Close Ad

The Spoon

Daily news and analysis about the food tech revolution

AI

Yummly App Adds New Features, Reminding Us It’s Still Around Six Years After Whirlpool Deal

Today, recipe and guided cooking app Yummly announced a refreshed set of features, including what it describes as AI-powered recipe recommendations, an improved meal planner feature, and integration with an upgraded Yummly thermometer.

Since Whirlpool acquired Yummly, the recipe recommendation and cooking guidance app has largely flown below the radar while adding periodic incremental improvements over the years. And as far as I can tell, the announced improvements are par for the course.

This includes the new and improved “AI-powered recipe recommendations,” which sounds a lot like the things the company was promoting almost five years ago when they were touting “AI-powered personalization.” It’s not immediately clear how these AI-powered recommendations differ from previous AI-enabled recommendations, but we’ll have to take the company’s word for it.

The app’s improved meal planner function looks like it’s primarily focused on further building out a shoppable recipe function, something that has become relatively common in recent years for many recipe apps as a way to monetize through affiliate marketing commerce. The Yummly meal-planning shoppable recipe meal planning capability is a premium feature for users through a monthly subscription.

Whirlpool is hardly mentioned in the release (outside of the About Yummly boilerplate at the bottom), and the only real evidence of the company’s influence is the integration with an improved Yummly Thermometer, which is a product that Whirlpool has gone through pains to integrate with a number of their appliances. According to the announcement, the new Yummly thermometer now has three sensors, up from the two sensors in the previous generation.

While Whirlpool seems content to let Yummly operate mainly as a standalone app with its own brand, it seems a far cry from when the company acquired the app and saw it as driving the digital transformation of the appliance giant’s product lineup. Outside of a big splash at CES 2019, which the company described as a “roll-out across multiple Whirlpool brands,” the app hasn’t added all that much in terms of feature sets beyond what it had five years ago, and there’s been scant evidence of any further integration – thermometer notwithstanding – with the broader Whirlpool family.

One reason the app has become something of an afterthought in Whirlpool might be that many of the original stakeholders have moved on. Yummly founder Dave Feller left soon after the deal was done, and Brian Whitlin, who drove much of the product innovation, left in 2021. Add in the fact that the acquisition’s primary champion within Whirlpool, Brett Dibkey – who drove much of Whirlpool’s digital transformation – left in 2020, and the company’s current caretaker mode makes sense.

Scentian Bio Raises $2.1M for Tech is Says Can Replicate Insect Smell Receptors

Scentian Bio, a biosensor startup that claims to have blended nature and technology by leveraging the olfactory capabilities of insects to develop a powerful new sensory tool, announced a $2.1 million seed funding today according to a release sent to The Spoon. The company’s new investors, which include Finistere Ventures and Toyota Ventures, will join the Bill and Melinda Gates Foundation, bringing their total backing up to $4.4 million.

Scentian says their technology relies on virtualizing insect olfactory receptors (iOR), using AI to process and interpret signals from its biosensors to replicate an insect’s neuronal network to interpret smells. According to Scentian, their sensors are a thousand times more sensitive than a dog’s nose and have dozens of unique receptors that effectively recognize millions of virtual organic compounds or VOCs.

The company says they’ve been running a trial with a large unnamed food brand, and based on its early success in this trial, the company plans to focus on quality control for the food industry initially. This company’s first digital biosensor, which is expected to launch commercially at the end of 2024, will provide quality control of key food ingredients’ smell and taste attributes. The company will focus on essential oils and expand later to other ingredients.

The pitch is compelling, but the company doesn’t explain precisely how their technology replicates insect smell detection. It says they combine “insect smell receptors” with sensing surfaces that create “the most sensitive digital fingerprint for smell.” It sounds good, but I have to wonder if it’s just a colorful way to describe a really powerful electronic nose.

ScentianBio - Unlocking the Language of Life

Amazon Details Usage of Generative AI-Created Synthetic Data to Train Just Walk Out Technology

For a while now, we’ve known the basic gist of how Amazon’s Just Walk Out technology works: A combination of computer vision, machine learning, and other sensor data helps enable a friction-free shopping experience in which customers pick items off the shelf and walk out the door without ever having to stop at a cash register.

But in a recent blog post by Amazon’s retail technology team, the company explained how it all worked in greater detail than we’ve seen in the past, including how the company has been using generative AI to train its Just Walk Out platform for long-tail cases that are rare but entirely possible in the unpredictable environment of retail.

According to Gérard Medioni, vice president and distinguished scientist at Amazon, the company uses a generative AI called a generative adversarial network (GAN) to create synthetic data for training Amazon’s Just Walk Out technology. The Just Walk Out team used datasets from millions of AI-generated synthetic images and video clips mimicking realistic, and sometimes rare, shopping scenarios, including variations in lighting, store layouts, and crowd sizes. According to Amazon, this training using generative AI-created synthetic visual data enables Just Walk Out to recognize and properly interpret millions of customer actions.

“When the customer exits, having an accurate account of their purchases is critical,” Medioni said.

The company also went into detail about how Just Walk Out and its Amazon One palm-based bioauthentication technology does – and don’t – work together. According to Amazon, the two systems operate independently of each other, keeping a person’s biometric information associated with their payment separate from Just Walk Out. When a shopper enters the store, the Just Walk Out system assigns the shopper a temporary numeric code, which serves as their unique digital signature for that shopping trip. When a shopper exits, the code disappears. When they come back, they get a new code.

Medioni says that Just Walk Out associates a person’s “pixels” to the one-time payment code assigned for that trip and the products they pick up off the shelf.

“Just Walk Out tech doesn’t collect any biometrics. All we need to know is where that person is on the floor, and where their hands are in relation to the store’s merchandise.”

According to Medioni, the system is sophisticated enough to track groups of shoppers assigned to a single payment instrument, and the system can create a single receipt for a group shopping trip.

“We had a tour bus that came in one day, and they had 90 people all paying with a single credit card,” Medioni adds. “Even if people leave the store separately and we can still keep track of the group’s purchases.”

While Amazon has shown mixed signals regarding its retail footprint, the company appears to remain interested in developing its technology platform for usage by other retailers. My guess is they’ll likely see some smaller retailers and non-grocers (like stadiums/sports venues) adopt the technology, but larger grocers will remain reticent to jump on board with technology developed by a competitor.

If you’re interested in how generative AI will change food retail, join us at the Spoon’s Food AI Summit on October 25th in Alameda!

Meet The Dutch Robotic Kitchen That Makes Five Thousand Meals Per Day

Last month, a Dutch startup named Eatch announced they had built a fully automated robotic kitchen that makes up to five thousand meals per day. The company’s new robot, designed to work in a high-production centralized kitchen, has been making meals in the Amsterdam market for food service and catering giant ISS for the past four months.

The Eatch robotic kitchen platform handles the entire meal production flow. It oils the cooking pans, dispenses refrigerated ingredients, adds spices, plates the food, and cleans the cooking pans when everything is done.

You can watch it in action in the video below:

Eatch - World's First Robotic Kitchen for Large-scale Cooking - Order meals at: maaltijden.eatch.me

Eatch’s robotic kitchen uses a pot system similar to those we’ve seen in the Spyce kitchen, Kitchen Robotics’ Beastro, and TechMagic’s pasta robot in Tokyo. The Eatch’s tilted pans rotate and toss the food inside, using an internal peg to push the food into a rotation and then drop from the top, creating a toss fry cooking motion common in stir fry kitchens.

What’s most impressive about the Eatch is the throughput, creating five thousand daily meals (and the company says it has the potential to produce up to 15 thousand per day), handling the entire production flow. Most robotic kitchens we’ve seen have production volumes much lower than this and often don’t incorporate plating and pot cleaning in the automation flow.

Company CEO Jelle Sijm told The Spoon that the company has approximately 10 employees and has raised €4.5 million. The company expansion plan includes working with partners who can handle the daily operations, and Eatch will provide the automation technology, software, and recipes. Sijm sees Eatch working with partners to produce food in centralized kitchens for contact caterers. Sijm says they are eyeing an American market entry and says the company is currently in talks with some grocery chains and contract caterers in the US.

With the Launch of Samsung Food, Samsung Hypes AI & Consolidates Food Features Acquired Over the Years

Over the years, Samsung has acquired and launched several products in an effort to become the king of the tech-powered kitchen.

First, there was the launch of the Family Hub refrigerator, the company’s attempt to create a smart fridge built around the company’s own operating system and packed with technology like fridge cams to identify food and help you with your shopping.

Then, there was the acquisition of Whisk, an intelligent food and shopping app that helped pioneer the shoppable recipe space. Whisk had not only amassed an extensive food database, which would ultimately become a foundation for some of Family Hub’s (now Bespoke Family Hub) shopping and recipe capabilities, but it also served up the foundational ‘Food AI’ that is now being pushed to the forefront by Samsung.

Then, there were various attempts to use AI through automation in the kitchen, as the company announced (and never released commercially) different cooking and kitchen-task robots at CES.

And we can’t forget that Samsung also took some of the smart home technology from its SmartThings smart home group (another Samsung acquisition) and paired it with Whisk’s recipe intelligence to create SmartThings Cooking, a guided cooking app.

This leads us to this week, in which Samsung announced what amounts to packaging up this collected knowledge and technology – save for (at least for now) the robotics – into a newly expanded app and platform called Samsung Food. Samsung Food, which the company describes as “a personalized, AI-powered food and recipe platform,” looks like a significant step forward for the company’s efforts to build a centralized digital food management app. It also is a logical move to consolidate much of the collected efforts under the Samsung brand after the company had collected a variety of platforms that served as a foundation for what we see today.

Let’s take a look at precisely what the company unveiled. In the announcement, Samsung detailed four primary areas of activity for Samsung Food: Recipe Exploration and Personalization, AI-Enhanced Meal Planning, Kitchen Connectivity, and Social Sharing.

For recipe exploration, Samsung looks like it’s essentially using what was an already somewhat evolved feature set in Whisk. Samsung says that it can save recipes to a user’s digital recipe box anytime and from anywhere, create shopping lists based on their ingredients, and is accessible via Family Hub. In addition to mobile devices, users can access Samsung Food with their Bespoke Family Hub fridges, which will provide recipe recommendations based on a list of available food items managed by the user and shoppable recipe capabilities.

With the Personalize Recipe function, Samsung Food looks like it builds on the personalization engine created by Whisk and plans to take it further through integration with Samsung Health. According to the announcement, by the end of this year, Samsung plans to integrate with Samsung Health to power suggestions for diet management. This integration will factor in info such as a user’s body mass index (BMI), body composition, and calorie consumption in pursuit of their health goals and efforts to maintain a balanced diet.

The AI-Enhanced meal planning feature looks like a longer-view planning feature that consolidates personalized recipe recommendations, and it will no doubt similarly benefit from the integration of Samsung Health.

With Connected Cooking, Samsung has rebranded and extended the features of the SmartThings Cooking app, adding new devices like the BeSpoke oven and incorporating some of the same guided cooking features.

And, of course, a consolidated food-related platform from Samsung wouldn’t be complete without a social media component. My guess is the Social Sharing feature – which will allow users to share with their community – is the least necessary addition to the app and will ultimately not be all that successful, as consumers will continue to use mainline social apps (TikTok, Instagram, Facebook) for their food-related social sharing.

The company also teased expanded computer vision capabilities in 2024 in the announcement. The company’s Vision AI technology “will enable Samsung Food to recognize food items and meals photographed through the camera and provide details about them, including nutrition information.”

Overall, I’m impressed with the overall cohesiveness and trajectory of what I see in Samsung Food. I think it’s a sign that Samsung – despite having the occasional misstep and strategic vagueness around their food vision – looks like they remain committed to becoming the leader of the future kitchen, something that they started way back in 2016 with the launch of the Family Hub line.

I Attended a Workshop on the Impact of AI on The Food World. Here’s What We Discussed

Last month, I headed down to San Luis Obispo to participate in a National Science Foundation-funded project analyzing the impact of automation and AI on the food system. I’d been invited to participate in a workshop headed up by Patrick Lin and Ryan Jenkins, professors at Cal Poly and the project leads.

The workshop was the first for the four-year project exploring the social and ethical impacts of automation and artificial intelligence in kitchens. The project endeavors to draw out the wide-ranging implications of this technology, exploring both the impact on commercial environments like restaurants and how automation could impact the longstanding tradition of home cooking and family meals.

“This project will help to draw out the hidden and very broad impacts of technology,” said Lin at the time of the project’s announcement. “By focusing on the trend of robot kitchens that’s just emerging from under the radar, there is still time for technical and policy interventions in order to maximize benefits and minimize harms and disruptions.” 

The two-day workshop included a cross-section of academic types, chefs and food service professionals, journalists, and technology experts. It was the first of three workshops across continents to gather insights and work towards producing a report and academic curriculum centered around the intersection of food and automation and AI.

The workshop, structured as a giant whiteboard session, included expert presentations and facilitated conversations. During and after each presentation, the participants shared their thoughts on potential impacts – both direct and cascading effects – that could result from the introduction of AI in its various forms over time. While much of the conversation focused more heavily on AI in the form of automation – i.e., cooking robots – AI in other forms, such as generative AI, was also discussed.

Below are some of the key themes discussed during the two days, as well as a few of my thoughts now that I’ve had time to think through the issues since the workshop.

I’d also love to hear your thoughts on this critical topic, so please send them along!

Finally, we’ll be discussing many of these same issues at the Food AI Summit on October 25th. If this is an issue critical to you and your company, make sure to join us!

Atrophying Cooking Skills

One of the concerns raised during the workshop was the potential loss of cooking skills and culinary knowledge as we rely increasingly on automation and AI to make our meals. While it was generally recognized that robotics could take over repetitive and tedious cooking tasks, some wondered if handing over the cooking process to machines could lead to a general loss of competency in culinary arts and a homogenization of meals produced by highly automated cooking.

It’s easy to see how highly automated food prep would be extremely popular; some would hand the entire process over to the machine. However, there’s a good chance that handing off the mundane parts of cooking would give home cooks, chefs, or food workers more time to focus on creating the special touches that often make a meal great. As we have seen with the advent of digital design and art tools, there’s a possibility that those who love making food could use technology to take their work to the next level.

The Loss of Together Time

Another concern raised across the two days was the impact on shared family time by handing over meal prep and cooking to robots. Parents and other caregivers often use time in the kitchen to share lessons to help children develop motor skills, understand their heritage and develop self-confidence. Over-automation of cooking could disrupt this transfer of knowledge. Cooking has also shown many positive mental health benefits for those involved.

I think these are valid concerns, as there is a real risk of losing some of the benefits of the shared cooking process due to automation. After all, there’s no replacement for a grandchild spending time with their grandma learning how to make her special cookies and the sharing of family history that comes along with such an activity.

However, a few counterpoints. First, no one says the act of hand-making that special recipe has to be a victim of technology, and, in some ways, I think the kitchen will prove to be one of the areas where some families will insist on preserving the art and act of doing the actual cooking themselves.

And as the world becomes more digital and automated, kitchens may be a refuge for many who find the hands-on nature of making food therapeutic and fulfilling. In other words, the kitchen may be the last true ‘maker space’ left in our homes, and many will look to protect and preserve that.

Finally, average meal times shrank 5% between 2006 and 2014, a much smaller decline than we’ve seen in meal prep times as the advent of ready-to-eat meals has become more popular over the past few decades. While automation may result in faster meals, people could spend nearly as much time – or maybe more – sitting around the dinner table.

A Loss of Authenticity, Creativity, and Happy Accidents

With AI, there’s a chance recipe creation algorithms may rely too heavily on existing data patterns and therefore lack originality. There was also the concern that AI systems may limit opportunities for spontaneous creativity and the type of “happy accidents” that often lead to new recipes. One workshop participant gave an example of mistakes leading to important new dishes, like the croissant.

There was also concern that using AI to generate meal plans or recipes could result in over-standardization and homogenization, particularly if the AI systems rely too narrowly on popular recipes, which could also reduce culinary diversity.

It’s a valid concern that AI systems will generalize based on limited data sets, often creating recipes or meal plans based on popular or trending food concepts. Anyone who listens to algorithm-generated playlists by Spotify or Pandora can attest to some off-note song recommendations, and I can see how that could easily be the case with food and recipe generation. However, good technology products allow humans to reject recommendations and fine-tune algorithms, which may allow for more personalized recommendations based on a particular user’s preferences.

There’s also a real possibility that AI could lead to new and intriguing food combinations. Chef Watson and other AIs have been able to create unexpected but interesting recipes based on intelligence built into the algorithms around flavor compounds. If a restaurant or home chef can leverage heretofore inaccessible deep insights based on science and flavor research built into AI systems to create their next masterpiece, the results could be exciting.

As for the impact on cultural diversity, I think it’s important to recognize that AI systems are known to have bias problems, often hewing more closely to the worldviews of their creators and their preferred datasets. Because the world of food is one of the most important pathways for under-represented voices to connect with broader audiences, it will be critical for us to guard against the loss of accessibility and equality in the culinary world as AI and automation tools become more commonplace.

However, food AIs could be built to emphasize unique and emerging food cultures, which could be a savvy move since millennials and younger generations celebrate new food discoveries, often from cultures outside their home markets. Also, many of the creators of new food automation technology are often from markets outside our own, emphasizing food types different from our traditional fare.

This is just a few of the themes discussed during the workshop. Other themes, such as job loss and the economic impacts of automation, were also explored in detail, and I’ll have more thoughts on that later this week.

Innit Debuts FoodLM to Power More Contextually Relevant Answers from Generative AI Platforms

Today Innit, a startup best known for its shoppable recipe and smart kitchen software solutions, announced the release of FoodLM, a software intelligence layer that helps power more contextually relevant food-related answers from generative AI large language models (LLMs).

The new platform, which itself is not a new LLM, is instead a software intelligence layer built to plug into existing LLMs to do pre and post-processing of queries to help provide better answers around a variety of food-related topics.

From the announcement:

FoodLM enables powerful semantic search for retailers to go beyond keywords and understand intent. Brands can provide consumers with highly personalized AI assistance from product selection through preparation and cooking. For health providers supporting patients with chronic diseases such as type 2 diabetes, FoodLM provides powerful science-backed assistance for healthy eating and food as medicine.

Innit CEO Kevin Brown described FoodLM as a “vertical AI” expert layer that can integrate into popular LLMs such as OpenAI’s GPT4 or Google’s PaLM. Brown compared FoodLM to what Google has done with Med-PaLM, which is Google’s medical knowledge layer that provides focused answers that are so contextually smart around medical information that it has started to pass the medical exams.

“You’re going to need the pairing of an LLM with expert training and expert systems to narrow it down for certain functions where it’s essential to be accurate,” Brown said.

The biggest concern with LLMs today is their tendency to hallucinate. Brown says that integrating with a vertical knowledge layer increases the likelihood of more relevant and accurate answers, ultimately leading to more trust in these systems.

“Food queries are one of the top use cases for LLMs, helping with tough problems like helping to manage people’s diets,” said Brown, “But only if you can trust them. If you can trust these systems and ensure they reflect key dietary and health factors, it becomes much more valuable.”

According to the company, answers are pre-processed and post-processed through FoodLM’s focused computation models, which it calls validators. The different validators within FoodLM include:

  • Nutrition & Diets: Analyzes more than 60 diets, allergies, lifestyles, and health profiles to provide detailed recommendations tailored to individual needs.
  • Health Conditions: Provides dietary guidelines, product scoring, and content specifically designed for conditions such as type 2 diabetes or hypertension.
  • Personalized Shopping: Automated grocery purchases, incorporating personalized scoring and selection of over three million grocery products worldwide.
  • Culinary & Cooking: Advanced logic to ensure that AI-generated recipes follow culinary guidelines and are cookable. Seamlessly integrates with smart kitchens, featuring automated cooking programs.

For now, Brown says FoodLM will be used by its partners through custom integrations via API. Over time, he sees the system as having a more approachable user interface where the system is used via a SaaS model.

From my perspective, FoodLM makes lots of sense for Innit. While we’ve already seen similar moves from some data-service and SaaS providers in the food space, Innit’s offering goes further and has more granular breakouts to provide specific contextualized offerings to power food-related services for their CPG, appliance, and health/wellness industries.

If you’re interested in the intersection of food and AI, make sure to check out The Spoon’s Food AI Summit, which is on October 25th in Alameda, California.

SEERGRILLS Unveils the Perfecta, an ‘AI-Powered’ Grill That Cooks the ‘Perfect Steak’ in Two Minutes

AI is seemingly everywhere nowadays, so it was only a matter of time before it would show up at the backyard BBQ to help us cook the perfect steak.

That’s the vision of a UK startup named SEERGRILLS, which debuted the Perfecta this week, which the company describes as the world’s first AI-powered grill. The grill combines high-temperature infrared cooking with its AI system called NeuralFire, which automates the cooking process.

According to SEERGRILLS CEO Suraj Sudera, the AI works through a combination of sensor data, cook preferences inputted by the user, and intelligence built into the software around different food types.

“The device will capture the starting temperature of, say, chicken breast and adjust the cooking in line with the preferences you’ve inputted in the device,” said Sudera. “Whether it’s a three-inch or five-inch chicken breast, it doesn’t matter. It will be whatever adjustments it needs, just like your cruise control on your car will adjust to keep you at the preferred speed.”

When a cook is done, users can rate the quality of the cook, which informs and optimizes the NeuralFire algorithm for the next cook. Suraj says that SEERGRILLS is also constantly updating its food database, so if, say, a new type of steak from Japan becomes popular, the AI engine will be updated to optimize the cook for that meat type. The company says its AI will also optimize to reach each type of meat’s sear and doneness, as well as help to perfect the Maillard reaction.

The hardware itself is somewhat unique compared to other infrared grills on the market in that it cooks meat vertically. The user puts the meat in a holder, which will sense the temperature and thickness of the meat. Once inserted, both sides are cooked simultaneously using infrared heat, powered by propane, which SEERGRILLS says can reach 1652ºF. According to the company, the grill can cook three ribeyes in one minute and fifty seconds, six burgers in a minute and thirty seconds, and four chicken breasts in two minutes and thirty seconds.

In addition to the grill itself, the company is also building accessories such as a rotisserie module, a pizza module, and a grill station. The company will start taking preorders in July and plans to begin shipping the Perfecta by the end of this year. Pricing for the grill and its accessories has not yet been disclosed.

🚀 Introducing Perfecta™ - The World’s First AI Powered Grill. 🚀

Shiru Used AI To Discover Its First Novel Ingredient in 3 Months. The Next One Will Go Even Faster

This week, novel ingredient discovery startup Shiru announced they have commercially launched their first ingredient, OleoPro, a plant-based fat ingredient the company says doesn’t have the environmental costs or health consequences of animal fat. As part of the announcement, the company disclosed that the company’s first commercial partner is Griffith Foods, a commercial food ingredient manufacturer.

As readers of The Spoon know, Shiru is part of a cohort of startups using AI to discover new ingredients more quickly than traditional methods. Unlike many first-generation synthetic bio products, OleoPro was developed using machine learning, enabling a multifold acceleration of the discovery and testing phase according to the company.

The company’s discovery timeframe for OleoPro took less than three months. According to the announcement, “Shiru’s biochemists and computational biologists used AI to scan and select nearly 10,000 formulations” in that time frame, and “then they determined the precise molecules that would combine to form an ingredient with the unique oil-holding protein scaffold of animal fat.” The entire discovery and commercialization process took 18 months from the project’s start, much shorter than the multi-year process typical of classical synthetic biology workflows.

And now, according to Shiru CEO Jasmin Hume, that time frame for discovery will compress even more now that the company has built out its machine learning model. Finding a new novel protein or functional ingredient will take “eight to 10 weeks is like what we’re comfortable with,” Hume told me in a recent interview. “And what that means is, it’s not just digital, but at eight weeks, we have up to half a dozen proteins that we’re making at a couple of grams. And so we go from totally digital to pilot-produced ingredients, not one but a couple that can work, in about eight weeks.”

“Instead of a half decade and more than a quarter billion dollars in R&D to ship a viable product, Shiru used AI to dramatically reduce the cost and time to market of an essential ingredient of plant-based meat to a matter of months and a few hundred thousand dollars – and the cost of protein discovery at Shiru continues to decline,” said Dr. Ranjani Varadan, Shiru Chief Scientific Officer, in the announcement. Varadan, who sat down with The Spoon last summer, was previously VP of R&D at Impossible Foods.

Podcast: Talking AI & Food With Evan Rapoport

In this week’s podcast, we talk food and AI with Evan Rapoport.

Over the past decade, Evan has led teams in Google Research and other organizations, looking at how AI could impact biodiversity and change. During our conversation, we talked about a project called Tidal, in which he and Google used AI technology like computer vision and applied it to aquaculture, and discussed the impact of AI more broadly on the food system and how Evan thinks newer technology, like generative AI, might have an impact sooner than we think on the world of food.

You can find the full conversation on Apple Podcasts, Spotify, or wherever you get your podcasts, or just click play below!

Eating Disorder Org’s AI Blunder is a Cautionary Tale About Embracing Tech for Fundamentally Human Roles

One of the ongoing debates in tech circles and beyond is how fast AI will replace humans in certain lines of work. One role where we’ve already seen organizations embrace the technology is in customer support, deploying AI-powered customer interfaces to act as the first line of contact to handle inbound queries and provide critical information to customers.

The only problem? Sometimes the information they provide is wrong and potentially harmful to an organization’s customers. To illustrate this, we need to look no further than this week’s news about efforts by the National Eating Disorder Association to use AI-powered chatbots to replace human workers for the organization’s call helpline. The group announced a chatbot named Tessa earlier this month after the helpline workers decided to unionize and, just a couple of weeks later, announced they would shut the chatbot down.

The quick about-face resulted from the chatbot giving out information that, according to NEDA, “was harmful and unrelated to the program.” This included giving weight loss advice to a body-positive activist named Sharon Maxwell, who has a history of eating disorders. Maxwell documented the interaction in which the bot told her to weigh herself daily and track her calories. In doing so, the bot went off-script since it was only supposed to walk users through the organization’s eating disorder prevention program and refer them to other resources.

While one has to question the decision-making of an organization that thought it could replace professionals trained to help those with sensitive health and mental wellness challenges, the example of NEDA is a cautionary tale for any organization eager to replace humans with AI. In the world of food and nutrition, AI can be a valuable tool to provide information to customers. However, the potential cost savings and efficiency the technology provides must be balanced against the need for a nuanced human understanding of the sensitive issues and the potential damage bad information could cause.

NEDA saw AI as a quick fix to what it saw as a nuisance in the form of real human workers and their pesky desire to organize a union to force change in the workplace. But unfortunately, in swapping out humans for a computer simulation of humans, the organization lost sight of the fact that serving their community requires a fundamentally human form of expression in empathy, something AI is famously bad at.

All forms of customer interaction are not created equal. An AI that asks if you want a drink with your burger at the drive-thru is probably going to be suitable in most scenarios, but even in those scenarios, it’s perhaps best to tightly guardrail the AI’s knowledge set and build in offramps to the system where customers can seamlessly be handed over to an actual human in case they have a specialized question or if there’s any potential for doing more harm than good during the interaction.

Recipe for Disaster? ChatGPT is Tasked to Create Unique, Tasty Dishes and Fails Miserably

So you think your newfound ability to prompt ChatGPT for AI-generated recipes could result in a culinary masterpiece?

Hold that thought, advises the World of Vegan, a popular wellness website focused on vegan living. The site recently undertook an intriguing experiment powered by generative AI, where they prompted ChatGPT to conjure over a hundred diverse recipes. The group prompted the AI bot to whip up new and innovative recipes for a variety of occasions ranging from date night dishes to brunch and dessert ideas. From there, the site’s chef team tested each recipe to see how they tasted.

The result? Not good.

All this prompting led to what the group described as “hilariously pitiful results.” With many of the recipes, the chef team at World of Vegan spotted ingredient formulations that “would clash right away and where the mishaps would occur.” The team also felt the recipes were largely “deceptive,” seeming ordinary at first glance but often described as “rich” and “decadent” when they were quite the contrary.

“I had a feeling ChatGPT would struggle with recipe development, since developing recipes is such a delicate mixture of fine art and science,” World of Vegan founder and chef Michelle Cehn told The Spoon. “But I was shocked by just how difficult it was to find a single spring recipe written by ChatGPT that worked with a passing grade. This is a crucial warning for both food bloggers seeking shortcuts and home cooks looking for quick recipes. You’ll save yourself a lot of trouble (and wasted time, energy, and money) by bypassing ChatGPT and opting for a trusted blogger’s highly-rated recipe instead.”

Image credit: Erin Wysocarski

One of the biggest fails cooked up by the World of Vegan team was a vegan scalloped potato dish (pictured above), which the recipe’s chef said had an ingredient list and cooking instructions that were out of order. The resulting dish had an off-putting color, a pungent sauce, and tasted bad.

According to World of Vegan, out of the 100 or so recipes the team cooked up, only one – a cauliflower taco dish – resulted in an appetizing result.

Cehn believes the resulting 1% success rate might be due to ChatGPT’s reliance on what is essentially flawed data, namely millions of subpar recipes drawn from the Internet. With this as its foundation, things are destined to go poorly once the bot is tasked to create a unique recipe.

“A human brain can’t access all that information, so people are likely independently (and unintentionally) creating duplicate recipes online. Since ChatGPT must create a truly unique recipe, it has to get a little weird to create one that’s not plagiarized.”

While one might expect a site focused on creating recipes to be skeptical about AI filling its shoes, I don’t doubt the poor results are that far off from what others may find if they conducted a similar experiment. Good recipes often result from lots of experimentation and applied knowledge, something that you don’t get when a bot freewheels up a new dish idea out of thin air.

And while a more specialized AI trained on the compatibility of various culinary ingredients – something akin to a chatbot based on Chef Watson – might yield better results, we don’t have that, at least not yet.

Bottom line: human-powered recipe creators are still necessary…for the time being.