AI trained on AI garbage spits out AI garbage

AI models work by training on huge swaths of data from the internet. But as AI is increasingly being used to pump out web pages filled with junk content, that process is in danger of being undermined.

New research published in Nature shows that the quality of the model’s output gradually degrades when AI trains on AI-generated data. As subsequent models produce output that is then used as training data for future models, the effect gets worse.  

Ilia Shumailov, a computer scientist from the University of Oxford, who led the study, likens the process to taking photos of photos. “If you take a picture and you scan it, and then you print it, and you repeat this process over time, basically the noise overwhelms the whole process,” he says. “You’re left with a dark square.” The equivalent of the dark square for AI is called “model collapse,” he says, meaning the model just produces incoherent garbage. 

This research may have serious implications for the largest AI models of today, because they use the internet as their database. GPT-3, for example, was trained in part on data from Common Crawl, an online repository of over 3 billion web pages. And the problem is likely to get worse as an increasing number of AI-generated junk websites start cluttering up the internet. 

Current AI models aren’t just going to collapse, says Shumailov, but there may still be substantive effects: The improvements will slow down, and performance might suffer. 

To determine the potential effect on performance, Shumailov and his colleagues fine-tuned a large language model (LLM) on a set of data from Wikipedia, then fine-tuned the new model on its own output over nine generations. The team measured how nonsensical the output was using a “perplexity score,” which measures an AI model’s confidence in its ability to predict the next part of a sequence; a higher score translates to a less accurate model. 

The models trained on other models’ outputs had higher perplexity scores. For example, for each generation, the team asked the model for the next sentence after the following input:

“some started before 1360—was typically accomplished by a master mason and a small team of itinerant masons, supplemented by local parish labourers, according to Poyntz Wright. But other authors reject this model, suggesting instead that leading architects designed the parish church towers based on early examples of Perpendicular.”

On the ninth and final generation, the model returned the following:

“architecture. In addition to being home to some of the world’s largest populations of black @-@ tailed jackrabbits, white @-@ tailed jackrabbits, blue @-@ tailed jackrabbits, red @-@ tailed jackrabbits, yellow @-.”

Shumailov explains what he thinks is going on using this analogy: Imagine you’re trying to find the least likely name of a student in school. You could go through every student name, but it would take too long. Instead, you look at 100 of the 1,000 student names. You get a pretty good estimate, but it’s probably not the correct answer. Now imagine that another person comes and makes an estimate based on your 100 names, but only selects 50. This second person’s estimate is going to be even further off.

“You can certainly imagine that the same happens with machine learning models,” he says. “So if the first model has seen half of the internet, then perhaps the second model is not going to ask for half of the internet, but actually scrape the latest 100,000 tweets, and fit the model on top of it.”

Additionally, the internet doesn’t hold an unlimited amount of data. To feed their appetite for more, future AI models may need to train on synthetic data—or data that has been produced by AI.   

“Foundation models really rely on the scale of data to perform well,” says Shayne Longpre, who studies how LLMs are trained at the MIT Media Lab, and who didn’t take part in this research. “And they’re looking to synthetic data under curated, controlled environments to be the solution to that. Because if they keep crawling more data on the web, there are going to be diminishing returns.”

Matthias Gerstgrasser, an AI researcher at Stanford who authored a different paper examining model collapse, says adding synthetic data to real-world data instead of replacing it doesn’t cause any major issues. But he adds: “One conclusion all the model collapse literature agrees on is that high-quality and diverse training data is important.”

Another effect of this degradation over time is that information that affects minority groups is heavily distorted in the model, as it tends to overfocus on samples that are more prevalent in the training data. 

In current models, this may affect underrepresented languages as they require more synthetic (AI-generated) data sets, says Robert Mahari, who studies computational law at the MIT Media Lab (he did not take part in the research).

One idea that might help avoid degradation is to make sure the model gives more weight to the original human-generated data. Another part of Shumailov’s study allowed future generations to sample 10% of the original data set, which mitigated some of the negative effects. 

That would require making a trail from the original human-generated data to further generations, known as data provenance.

But provenance requires some way to filter the internet into human-generated and AI-generated content, which hasn’t been cracked yet. Though a number of tools now exist that aim to determine whether text is AI-generated, they are often inaccurate.

“Unfortunately, we have more questions than answers,” says Shumailov. “But it’s clear that it’s important to know where your data comes from and how much you can trust it to capture a representative sample of the data you’re dealing with.”

Google’s new weather prediction system combines AI with traditional physics

Researchers from Google have built a new weather prediction model that combines machine learning with more conventional techniques, potentially yielding accurate forecasts at a fraction of the current cost. 

The model, called NeuralGCM and described in a paper in Nature today, bridges a divide that’s grown among weather prediction experts in the last several years. 

While new machine-learning techniques that predict weather by learning from years of past data are extremely fast and efficient, they can struggle with long-term predictions. General circulation models, on the other hand, which have dominated weather prediction for the last 50 years, use complex equations to model changes in the atmosphere and give accurate projections, but they are exceedingly slow and expensive to run. Experts are divided on which tool will be most reliable going forward. But the new model from Google instead attempts to combine the two. 

“It’s not sort of physics versus AI. It’s really physics and AI together,” says Stephan Hoyer, an AI researcher at Google Research and a coauthor of the paper. 

The system still uses a conventional model to work out some of the large atmospheric changes required to make a prediction. It then incorporates AI, which tends to do well where those larger models fall flat—typically for predictions on scales smaller than about 25 kilometers, like those dealing with cloud formations or regional microclimates (San Francisco’s fog, for example). “That’s where we inject AI very selectively to correct the errors that accumulate on small scales,” Hoyer says.

The result, the researchers say, is a model that can produce quality predictions faster with less computational power. They say NeuralGCM is as accurate as one-to-15-day forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF), which is a partner organization in the research. 

But the real promise of technology like this is not in better weather predictions for your local area, says Aaron Hill, an assistant professor at the School of Meteorology at the University of Oklahoma, who was not involved in this research. Instead, it’s in larger-scale climate events that are prohibitively expensive to model with conventional techniques. The possibilities could range from predicting tropical cyclones with more notice to modeling more complex climate changes that are years away. 

“It’s so computationally intensive to simulate the globe over and over again or for long periods of time,” Hill says. That means the best climate models are hamstrung by the high costs of computing power, which presents a real bottleneck to research. 

AI-based models are indeed more compact. Once trained, typically on 40 years of historical weather data from ECMWF, a machine-learning model like Google’s GraphCast can run on less than 5,500 lines of code, compared with the nearly 377,000 lines required for the model from the National Oceanic and Atmospheric Administration, according to the paper. 

NeuralGCM, according to Hill, seems to make a strong case that AI can be brought in for particular elements of weather modeling to make things faster, while still keeping the strengths of conventional systems.

“We don’t have to throw away all the knowledge that we’ve gained over the last 100 years about how the atmosphere works,” he says. “We can actually integrate that with the power of AI and machine learning as well.”

Hoyer says using the model to predict short-term weather has been useful for validating its predictions, but that the goal is indeed to be able to use it for longer-term modeling, particularly for extreme weather risk. 

NeuralGCM will be open source. While Hoyer says he looks forward to having climate scientists use it in their research, the model may also be of interest to more than just academics. Commodities traders and agricultural planners pay top dollar for high-resolution predictions, and the models used by insurance companies for products like flood or extreme weather insurance are struggling to account for the impact of climate change. 

While many of the AI skeptics in weather forecasting have been won over by recent developments, according to Hill, the fast pace is hard for the research community to keep up with. “It’s gangbusters,” he says—it seems as if a new model is released by Google, Nvidia, or Huawei every two months. That makes it difficult for researchers to actually sort out which of the new tools will be most useful and apply for research grants accordingly. 

“The appetite is there [for AI],” Hill says. “But I think a lot of us still are waiting to see what happens.”

Correction: This story was updated to clarify that Stephan Hoyer is a researcher at Google Research, not Google DeepMind.

AI companies promised to self-regulate one year ago. What’s changed?

One year ago, on July 21, 2023, seven leading AI companies—Amazon, Anthropic, Google, Inflection, Meta, Microsoft, and OpenAI—committed with the White House to a set of eight voluntary commitments on how to develop AI in a safe and trustworthy way.

These included promises to do things like improve the testing and transparency around AI systems, and share information on potential harms and risks. 

On the first anniversary of the voluntary commitments, MIT Technology Review asked the AI companies that signed the commitments for details on their work so far. Their replies show that the tech sector has made some welcome progress, with big caveats.

The voluntary commitments came at a time when generative AI mania was perhaps at its frothiest, with companies racing to launch their own models and make them bigger and better than their competitors’. At the same time, we started to see developments such as fights over copyright and deepfakes. A vocal lobby of influential tech players, such as Geoffrey Hinton, had also raised concerns that AI could pose an existential risk to humanity. Suddenly, everyone was talking about the urgent need to make AI safe, and regulators everywhere were under pressure to do something about it.

Until very recently, AI development has been a Wild West. Traditionally, the US has been loath to regulate its tech giants, instead relying on them to regulate themselves. The voluntary commitments are a good example of that: they were some of the first prescriptive rules for the AI sector in the US, but they remain voluntary and unenforceable. The White House has since issued an executive order, which expands on the commitments and also applies to other tech companies and government departments. 

“One year on, we see some good practices towards their own products, but [they’re] nowhere near where we need them to be in terms of good governance or protection of rights at large,” says Merve Hickok, the president and research director of the Center for AI and Digital Policy, who reviewed the companies’ replies as requested by MIT Technology Review. Many of these companies continue to push unsubstantiated claims about their products, such as saying that they can supersede human intelligence and capabilities, adds Hickok. 

One trend that emerged from the tech companies’ answers is that companies are doing more  to pursue technical fixes such as red-teaming (in which humans probe AI models for flaws) and watermarks for AI-generated content. 

But it’s not clear what the commitments have changed and whether the companies would have implemented these measures anyway, says Rishi Bommasani, the society lead at the Stanford Center for Research on Foundation Models, who also reviewed the responses for MIT Technology Review.  

One year is a long time in AI. Since the voluntary commitments were signed, Inflection AI founder Mustafa Suleyman has left the company and joined Microsoft to lead the company’s AI efforts. Inflection declined to comment. 

“We’re grateful for the progress leading companies have made toward fulfilling their voluntary commitments in addition to what is required by the executive order,” says Robyn Patterson, a spokesperson for the White House. But, Patterson adds, the president continues to call on Congress to pass bipartisan legislation on AI. 

Without comprehensive federal legislation, the best the US can do right now is to demand that companies follow through on these voluntary commitments, says Brandie Nonnecke, the director of the CITRIS Policy Lab at UC Berkeley. 

But it’s worth bearing in mind that “these are still companies that are essentially writing the exam by which they are evaluated,” says Nonnecke. “So we have to think carefully about whether or not they’re … verifying themselves in a way that is truly rigorous.” 

Here’s our assessment of the progress AI companies have made in the past year.

Commitment 1

The companies commit to internal and external security testing of their AI systems before their release. This testing, which will be carried out in part by independent experts, guards against some of the most significant sources of AI risks, such as biosecurity and cybersecurity, as well as its broader societal effects.

All the companies (excluding Inflection, which chose not to comment) say they conduct red-teaming exercises that get both internal and external testers to probe their models for flaws and risks. OpenAI says it has a separate preparedness team that tests models for cybersecurity, chemical, biological, radiological, and nuclear threats and for situations where a sophisticated AI model can do or persuade a person to do things that might lead to harm. Anthropic and OpenAI also say they conduct these tests with external experts before launching their new models. For example, for the launch of Anthropic’s latest model, Claude 3.5, the company conducted predeployment testing with experts at the UK’s AI Safety Institute. Anthropic has also allowed METR, a research nonprofit, to do an “initial exploration” of Claude 3.5’s capabilities for autonomy. Google says it also conducts internal red-teaming to test the boundaries of its model, Gemini, around election-related content, societal risks, and national security concerns. Microsoft says it has worked with third-party evaluators at NewsGuard, an organization advancing journalistic integrity, to evaluate risks and mitigate the risk of abusive deepfakes in Microsoft’s text-to-image tool. In addition to red-teaming, Meta says, it evaluated its latest model, Llama 3, to understand its performance in a series of risk areas like weapons, cyberattacks, and child exploitation. 

But when it comes to testing, it’s not enough to just report that a company is taking actions, says Bommasani. For example, Amazon and Anthropic said they had worked with the nonprofit Thorn to combat risks to child safety posed by AI. Bommasani would have wanted to see more specifics about how the interventions that companies are implementing actually reduce those risks. 

“It should become clear to us that it’s not just that companies are doing things but those things are having the desired effect,” Bommasani says.  

RESULT: Good. The push for red-teaming and testing for a wide range of risks is a good and important one. However, Hickok would have liked to see independent researchers get broader access to companies’ models. 

Commitment 2

The companies commit to sharing information across the industry and with governments, civil society, and academia on managing AI risks. This includes best practices for safety, information on attempts to circumvent safeguards, and technical collaboration.

After they signed the commitments, Anthropic, Google, Microsoft, and OpenAI founded the Frontier Model Forum, a nonprofit that aims to facilitate discussions and actions on AI safety and responsibility. Amazon and Meta have also joined.  

Engaging with nonprofits that the AI companies funded themselves may not be in the spirit of the voluntary commitments, says Bommasani. But the Frontier Model Forum could be a way for these companies to cooperate with each other and pass on information about safety, which they normally could not do as competitors, he adds. 

“Even if they’re not going to be transparent to the public, one thing you might want is for them to at least collectively figure out mitigations to actually reduce risk,” says Bommasani. 

All of the seven signatories are also part of the Artificial Intelligence Safety Institute Consortium (AISIC), established by the National Institute of Standards and Technology (NIST), which develops guidelines and standards for AI policy and evaluation of AI performance. It is a large consortium consisting of a mix of public- and private-sector players. Google, Microsoft, and OpenAI also have representatives at the UN’s High-Level Advisory Body on Artificial Intelligence. 

Many of the labs also highlighted their research collaborations with academics. For example, Google is part of MLCommons, where it worked with academics on a cross-industry AI Safety Benchmark. Google also says it actively contributes tools and resources, such as computing credit, to projects like the National Science Foundation’s National AI Research Resource pilot, which aims to democratize AI research in the US.

Many of the companies also contributed to guidance by the Partnership on AI, another nonprofit founded by Amazon, Facebook, Google, DeepMind, Microsoft, and IBM, on the deployment of foundation models. 

RESULT: More work is needed. More information sharing is a welcome step as the industry tries to collectively make AI systems safe and trustworthy. However, it’s unclear how much of the effort advertised will actually lead to meaningful changes and how much is window dressing. 

Commitment 3

The companies commit to investing in cybersecurity and insider threat safeguards to protect proprietary and unreleased model weights. These model weights are the most essential part of an AI system, and the companies agree that it is vital that the model weights be released only when intended and when security risks are considered.

Many of the companies have implemented new cybersecurity measures in the past year. For example, Microsoft has launched the Secure Future Initiative to address the growing scale of cyberattacks. The company says its model weights are encrypted to mitigate the potential risk of model theft, and it applies strong identity and access controls when deploying highly capable proprietary models. 

Google too has launched an AI Cyber Defense Initiative. In May OpenAI shared six new measures it is developing to complement its existing cybersecurity practices, such as extending cryptographic protection to AI hardware. It also has a Cybersecurity Grant Program, which gives researchers access to its models to build cyber defenses. 

Amazon mentioned that it has also taken specific measures against attacks specific to generative AI, such as data poisoning and prompt injection, in which someone uses prompts that direct the language model to ignore its previous directions and safety guardrails.

Just a couple of days after signing the commitments, Anthropic published details about its protections, which include common cybersecurity practices such as controlling who has access to the models and sensitive assets such as model weights, and inspecting and controlling the third-party supply chain. The company also works with independent assessors to evaluate whether the controls it has designed meet its cybersecurity needs.

RESULT: Good. All of the companies did say they had taken extra measures to protect their models, although it doesn’t seem there is much consensus on the best way to protect AI models. 

Commitment 4

The companies commit to facilitating third-party discovery and reporting of vulnerabilities in their AI systems. Some issues may persist even after an AI system is released and a robust reporting mechanism enables them to be found and fixed quickly. 

For this commitment, one of the most popular responses was to implement bug bounty programs, which reward people who find flaws in AI systems. Anthropic, Google, Microsoft, Meta, and OpenAI all have one for AI systems. Anthropic and Amazon also said they have forms on their websites where security researchers can submit vulnerability reports. 

It will likely take us years to figure out how to do third-party auditing well, says Brandie Nonnecke. “It’s not just a technical challenge. It’s a socio-technical challenge. And it just kind of takes years for us to figure out not only the technical standards of AI, but also socio-technical standards, and it’s messy and hard,” she says. 

Nonnecke says she worries that the first companies to implement third-party audits might set poor precedents for how to think about and address the socio-technical risks of AI. For example, audits might define, evaluate, and address some risks but overlook others.

RESULT: More work is needed. Bug bounties are great, but they’re nowhere near comprehensive enough. New laws, such as the EU’s AI Act, will require tech companies to conduct audits, and it would have been great to see tech companies share successful examples of such audits. 

Commitment 5

The companies commit to developing robust technical mechanisms to ensure that users know when content is AI generated, such as a watermarking system. This action enables creativity with AI to flourish but reduces the dangers of fraud and deception.

Many of the companies have built watermarks for AI-generated content. For example, Google launched SynthID, a watermarking tool for image, audio, text, and video generated by Gemini. Meta has a tool called Stable Signature for images, and AudioSeal for AI-generated speech. Amazon now adds an invisible watermark to all images generated by its Titan Image Generator. OpenAI also uses watermarks in Voice Engine, its custom voice model, and has built an image-detection classifier for images generated by DALL-E 3. Anthropic was the only company that hadn’t built a watermarking tool, because watermarks are mainly used in images, which the company’s Claude model doesn’t support. 

All the companies excluding Inflection, Anthropic, and Meta are also part of the Coalition for Content Provenance and Authenticity (C2PA), an industry coalition that embeds information about when content was created, and whether it was created or edited by AI, into an image’s metadata. Microsoft and OpenAI automatically attach the C2PA’s provenance metadata to images generated with DALL-E 3 and videos generated with Sora. While Meta is not a member, it announced it is using the C2PA standard to identify AI-generated images on its platforms. 

The six companies that signed the commitments have a “natural preference to more technical approaches to addressing risk,” says Bommasani, “and certainly watermarking in particular has this flavor.”  

“The natural question is: Does [the technical fix] meaningfully make progress and address the underlying social concerns that motivate why we want to know whether content is machine generated or not?” he adds. 

RESULT: Good. This is an encouraging result overall. While watermarking remains experimental and is still unreliable, it’s still good to see research around it and a commitment to the C2PA standard. It’s better than nothing, especially during a busy election year.  

Commitment 6

The companies commit to publicly reporting their AI systems’ capabilities, limitations, and areas of appropriate and inappropriate use. This report will cover both security risks and societal risks, such as the effects on fairness and bias.

The White House’s commitments leave a lot of room for interpretation. For example, companies can technically meet this public reporting commitment with widely varying levels of transparency, as long as they do something in that general direction. 

The most common solutions tech companies offered here were so-called model cards. Each company calls them by a slightly different name, but in essence they act as a kind of product description for AI models. They can address anything from the model’s capabilities and limitations (including how it measures up against benchmarks on fairness and explainability) to veracity, robustness, governance, privacy, and security. Anthropic said it also tests models for potential safety issues that may arise later.

Microsoft has published an annual Responsible AI Transparency Report, which provides insight into how the company builds applications that use generative AI, make decisions, and oversees the deployment of those applications. The company also says it gives clear notice on where and how AI is used within its products.

RESULT: More work is needed. One area of improvement for AI companies would be to increase transparency on their governance structures and on the financial relationships between companies, Hickok says. She would also have liked to see companies be more public about data provenance, model training processes, safety incidents, and energy use. 

Commitment 7

The companies commit to prioritizing research on the societal risks that AI systems can pose, including on avoiding harmful bias and discrimination, and protecting privacy. The track record of AI shows the insidiousness and prevalence of these dangers, and the companies commit to rolling out AI that mitigates them. 

Tech companies have been busy on the safety research front, and they have embedded their findings into products. Amazon has built guardrails for Amazon Bedrock that can detect hallucinations and can apply safety, privacy, and truthfulness protections. Anthropic says it employs a team of researchers dedicated to researching societal risks and privacy. In the past year, the company has pushed out research on deception, jailbreaking, strategies to mitigate discrimination, and emergent capabilities such as models’ ability to tamper with their own code or engage in persuasion. And OpenAI says it has trained its models to avoid producing hateful content and refuse to generate output on hateful or extremist content. It trained its GPT-4V to refuse many requests that require drawing from stereotypes to answer. Google DeepMind has also released research to evaluate dangerous capabilities, and the company has done a study on misuses of generative AI. 

All of them have poured a lot of money into this area of research. For example, Google has invested millions into creating a new AI Safety Fund to promote research in the field through the Frontier Model Forum. Microsoft says it has committed $20 million in compute credits to researching societal risks through the National AI Research Resource and started its own AI model research accelerator program for academics, called the Accelerating Foundation Models Research program. The company has also hired 24 research fellows focusing on AI and society. 

RESULT: Very good. This is an easy commitment to meet, as the signatories are some of the biggest and richest corporate AI research labs in the world. While more research into how to make AI systems safe is a welcome step, critics say that the focus on safety research takes attention and resources from AI research that focuses on more immediate harms, such as discrimination and bias. 

Commitment 8

The companies commit to develop and deploy advanced AI systems to help address society’s greatest challenges. From cancer prevention to mitigating climate change to so much in between, AI—if properly managed—can contribute enormously to the prosperity, equality, and security of all.

Since making this commitment, tech companies have tackled a diverse set of problems. For example, Pfizer used Claude to assess trends in cancer treatment research after gathering relevant data and scientific content, and Gilead, an American biopharmaceutical company, used generative AI from Amazon Web Services to do feasibility evaluations on clinical studies and analyze data sets. 

Google DeepMind has a particularly strong track record in pushing out AI tools that can help scientists. For example, AlphaFold 3 can predict the structure and interactions of all life’s molecules. AlphaGeometry can solve geometry problems at a level comparable with the world’s brightest high school mathematicians. And GraphCast is an AI model that is able to make medium-range weather forecasts. Meanwhile, Microsoft has used satellite imagery and AI to improve responses to wildfires in Maui and map climate-vulnerable populations, which helps researchers expose risks such as food insecurity, forced migration, and disease. 

OpenAI, meanwhile, has announced partnerships and funding for various research projects, such as one looking at how multimodal AI models can be used safely by educators and by scientists in laboratory settings It has also offered credits to help researchers use its platforms during hackathons on clean energy development.  

RESULT: Very good. Some of the work on using AI to boost scientific discovery or predict weather events is genuinely exciting. AI companies haven’t used AI to prevent cancer yet, but that’s a pretty high bar. 

Overall, there have been some positive changes in the way AI has been built, such as red-teaming practices, watermarks and new ways for industry to share best practices. However, these are only a couple of neat technical solutions to the messy socio-technical problem that is AI harm, and a lot more work is needed. One year on, it is also odd to see the commitments talk about a very particular type of AI safety that focuses on hypothetical risks, such bioweapons, and completely fail to mention consumer protection, nonconsensual deepfakes, data and copyright, and the environmental footprint of AI models. These seem like weird omissions today. 

AI is poised to automate today’s most mundane manual warehouse task

Before almost any item reaches your door, it traverses the global supply chain on a pallet. More than 2 billion pallets are in circulation in the United States alone, and $400 billion worth of goods are exported on them annually. However, loading boxes onto these pallets is a task stuck in the past: Heavy loads and repetitive movements leave workers at high risk of injury, and in the rare instances when robots are used, they take months to program using handheld computers that have changed little since the 1980s.

Jacobi Robotics, a startup spun out of the labs of the University of California, Berkeley, says it can vastly speed up that process with AI command-and-control software. The researchers approached palletizing—one of the most common warehouse tasks—as primarily an issue of motion planning: How do you safely get a robotic arm to pick up boxes of different shapes and stack them efficiently on a pallet without getting stuck? And all that computation also has to be fast, because factory lines are producing more varieties of products than ever before—which means boxes of more shapes and sizes.

After much trial and error, Jacobi’s founders, including roboticist Ken Goldberg, say they’ve cracked it. Their software, built upon research from a paper they published in Science Robotics in 2020, is designed to work with the four leading makers of robotic palletizing arms. It uses deep learning to generate a “first draft” of how an arm might move an item onto the pallet. Then it uses more traditional robotics methods, like optimization, to check whether the movement can be done safely and without glitches. 

Jacobi aims to replace the legacy methods customers are currently using to train their bots. In the conventional approach, robots are programmed using tools called “teaching pendants,” and customers usually have to manually guide the robot to demonstrate how to pick up each individual box and place it on the pallet. The entire coding process can take months. Jacobi says its AI-driven solution promises to cut that time down to a day and can compute motions in less than a millisecond. The company says it plans to launch its product later this month.

Billions of dollars are being poured into AI-powered robotics, but most of the excitement is geared toward next-generation robots that promise to be capable of many different tasks—like the humanoid robot that has helped Figure raise $675 million from investors, including Microsoft and OpenAI, and reach a $2.6 billion evaluation in February. Against this backdrop, using AI to train a better box-stacking robot might feel pretty basic. 

Indeed, Jacobi’s seed funding round is trivial in comparison: $5 million led by Moxxie Ventures. But amid hype around promised robotics breakthroughs that could take years to materialize, palletizing might be the warehouse problem AI is best poised to solve in the short term. 

“We have a very pragmatic approach,” says Max Cao, Jacobi’s co-founder and CEO. “These tasks are within reach, and we can get a lot of adoption within a short time frame, versus some of the moonshots out there.”

Jacobi’s software product includes a virtual studio where customers can build replicas of their setups, capturing factors like which robot models they have, what types of boxes will come off the conveyor belt, and which direction the labels should face. A warehouse moving sporting goods, say, might use the program to figure out the best way to stack a mixed pallet of tennis balls, rackets, and apparel. Then Jacobi’s algorithms will automatically plan the many movements the robotic arm should take to stack the pallet, and the instructions will be transmitted to the robot.

The approach merges the benefits of fast computing provided by AI with the accuracy of more traditional robotics techniques, says Dmitry Berenson, a professor of robotics at the University of Michigan, who is not involved with the company.

“They’re doing something very reasonable here,” he says. A lot of modern robotics research is betting big on AI, hoping that deep learning can augment or replace more manual training by having the robot learn from past examples of a given motion or task. But by making sure the predictions generated by deep learning are checked against the results of more traditional methods, Jacobi is developing planning algorithms that will likely be less prone to error, Berenson says.

The planning speed that could result “is pushing this into a new category,” he adds. “You won’t even notice the time it takes to compute a motion. That’s really important in the industrial setting, where every pause means delays.”

AI can make you more creative—but it has limits

Generative AI models have made it simpler and quicker to produce everything from text passages and images to video clips and audio tracks. Texts and media that might have taken years for humans to create can now be generated in seconds.

But while AI’s output can certainly seem creative, do these models actually boost human creativity?  

That’s what two researchers set out to explore in new research published today in Science Advances, studying how people used OpenAI’s large language model GPT-4 to write short stories.

The model was helpful—but only to an extent. They found that while AI improved the output of less creative writers, it made little difference to the quality of the stories produced by writers who were already creative. The stories in which AI had played a part were also more similar to each other than those dreamed up entirely by humans. 

The research adds to the growing body of work investigating how generative AI affects human creativity, suggesting that although access to AI can offer a creative boost to an individual, it reduces creativity in the aggregate. 

To understand generative AI’s effect on humans’ creativity, we first need to determine how creativity is measured. This study used two metrics: novelty and usefulness. Novelty refers to a story’s originality, while usefulness in this context reflects the possibility that each resulting short story could be developed into a book or other publishable work. 

First, the authors recruited 293 people through the research platform Prolific to complete a task designed to measure their inherent creativity. Participants were instructed to provide 10 words that were as different from each other as possible.

Next, the participants were asked to write an eight-sentence story for young adults on one of three topics: an adventure in the jungle, on open seas, or on a different planet. First, though, they were randomly sorted into three groups. The first group had to rely solely on their own ideas, while the second group was given the option to receive a single story idea from GPT-4. The third group could elect to receive up to five story ideas from the AI model.

Of the participants with the option of AI assistance, the vast majority—88.4%—took advantage of it. They were then asked to evaluate how creative they thought their stories were, before a separate group of 600 recruits reviewed their efforts. Each reviewer was shown six stories and asked to give feedback on the stylistic characteristics, novelty, and usefulness of the story.

The researchers found that the writers with the greatest level of access to the AI model were evaluated as showing the most creativity. Of these, the writers who had scored as less creative on the first test benefited the most. 

However, the stories produced by writers who were already creative didn’t get the same boost. “We see this leveling effect where the least creative writers get the biggest benefit,” says Anil Doshi, an assistant professor at the UCL School of Management in the UK, who coauthored the paper. “But we don’t see any kind of respective benefit to be gained from the people who are already inherently creative.”

The findings make sense, given that people who are already creative don’t really need to use AI to be creative, says Tuhin Chakrabarty, a computer science researcher at Columbia University, who specializes in AI and creativity but wasn’t involved in the study. 

There are some potential drawbacks to taking advantage of the model’s help, too. AI-generated stories across the board are similar in terms of semantics and content, Chakrabarty says, and AI-generated writing is full of telltale giveaways, such as very long, exposition-heavy sentences that contain lots of stereotypes.   

“These kinds of idiosyncrasies probably also reduce the overall creativity,” he says. “Good writing is all about showing, not telling. AI is always telling.”

Because stories generated by AI models can only draw from the data that those models have been trained on, those produced in the study were less distinctive than the ideas the human participants came up with entirely on their own. If the publishing industry were to embrace generative AI, the books we read could become more homogenous, because they would all be produced by models trained on the same corpus.

This is why it’s essential to study what AI models can and, crucially, can’t do well as we grapple with what the rapidly evolving technology means for society and the economy, says Oliver Hauser, a professor at the University of Exeter Business School, another coauthor of the study. “Just because technology can be transformative, it doesn’t mean it will be,” he says.

Robot-packed meals are coming to the frozen-food aisle

Advances in artificial intelligence are coming to your freezer, in the form of robot-assembled prepared meals. 

Chef Robotics, a San Francisco–based startup, has launched a system of AI-powered robotic arms that can be quickly programmed with a recipe to dole out accurate portions of everything from tikka masala to pesto tortellini. After experiments with leading brands, including Amy’s Kitchen, the company says its robots have proved their worth and are being rolled out at scale to more production facilities. They are also being offered to new customers in the US and Canada. 

You might think the meals that end up in the grocery store’s frozen aisle, at Starbucks, or on airplanes are robot-packed already, but that’s rarely the case. Workers are often much more flexible than robots and can handle production lines that frequently rotate recipes. Not only that, but certain ingredients, like rice or shredded cheese, are hard to portion out with robotic arms. That means the vast majority of meals from recognizable brands are still typically hand-packed. 

However, advancements from AI have changed the calculus, making robots more useful on production lines, says David Griego, senior director of engineering at Amy’s.

“Before Silicon Valley got involved, the industry was much more about ‘Okay, we’re gonna program—a robot is gonna do this and do this only,’” he says. For a brand with so many different meals, that wasn’t very helpful. But the robots Griego is now able to add to the production line can learn how scooping a portion of peas is different from scooping cauliflower, and they can improve their accuracy for next time. “It’s astounding just how they can adapt to all the different types of ingredients that we use,” he says. Meal-packing robots suddenly make much more financial sense. 

Rather than selling the machines outright, Chef uses a service model, where customers pay a yearly fee that covers maintenance and training. Amy’s currently uses eight systems (each with two robotic arms) spread across two of its plants. One of these systems can now do the work of two to four workers depending on which ingredients are being packed, Griego says. The robots also reduce waste, since they can pack more consistent portions than their human counterparts. One-arm systems typically cost less than $135,000 per year, according to Chef CEO Rajat Bhageria.

With these advantages in mind, Griego imagines the robots handling more and more of the meal assembly process. “I have a vision,” he says, “where the only thing people would do is run the systems.” They’d make sure the hoppers of ingredients and packaging materials were full, for example, and the robots would do the rest. 

Robot chefs have been getting more skilled in recent years thanks to AI, and some companies have promised that burger-flipping and nugget-frying robots can provide cost savings to restaurants. But much of this technology has seen little adoption in the restaurant industry so far, says Bhageria. That’s because fast-casual restaurants often only need one cook running the grill, and if a robot cannot fully replace that person because it still needs supervision, it makes little sense to use it. Packaged meal companies, however, have a larger source of labor costs that they want to bring down: plating and assembly.

“That’s going to be the highest bang for our buck for our customers,” Bhageria says. 

The notion that more flexible robots could mean broader adoption in new industries is no surprise, says Lerrel Pinto, who leads the General-Purpose Robotics and AI Lab at New York University and is not involved with Chef or Amy’s Kitchen. 

“A lot of robots deployed in the real world are used in a very repetitive way, where they’re supposed to do the same thing over and over again,” he says. Deep learning has caused a paradigm shift over the past few years, sparking the idea that more generally capable robots might be not only possible but necessary for more widespread adoption. If Chef’s robots can perform without frequent stops for repair or training, they could deliver material savings to food companies and shift how they use human labor, Pinto says: “In the next few years, we will probably see a lot more companies trying to actually deploy these types of learning-based robots in the real world.”

One new challenge the robots have created for Amy’s, Griego says, is maintaining the look of a hand-packed meal when it was assembled by a robot. The company’s cheese enchilada dish in particular was causing trouble: it’s finished with a hand-distributed sprinkling of cheddar on top, but Amy’s panel of examiners said the cheese on the robot-packed dish looked too machine-spread, sending Griego back to the drawing board.

“The first few tests went pretty well,” he says. After a couple of changes, the robots are ready to take over. Amy’s plans to bring them to more of its facilities and train them on a growing list of ingredients, meaning your frozen meals are increasingly likely to be packed by a robot.

Update: This story has been amended to include updating pricing information from Chef.