The Limits of Prediction

This interview is part of our extended Q&A series exploring AI and food safety.

AI is shifting food safety toward a more predictive approach, but that shift is still incomplete. While these tools can flag signals that suggest risk is increasing, they are not yet reliable enough to function as stand-alone early warning systems. In this Q&A, Cornell University Professor Martin Wiedmann and PhD candidate YeonJin Jung explain where predictive modeling adds value and why its limitations still matter.

How close are we to predicting contamination risk before positives occur using AI tools and/or AI-driven predictive modeling?

We are moving from answering a reactive question—“Do we have contamination?”—to a more proactive one: “Given the current condition, what is the probability that contamination risk is increasing?”

That is an important shift, but we have not reached a point where contamination events can be anticipated with full confidence. Currently, AI tools and predictive models can help flag elevated risk before a lab–positive result, but their performance depends heavily on the quantity, quality, and relevance of the underlying data, as well as whether the model has been validated for a specific process or environment.

Early risk detection is more feasible in settings with relatively large, structured datasets, for example, raw poultry contamination with Salmonella. Prediction becomes more difficult when contamination events are rare, because rare-event data are inherently harder to use for model training and validation.

Overall, AI is bringing the industry closer to estimating when contamination risk may be elevated, but current systems are not yet reliable enough to function as stand-alone early warning tools. Broader adoption will likely depend on better-curated datasets and careful integration with existing food safety programs, where AI strengthens decision-making rather than replaces it.

We’re moving from asking, ‘Do we have contamination?’ to, ‘What is the probability that risk is increasing?’

Martin Wiedmann

Cornell University

What limitations exist in pathogen prediction models today, and how can AI still benefit food processors in predicting risk?

A major limitation of current pathogen prediction models is that contamination events are often rare. In food safety, a model that appears highly accurate can still fail if it misses an impactful event.

These models can also struggle with unusual events that are not well represented in historical data, such as infrastructure failures or atypical traffic patterns. That means evaluating false negatives, sensitivity, and performance on rare events is just as important as overall model accuracy.

Another limitation is that many models can flag elevated risk but may not be able to explain why. As a result, food safety professionals still need to evaluate the signal, investigate the context, and determine appropriate interventions.

Despite these limitations, AI can still provide meaningful value. It can improve risk prioritization, support traceability efforts, and accelerate root-cause analysis. Even when a model cannot predict every contamination event, it can help identify patterns associated with increasing risk and focus attention on high-priority areas—reducing the time needed to investigate likely contamination sources. Its value lies in strengthening decision-making within existing food safety systems.

How should companies validate an algorithm that is making or influencing food safety decisions?

Companies should validate any algorithm for its specific context of use. This includes confirming that the training and test data are relevant to the actual facility, product, and process, and that the data are complete, accurate, and governed over time.

Model performance should be evaluated not only against historical cases, but also under realistic failure scenarios and atypical operating conditions, because those are often the most important situations in food safety decision-making.

Companies should establish predefined criteria for acceptable model performance and document the model’s intended use, limitations, and required level of human oversight.

An effective strategy is to deploy the model in parallel with the existing FSQA program, allowing companies to compare outputs against current decision frameworks and assess whether the algorithm provides practical value.

Human experts must remain in the loop, particularly when making decisions about interventions and corrective actions, as current AI systems are not intended to replace professional accountability.

AI systems depend heavily on data. What are the biggest limitations in the data that food safety programs currently collect?

The main limitations include data availability, standardization, quality, and sharing.

There are often not enough outcome data for the specific hazard, product, and process combination a company wants to model. For example, there may not be sufficient structured data to support reliable prediction of Listeria risk in a particular ready-to-eat meat operation.

Even when data exist, they are often inconsistently formatted, missing metadata, or dispersed across multiple systems—such as paper records, handwritten logs, and separate digital platforms—which limits integration.

Another challenge is that relevant data are often siloed across companies and facilities, with limited mechanisms for sharing information. This restricts the ability to identify broader contamination patterns and build more robust predictive models.

The issue is not just how much data are collected, but how usable, interoperable, and shareable those data are.

Five years from now, what role do you realistically expect AI to play inside food safety programs?

AI’s most practical roles will likely include reviewing and comparing food safety documents, drafting and updating SOPs and training materials, translating content, and helping employees access and query existing procedures.

It will also play a role in triaging large volumes of data and supporting multistep workflows—such as comparing regulatory or customer requirements against current programs, drafting revisions, and assigning follow-up tasks. This could reduce administrative burden and improve consistency in recordkeeping.

Another area of growth is video-language models for verification and anomaly detection. AI-enabled monitoring could help identify when facilities are not operating “as usual”—for example, detecting shifts in humidity, employee movement, or sanitation performance.

Overall, the most realistic expectation is that AI will help food safety programs become faster and more consistent in supporting decisions, rather than replacing human accountability.

All interviews have been edited for clarity and brevity.