As a cottage food producer, making safe, delicious and unique foods is your business. Since 2017, the number of newly registered cottage food producers has increased by almost 40% each year in Minnesota! As the number of registered cottage food producers continues to grow, this might mean new competition at farmers markets, community events or online. You may find yourself searching for or developing new recipes for your cottage food products.
Generative AI (artificial intelligence) or AI systems can be sources of recipes and food preservation information, but are the results safe? Many AI systems are free to use. For the purposes of this article, I used three different AI systems to search for recipes and information about food preservation: Ask Meta in the explore function of Instagram, AI Overview in the Google search engine and ChatGPT. The goal of this testing was not to compare different AI systems against each other, but to see if the search or question results aligned with research-based food preservation and cooking procedures.
Food safety and AI-generated recipes
Using Ask Meta, I searched “how to pickle cucumbers.” The results were a mix of how-to videos, images and tips. There were some interesting aromatics and spice combinations that could be used without impacting the processing time of the jars. However, I saw terminology frequently misused, such as quick-process pickles (fresh-pack) when the content creator was actually talking about refrigerator pickles. This can have an impact on food safety because one version (quick-process pickles) is processed to become shelf-stable and the other is not. I also viewed videos demonstrating how to ferment cucumbers, and many of these included questionable types of ingredients, proportions of ingredients or procedures that could lead to foodborne illness.
The AI Overview in the Google search engine summarizes or creates a snapshot of what it identifies as key information and then provides links for you to read more. One of my searches was “how to can asparagus” and while the results were not entirely inaccurate, they were incomplete. AI Overview was correct that asparagus can only be pressure canned or pickled (acidified) and then water bath canned. However, it did not specify the type of pressure canner (dial-gauge or weighted gauge) to use or elevation and PSI requirements for safe processing. The instructions for pickling and canning asparagus were very vague: no ingredient portions, the jar size was not included and the headspace was incorrect. This lack of information could result in an unsafe product even if the jar was processed for the right amount of time. Lastly, of the three links for more information highlighted by AI Overview, only one was from a scientific, research-based source. The other two links were to recipe blogs.
When I used ChatGPT, I entered multiple “write a recipe for” prompts and asked specific food preservation questions. The results for each prompt were detailed, however, not always correct. If you are new to or inexperienced with a certain process, you might not question an ingredient or a step because the results appear to be thorough.
For example, when I asked ChatGPT to “write a recipe for preserved rhubarb,” at first glance the process looked complete. However, when I analyzed the ingredient quantities and each step, it seemed that the AI was combining information about making a rhubarb jam and canning stewed rhubarb. This preserved rhubarb recipe had some food safety risks: not enough sugar, no testing steps and inaccurate headspace. Similarly, when I asked ChatGPT “how do I can or preserve homemade tomato sauce?”, the information appeared correct at first. The steps included adding an acidic ingredient, but the amount of the acidic ingredient per quart jar was incorrect. The boiling water bath canner instructions were correct and it included a note about elevation. However, the pressure canning instructions lacked details about the type of canner and PSI based on elevation.
After many simple prompts, I began to phrase my questions to include key food safety words. I asked ChatGPT to “write a recipe for homemade hot sauce that can be safely canned and stored in my pantry.” Again, at first glance, the recipe looked promising. However, when I compared the amount of vinegar listed to tested recipes with similar ingredients (peppers, onion, garlic) and their quantities, it was clear that this AI-generated recipe did not contain enough vinegar to lower the pH to 4.6 or below. The processing steps were for a water bath canner but did not specify the jar size.
Lastly, I asked ChatGPT to write different recipes with a water activity of 0.85 or below, such as pumpkin bread. For the pumpkin bread, it named the recipe “Low Water Activity Pumpkin Bread.” Included at the end of the response was a note section about measuring the water activity with a water activity meter. The recipe did not include a final water activity value. I then asked ChatGPT to write a recipe for pumpkin bread. The pumpkin bread recipes were identical with three exceptions. The “Low Water Activity Pumpkin Bread” recipe used ½ cup less pumpkin puree, substituted sour cream or yogurt for the milk, and had a baking time of 60 to 70 minutes instead of 60 to 65 minutes. The “Low Water Activity Pumpkin Bread” recipe written by ChatGPT may have a water activity of 0.85 or less as the recipe is similar to an existing pumpkin bread recipe that has been previously lab-tested. However, the AI-generated recipe did not include a final water activity value.
Use with caution
Generative AI systems can be a way to generate new recipe ideas or for flavor inspiration. However, the results should be compared against research-based procedures and recipes, as well as, testing your final product for pH and/or water activity. The AI systems mentioned in this article obtain information from any web page regardless of its credibility. This can contribute to incomplete or inaccurate responses to your questions or prompts. At this time, if you choose to use AI systems, use them with caution. As a cottage food producer, it is important that you apply your knowledge of food safety controls and the roles specific ingredients play in recipes and food preservation methods. Critically evaluate the results from AI sources. If you are developing a new recipe, have the final product tested in a food testing lab.
Research-based recipes and resources include: the USDA Complete Guide to Home Canning, the National Center for Home Food Preservation, Ball publications from 1994 or newer, and education institutions like Extension. When searching online, you can refine the search results to show only sources from research-based institutions. In the search engine enter your question or prompt followed by “site:.edu.” For example: “how to ferment cucumbers site:.edu.”
Have a question you want answered?
Email your question(s) to the Food Safety Team at exfdsafe@umn.edu. It may be featured in this newsletter to benefit all Minnesota cottage food producers.