 out of 56 papers examining microbiological hazards (25%), 11 out of 27 papers (41%) in the food fraud domain, and 12 out of 29 papers (41%) in food safety in a broader sense. The higher score of the last two domains could be due to the type of data used, namely text sources like website texts, food safety regulatory reports, including food recalls, news websites, and social media posts. Progress in computational capabilities, availability of large datasets, and advancements in AI methodologies are expected to drive the further adoption of deep learning in food safety research. Limitations In this review, we did not systematically compare model performance across studies, as many food safety datasets contain few positive cases, resulting in strong class imbalance and sparse coverage of predictor combinations, which makes reported performance metrics (e.g., accuracy) difficult to interpret and compare between studies. A second limitation of this review is that the literature search was restricted to Scopus, which was selected for its broad interdisciplinary coverage and compatibility with ASReview export. While this choice may have resulted in a small number of relevant papers being missed, we expect the impact to be limited given the substantial overlap in indexing with other major databases. Lastly, another limitation of our systematic review was the limited finding of papers examining the application of AI in food safety within food production, processing, and manufacturing contexts and their businesses. This particular gap may be attributed to the fact that the study’s focus was mainly on peer-reviewed scientific literature where the sector may not use scientific literature as the main information-exchange platform, but rather white papers. Additionally, proprietary or commercial concerns may discourage companies from publishing their AI-related findings, which limits the visibility of such research. Commercial applications of AI in food safety Although the focus of this review was scientific literature, a growing number of commercial platforms apply data-driven and AI-enabled approaches to support food safety management and regulatory compliance in practice. Many commercially available tools align primarily support regulatory efficiency, for example, through compliance workflows and regulatory intelligence (e.g., SGS Digicomply, IONI, iComplai). Other platforms primarily support risk analysis, for example, via supplier, specification, and documentation management across supply chains (e.g., TraceGains, Trustwell, Cloudtheapp, iFoodDS). In addition, predictive analytics providers such as Agroknow and Cropin are positioned around risk monitoring and decision support that aligns with risk forecasting. While these commercially available applications demonstrate increasing uptake of AI-enabled decision support in operational workflows, underlying model details and independent evaluations are often not publicly available, limiting direct comparison with peer-reviewed approaches. Commercial translation appears to differ between research domains. Comparatively few commercial tools focus on microbiological hazards, suggesting that AI applications in this area are still relatively far from widespread practical implementation. Conversely, within the chemical hazards domain, mycotoxin forecasting provides a clear example of commercial translation, as contamination risk is closely linked to climate and weather conditions during flowering and crop development. Commercial tools integrate environmental and crop-related information to generate actionable outputs for risk forecasting and priority setting for inspection and targeted testing, and can also support procurement-related risk analysis (e.g., DSM-Firmenich Mycotoxin Prediction Service; Adisseo MycoMan). Commercial uptake seems stronger in domains where AI can directly improve operational decision-making and resource allocation, such as spoilage detection and shelf-life monitoring using AI-enabled sensing (e.g., electronic nose approaches). Several vendors provide e-nose and digital olfaction tools positioned for quality control and screening workflows (e.g., Alpha MOS HERACLES, AIRSENSE Analytics, Aryballe NeOse). Finally, prediction and monitoring of contamination signals and outbreaks across the food chain are increasingly operationalized through systems that support early detection and situational awareness; for example, MedISys illustrates large-scale horizon scanning using automated processing of online information sources, aligning with risk forecasting and decision support for authorities28,29. Overall, commercial uptake of AI in food safety appears strongest in implementation contexts related to risk analysis, risk forecasting, and regulatory efficiency, with some tools also supporting priority setting for inspection and border testing, whereas fewer widely adopted applications focus directly on aid for laboratory testing and analysis for routine pathogen testing under controlled laboratory conditions. Addressing unique challenges for AI in food safety Data limitations There are a number of factors that may impact the use of AI in the area of food safety. One factor is the availability and quality of the data that can be used. Many food safety related data sets are often incomplete and possibly imbalanced. For example, the majority of analytical results typically show a low concentration of food safety hazard contamination, with only a small proportion of results exceeding the maximum limits30. It is important to note that this imbalance can impede the performance of classification or forecasting models. Specifically, these models may become very good at recognizing cases where contamination is low, but struggle to accurately identify cases where contamination levels are high and potentially unsafe. Since detecting these higher-risk cases is crucial in food safety research, special techniques are needed to improve model accuracy and reliability. There are several solutions to mitigate this problem. Mitigation strategies include weighing of samples to decrease the effect of the extreme class imbalance, such as weighted Bayesian networks30 or synthetic minority oversampling technique (SMOTE25), although there is also evidence that using such techniques might cause miscalibration to prediction models31. Other mitigation strategies can include incorporating the imbalance in the loss function or using synthetic data as augmentations. An example of such synthetic data generation is using a Generative Adversarial Network (GAN). A GAN can generate realistic data for the underrepresented class, for example, a sample exceeding the legal limit of contamination. GANs can be trained on existing data to produce synthetic samples which mimic the distribution and features of the minority class, making the dataset more representative32,33. Data governance and accessibility In addition to imbalanced data sets, missing some critical metadata can be an issue. Metadata is data providing information about the data, such as the origin of samples, sampling locations, and sampling conditions. This information is crucial for studies that aim to trace contaminations, examine risk factors, or forecast future risks. Standardized data collection following the FAIR principles (Findable, Accessible, Interoperable, and Reusable) would increase traceability and can help the field move forward. AI is often developed on already existing (e.g., retrospective) data. Collecting data with the development of AI in mind can increase effectiveness of AI models and can also form the solution to some of the beforementioned issues with metadata and traceability. Furthermore, a lot of food safety data is not openly accessible. Since many AI models rely on large high-quality datasets for training and validation, the availability of data becomes key, and thus promotion of sharing data across institutions ensures access is necessary to have AI to use diverse and comprehensive information. Collaborative data sharing is also a solution to duplication of efforts, it speeds up innovation, and it fosters global standardization in AI-driven food safety practices. Federated learning However, it is also true that openly sharing data is not a straightforward issue. Food safety data are often sensitive because of privacy information (e.g., names or addresses of people with a foodborne disease), because of commercial interests (e.g., proprietary data of a food manufacturer), or because they are obtained in monitoring by competent authorities. One of the most promising solutions to still use this data without unveiling its sensitive characteristics is federated learning34. Rather than centralizing all data for algorithm optimization, the model interacts with individual “data stations,” similar to a train stopping at various railroad stations. This decentralized approach is gaining popularity in food safety, where the sensitive nature of the data makes traditional centralization challenging35. AI Ethics The use of AI in food safety can also raise ethical considerations, particularly because AI systems may influence decisions with direct implications for public health and regulatory actions36. In response to concerns around fairness, accountability, and transparency, AI governance is increasingly being formalized through legislative and policy initiatives worldwide, such as the European Union’s AI Act37. Ethical governance frameworks aim to guide responsible AI development and deployment by addressing risks related to privacy, security, transparency, accountability, and potential misuse, thereby supporting public trust38. A key ethical challenge is bias: AI systems can reproduce or amplify biases present in historical datasets, which may result in systematically unfair outcomes if not identified and mitigated39. Explainable AI In addition to ethical and governance considerations, ensuring the reliability and trustworthiness of AI models remains a key challenge for applying AI in food safety research. Deep learning models, particularly neural networks, are often viewed as “black boxes” due to their intricate and layered structure, making it challenging to fully understand how they make decisions, even when each layer is examined40. There is growing concern that these models may harbor hidden biases that remain undetected, potentially leading to harmful outcomes. A solution for more trustworthy AI is the field of explainable artificial intelligence (XAI)41–43. XAI aims to increase the transparency of AI by striving to examine what parameters drive decision-making. Discovering such parameters can be crucial for unravelling the sources of bias and ensuring that only relevant information is used to draw the conclusion. This becomes especially important in high-stakes decision-making, such as food safety, where regulations like the European Union’s AI Act and General Data Protection Regulation (GDPR, Article 15) grant individuals the right to understand how decisions are made. XAI has already found widespread use in sectors like medicine40 and is expected to become increasingly essential in food safety to ensure transparency and accountability in AI-driven decisions. Food systems approach Food systems are complex with multiple stakeholders that all influence the system, and therefore also all influence the safety of the food44. AI-based assessment of food safety hazards can therefore benefit from using a holistic or systems approach45, using data from multiple sources and even multiple modalities (e.g., genomic data and image data). Fusing different data types can be challenging, and multiple approaches exists, such as fusing data before AI modelling (i.e., early fusion) or training multiple data-specific AI models and combining their output (i.e., late fusion46). Conclusion In conclusion, AI is increasingly used in food safety across various domains and implementation contexts. Despite several unique challenges in the use of AI in food safety, such as the quality and availability of (meta)data, issues related to data privacy, and the interpretability of AI models, the potential benefits are vast. As AI techniques such as federated learning and explainable AI continue to evolve, they offer promising solutions to address these challenges and further enhance the reliability and applicability of AI in food safety research.