Testing and Analysis

By Angela Anandappa, Ph.D., Founding Executive Director and President, Alliance for Advanced Sanitation and Niam Abeysiriwardena, M.S., Research Engineer, Harvard Medical School and Massachusetts General Hospital

Big Data, AI, and the Coming Philosophical Challenges with Food Safety

Improvements on the horizon will make AI systems more useful for food safety applications

Holographic AI logo glowing on a vibrant, abstract circuit board background.

Image credit: Vertigo3d/E+ via Getty Images

SCROLL DOWN

Are you using your data wisely?

You hope you've done a good job. You probably think you've done what is necessary to protect your company's data, or even your personal data, in a secure manner. Storing confidential data in secure locations (away from inquisitive eyes), backing up important data, adding password protection, installing firewalls, encrypting data, and hiding important information is the least you must do. But let's talk about something else: how we think about our data, and what we can do to apply strategy, critical thinking, and wisdom to how we generate, store, use, protect, and share our data.

Wrestling with which data to gather, how to organize and store the data, and how to make decisions with the data are questions the food industry has been considering for some time now. Given the sheer size of the food industry, stakeholders' ways of working with data are just as diverse as the companies and types of food around the world. From complex systems for all aspects of the business to nothing more than a spreadsheet, you can be a minimalist user of data or have computing systems worth some admiration and envy.

The number and types of data used in running a successful business are extraordinarily varied. Consider handwashing data and water quality metrics, or pathogen monitoring data and surveillance systems that identify suppliers likely to commit fraud. Regulatory pressure can be a motivator for organizations to test certain analytical methods and perform data correlations, or it can compel them to share information with external stakeholders. All these types of data provide as complete or incomplete a picture of the operation as the systems in which they are collected and analyzed. In the short term and within the upcoming years, your data and decision-making will determine the success of the business as a whole and, in some cases, could lead to the shuttering of some companies. Does that seem extreme? Sure, but let's discuss why, so that you can take measures that better protect your company.

By May 2025, the number of companies providing artificial intelligence (AI)-based services or developing AI was upwards of 70,000 globally.1 Those pitching services to food and agriculture businesses range in their offerings. Some specifically serve food manufacturers' quality and safety needs by offering to organize and provide greater transparency through completing simple tasks like scanning, collating, and organizing business documents. Many others hope to be the "crystal ball" through which users may see more deeply into the supply chain and receive alerts on various types of issues, such as crime and fraud regulatory actions, climate effects on agriculture, training logs, and systems for maintaining and prompting maintenance activities. These new solutions could assist organizations in many ways, and while this is an exciting prospect, this article aims to address a few key points to consider when vetting solutions.

Transparency

AI companies and AI service providers should be evaluated on how transparent they are in describing how their product works. What appears to be magic in the age of AI could be fiction dressed up as information and direction. Generative AI is, ultimately, many mathematical calculations that filter one set of information and spit out answers the system "thinks" you want. When these AI systems malfunction, users do not get a chance to see what went wrong and have no way of controlling the algorithms themselves.

For example, sensor data for controlling temperatures within a facility could be fed into an AI system that helps reduce energy costs. How the system obtains the sensor data, how the sensor itself functions, and how the AI makes decisions ideally would be known by the personnel managing the system. This way, faulty sensor data could be diagnosed, or temperature anomalies like an overheated machine located close to a sensor could be identified in a timely manner. With decisions made by an AI that is neither understood by, nor owned by, your organization, identifying and troubleshooting issues becomes that much harder. This is an overly simple example akin to using a thermostat, but integration of high-tech sensors in high-value production areas requires a higher investment in knowledge of how that system works to ensure that it operates properly and smoothly.

Consider systems that provide surveillance data on market trends, economic indicators, chemical samples, analysis of reports, results of microbiological samples, contamination risks, or even employees. The bigger the decision that is made using these AI systems, the bigger the impact—good or bad—on the lives of customers, employees, or the business.

Cybersecurity

Two words: Multi-disciplinary team. Food safety professionals must manage a food defense plan that should be audited by third parties. However, in a world where more complexity is introduced through the use of SaaS (Software as a Service) products like computational systems, AI services, and various subscription services, company employees and auditors need to become savvier at knowing which security questions to ask. They must also have more in-depth knowledge to troubleshoot how these systems can fail. IBM notes a growing concern in cybersecurity related to manufacturing operations and estimates the cost of a data breach at $4.9 million USD per attack.2 Cyberattacks as simple as phishing emails and ransomware, or as complex as those involving data leaks, attacks on utilities and power supplies, or the transmission of information outside of defined boundaries, can cause havoc if not actively prevented.

A recent study revealed that most of the 32,352 employees in the 47 countries surveyed reported using some form of AI in their daily work to "increase efficiency" (67 percent), "gain access to information" (61 percent), "be creative and innovative" (59 percent), or to "increase the quality of their work" (58 percent).3 A concerning number of these employees were using AI in highly risky ways by uploading sensitive information into public AI systems and having AI check them, and those same individuals further went on to cover up that AI was being used. Using AI to do everyday tasks like writing, editing, conceptualizing, developing marketing and sales plans, and devising strategies is common. An additional challenge is that over 42 percent of employees opted to use free public AI tools like ChatGPT, despite having access to paid, firewalled alternatives provided by their employers.

FIGURE 1. Cut Fresh has a number of best practices, auditing, and certification efforts in place to ensure the highest standards of food safety, food quality, and sanitation (Image courtesy of Cut Fresh LLC / Anna Patsakham)

"Even FDA has announced its own internal LLM for processing scientific studies as part of its efficiency initiative."
Monochrome photography, Parallel, Black, Black-and-white, Line, White

LLMs (Large Language Models)

Widespread availability of public LLMs has provided everyone access to using a form of AI in their daily lives. Many managers are now encouraging employees to explore LLMs in response to the surge of available technology. Some companies have taken the leap to creating their own firewall-protected LLMs for use within their companies. Even the Food and Drug Administration (FDA) has announced its own internal LLM for processing scientific studies as part of its efficiency initiative.

AI encompasses a lot of different things, but the solutions most people are referring to in AI involve Large Language Models (LLMs), such as ChatGPT (OpenAI), CoPilot (Microsoft), Gemini (Google), or Grok (xAI). LLMs came to the forefront of the discussion in late 2022, and since then people have found all sorts of creative uses for them. As part of their development, they incorporate something from nearly every discipline. Their surprising utility and versatility come from mimicking the way humans think "out loud." AI itself is nothing new; it really refers to any attempt to make a machine capable of something that we otherwise expect from humans, from simply searching for information to making complicated decisions.

Machine Learning is the way we create AI models, by getting the machine to learn from some dataset, as opposed to telling it how to do the task in painstaking detail. AI models are the result—algorithms that have been trained with a particular set of data. LLMs use large amounts of text as their dataset, including books, articles, and even conversations that have occurred on the internet, and so those LLMs are able to predict parts of human language. This is where the power of LLMs comes from: being able to work with human language and having already learned from human language resources discussing many topics, so that when it predicts what human language looks like, it also often predicts true sentences and valid lines of reasoning.

LLMs are being used everywhere that language has already been used. People are already promoting solutions for personalized marketing, writing e-mails, even simulating a conversation. In food, language is important for highly critical things: creating food safety plans, writing procedures, developing contingency plans, and communicating with suppliers and customers, to name a few. The solutions that are easiest to create using popularly available technologies are also those areas we do not want to get wrong.

One thing to keep in mind is that LLMs are not primarily trained to be truthful or accurate; they learn those things almost by mistake from the books, articles, and internet chats they use as input. The companies that develop the LLMs will nudge them toward truthfulness and accuracy, but they are not dedicated to these principles, and they are almost certainly lacking when it comes to technical food safety knowledge. People are trying to improve the accuracy by letting the LLMs look at references for these critical facts, but that does not change the lack of reasoning behind the choices LLMs would make in writing these critical documents. If we ask the LLM why it made such a choice, it can only make up an answer.

Validation

When we validate, we determine that a process accomplishes what we need it to do. AI complicates this by adding uncertainty. Instead of being certain that we can achieve a 5-log reduction, for AI we can only say that it will give the correct answer 95 percent of the time. When we perform verification, we would ordinarily be checking whether we are reaching the right temperatures, but when it comes to AI, the questions we need to ask for verification become much more complicated, or may not even make sense. For LLMs, we do not know what kinds of input make them "break," so we cannot check in advance to ensure that our AI systems will work as expected.

Although LLMs are the focus of hype right now, they have not yet made a huge change in automation. Even now, over 99 percent of what has been automated in total has not been automated using LLMs, but instead with older methods developed over the past decades and centuries that we find more familiar and more explainable. When the discussion of big data being used for food safety comes up, we are far from scratching the surface on what can make significant change in manufacturing.

We can test and validate models for specific applications, like computer vision models for identifying problematic produce, models that can improve detection of physical hazards, and even models that can predict changes in the microbiome. What we cannot validate are LLMs that we substitute in place of human reasoning. We can only hold accountable the people that implement and use these models, as we cannot actually ask them about their decision-making after the fact—only ask them for a justification in the process of deciding. This is a fundamental limitation in the way these models work.

Manufacturing automation and agricultural automation have already been well-developed in much of the global north and have already made their impact on manufacturing. Current challenges in labor and opportunities for robotics and automation reside in the underbelly of the food industry where difficult, dangerous, and unappealing work in unpleasant environments is still vital to the food supply.

"Better ways are needed to seamlessly communicate data between companies within the same supply chain and enable them to speak the same language—not only to ensure the safety of our food supply, but also to ensure efficiency."
Monochrome photography, Parallel, Black, Black-and-white, Line, White

Smoke and Mirrors

The psychology behind using an ordinary reference work for data versus the output of a Generative AI that has performed a search is different. The more conversational nature of the Generative AI can lead to a false sense of security in the results, and in the long term makes the human operator less vigilant in detecting mistakes. This can even lead to a highly skilled individual's accuracy suffering when relying on Generative AI. From the perspective of management, the flaws in such an approach can be hidden for a long time. It may appear that an expert can be substituted with someone less knowledgeable who is aided by AI, but the consequences will be noticed only later, when it is too late.

We have been living in the age of data, but the past year has opened the floodgates for data-related technologies to spawn new tools at a staggeringly accelerated rate. There is no easy way to critically review many tools in tandem without having in-depth knowledge of the technologies themselves, communications protocols, security systems, fitness for purpose, and subtle nuances in applications that might be known to some, but not all, members of a company.

Better ways are needed to seamlessly communicate data between companies within the same supply chain and enable them to speak the same language—not only to ensure the safety of our food supply, but also to ensure efficiency. Due to the fear of data abuse and leaks, many companies are still hesitant to share data for liability reasons. As we move forward with adopting more tools, sharing pertinent data with allies, customers, and suppliers will set the benchmark for efficiency and interoperability. This is vital for advancing digital strategies for better traceability, implementing quicker recalls, investigating product defects, and reducing foodborne illnesses. Beyond this, advanced digital strategies can aid in understanding customer preferences and patterns and linking these to specific challenges at the retail or consumer end. This can provide a wealth of opportunities to streamline supply chains and sell more products that consumers want.

Reputational risks can be managed with better data-sharing, when done right. Opening the door to shared lab results or shared sampling plans is a carefully growing area of liability protection, while sharing the responsibility for testing puts human illness prevention and efficiency as priorities. Companies are working to understand better ways to move forward with these approaches.

Additionally, large sources of consumer data enable advanced techniques for product developers to tailor colors, dyes, and small-scale manufacturing innovations to add tremendous value to high-end products for nutrition and medical purposes. This is a burgeoning, different business model that poses its own challenges. Distributed manufacturing and more localized supply chains, while desirable for marketing and claims purposes, offer new models for considering food safety challenges in a new light. One of these challenges applies more broadly across industry—the opportunity to use robots in much of assembly, customization, and delivery to the customer. While kiosk-based robotic devices can look like fancy, all-inclusive, clean-in-place operations, they will inevitably require complete teardown and cleaning of all parts, including those that perform the cleaning function. No robotic systems have yet been made to address the real challenges associated with cleaning—a chore no one wants to do.

“Each step in the 'rings of defense' must be used consciously and with proper consideration, as it may be obvious what caused the noncompliance, and its resolution may be straightforward, simple, and not costly.”
Tints and shades, Monochrome photography, Black, Black-and-white, Line, Style
Monochrome photography, Parallel, Black, Black-and-white, Line, White

Access is Power

For the past several decades, AI was used in the background, with limited access beyond tasks such as mapping services or editing photos on smart phones. However, AI has been a concept in development for much of our lives. The desire to make a machine that can "think" was conceptualized prior to the 1950s. Both past and current iterations of AI technologies rely on mathematical switches and signals that turn on and off. At high speeds, billions of switches work to show us pictures, letters, words, and sounds, and attempt to communicate with us as though the system has life. The more closely the system mimics human behavior, language, and thought, the easier it is for us to believe it and do less thinking and questioning. We trust it more than we should, and AI systems are getting better at convincing us that they should be trusted. The complexity of human intelligence is far beyond what any current LLM is capable of, and in no way does any current form of AI match up with human intelligence.

Although flawed, AI is continually and rapidly being improved, with errors being fixed. Improvements on the horizon will make AI systems more useful, with fewer flaws. Our lives can be made easier by technologies that use AI, and access to AI technologies will put more power in more hands. The closer we are to experiencing more functional AI, the more we must remember that to be human requires us to be responsible in our decisions. When we make the food supply safer, we do so for our customers and consumers.

Safer food is a responsibility, and we are called to protect the food supply for all of us. Even now, with more technology power available, that power must be used wisely and with good intention. We cannot use future improvements to justify present adoption of immature technologies, and certainly not when lives are at stake.

References

  1. Shrivastava, R, Ed. "AI 50 List, 2025." Forbes. April 10, 2025. https://www.forbes.com/lists/ai50/.
  2. IBM. "Cost of a Data Breach Report 2024." https://www.ibm.com/reports/data-breach.
  3. Gillespie, N., S. Lockey, T. Ward, A. Macdade, and G. Hassed. "Trust, attitudes and use of artificial intelligence: A global study 2025." The University of Melbourne and KPMG. April 28, 2025. https://figshare.unimelb.edu.au/articles/report/Trust_attitudes_and_use_of_artificial_intelligence_A_global_study_2025/28822919?file=54013232.
  4. Gillespie, N. and S. Lockey. "Major survey finds most people use AI regularly at work—but almost half admit to doing so inappropriately." The Conversation. April 28, 2025. https://theconversation.com/major-survey-finds-most-people-use-ai-regularly-at-work-but-almost-half-admit-to-doing-so-inappropriately-255405.

Angela Anandappa, Ph.D. is the Founding Executive Director and CEO of the Alliance for Advanced Sanitation. She is a food safety expert, food microbiologist, and food industry leader with over 25 years of experience in food safety. She holds a Ph.D. in Food Safety, an M.S. degree in Food Microbiology, and a B.S. degree in Biology.

Niam Abeysiriwardena, M.S. is a Machine Learning Engineer working at Massachusetts General Hospital. He holds an M.S. degree in Artificial Intelligence and Machine Learning and a B.A. degree in Neuroscience and Computer Science.

AUGUST/SEPTEMBER 2025

Font, Line, Text