Hello and welcome to Eye on AI. In this edition: A new study suggests AI can be a team player…OpenAI promotes its COO while CEO Sam Altman shifts focus…Apple shakes up its AI team amid frustration over delayed Apple Intelligence features…a revolutionary new AI weather forecasting method…and AI transforms architecture.

Evidence of AI’s positive impact on productivity continues to mount. But while many executives view AI as ultimately a substitute for human labor, hoping it will eventually fully automate tasks and save on headcount, the data suggests that this is not the best way to think about the technology. Yes, in a few cases, AI can fully automate some tasks. But in most cases, today’s AI systems—including the so-called “AI agents” from the likes of Salesforce, ServiceNow, Microsoft, and Google—aren’t yet capable or reliable enough to do this. Instead, AI systems should be thought of as a complement to human labor—a way to lift the performance of people, not to replace them.

The latest support for this view comes from a fascinating study by a group of researchers—from Harvard, the University of Pennsylvania’s Wharton School of Business, ESSEC Business School in France, and consumer products giant Procter & Gamble—and published as a working paper on the research repository SSRN. (The authors include Wharton’s Ethan Mollick, who has attracted a huge social media following for his tips on how to use AI effectively in business.)

In 2024, the researchers conducted a one-day virtual product development workshop at P&G, with the process designed to mirror the one that consumer products behemoth famously uses—except this time with an AI twist. In particular, this workshop involved the “seed” stage of product development—which is about brainstorming lots of possible new product ideas and incubating them to the point where a decision can be made on whether to test them at a larger scale. P&G normally assigns two-person teams consisting of one Commercial operations person and one R&D expert to work together on brainstorming ideas. In this case, the researchers took 776 P&G employees from Commercial and R&D and randomly assigned them to do one of the following: work alone; work alone but with access to a generative AI assistant based on OpenAI’s GPT-4 model; work in the usual two-person brainstorming team consisting of one Commercial and one R&D person; or work in the usual two-person configuration but with access to the AI assistant.

The groups were then tasked with coming up with new ideas for consumer products in the various P&G divisions in which they worked (baby care, feminine care, grooming, and oral care). These ideas were then assessed by human judges with both relevant business and technology expertise.

AI lifts individual performance—by a lot

Two heads are generally better than one, so it is perhaps not surprising that individuals working alone and without access to AI did the worst. But it turned out that individuals assisted by AI performed, on average, better than two-person teams without AI. In fact, the performance of these AI-assisted individuals was not statistically worse than two-person teams working with AI. This might lead one to conclude that AI can indeed be a good substitute for human labor—enabling a company like Procter & Gamble to reduce its two-person product teams to just single individuals brainstorming with the help of AI.

There were some other big benefits to the individuals working with AI, too. Individuals working with AI were able to work faster—taking more than 16% less time to come up with an idea compared to people working without AI, while teams working with AI were about 12% faster.

Working with AI was also better than “bowling alone”—individuals reported more positive emotions and fewer negative ones during the product ideation process than the unassisted lone wolves.

Importantly, people working alone tended to come up with ideas that fit primarily into their professional silos—commercial people favoring product innovations that were mostly about novel commercial ideas (changes in branding, packaging, or marketing strategy) while the R&D specialists favored technological innovations. But when assisted by AI, these individuals achieved blended approaches, combining both technical innovation and commercial innovation—just like the human-human pairings did. “This suggests AI serves not just as an information provider but as an effective boundary-spanning mechanism, helping professionals reason across traditional domain boundaries and approach problems more holistically,” the researchers wrote.

Helping teams to be extraordinary

But, before you jump to the conclusion that AI should be used to reduce team sizes, it is important to point out perhaps the most interesting finding of the whole study: The two person teams working with AI produced far more ideas that the human experts rated as “exceptional”—the 10% that they judged most likely to lead to truly breakout products. And the human teams assisted by AI also reported the most enjoyment from working on the task, compared to the other groups.

Blogging about the findings, Mollick wrote that “organizations have primarily viewed AI as just another productivity tool, like a better calculator or spreadsheet,” but that employees were often using “AI for critical thinking and complex problem solving, not just routine productivity tasks.” AI could be seen as another member of the team—as a collaborator—not just another tool, he wrote. “Companies that focus solely on efficiency gains from AI will not only find workers unwilling to share their AI discoveries for fear of making themselves redundant but will also miss the opportunity to think bigger about the future of work,” he wrote. He encouraged organizations to reimagine work and management structures, not just seek to automate existing processes.

I am sure this is correct. Unfortunately, the temptation for many managers will be to grab at the labor and time savings AI offers, since there is an obvious and immediate pay-off in labor savings. It will take braver executives to argue for keeping people in place but using AI to empower them to be exceptional.

With that, here’s the rest of this week’s AI news.

Jeremy Kahn
jeremy.kahn@fortune.com
@jeremyakahn

Before we get to the news, if you’re interested in learning more about how AI will impact your business, the economy, and our societies (and given that you’re reading this newsletter, you probably are), please consider joining me at the Fortune Brainstorm AI London 2025 conference. The conference is being held May 6-7 at the Rosewood Hotel in London. Confirmed speakers include Cohere CEO Aidan Gomez, Mastercard chief product officer Jorn Lambert, eBay chief AI officer Nitzan Mekel, Sequoia partner Shaun Maguire, noted tech analyst Benedict Evans, and many more. I’ll be there, of course. I hope to see you there too. You can apply to attend here.

And if I miss you in London, why not consider joining me in Singapore on July 22 and 23 for Fortune Brainstorm AI Singapore. You can learn more about that event here.

AI IN THE NEWS

OpenAI promotes a trio of top execs, announces shift in CEO Sam Altman’s focus. The company said it was expanding chief operating officer Brad Lightcap’s role to include “business and day-to-day operations” including business strategy, key partnerships, and AI infrastructure. Altman meanwhile will focus more on product and research. The company also announced an expanded role for Mark Chen, officially announcing him as the company’s chief research officer, and saying he will “tightly integrate research and product development, enabling faster translation of research into products.” (Chen was named OpenAI’s senior vice president of research after the previous research head, Bob McGrew, left OpenAI in 2024, on the same day former chief technology officer Mira Murati also announced her departure from the company. According to Chen’s LinkedIn profile, he has been serving as chief research officer since January.) OpenAI also announced Julia Villagra as its new chief people officer—although her LinkedIn profile says she’d already been in that job since June 2024. You can read more here from CNBC.

Apple sued for false advertising over delayed Apple Intelligence features. A class-action lawsuit has been filed against the tech giant on behalf of consumers who claim they were misled into buying Apple’s latest iPhone models based on the promises of Apple Intelligence capabilities that the company said would be available, and touted in a major ad campaign—but which have now been postponed until at least 2026. The suit, filed in federal court in San Jose, Calif., seeks unspecified financial damages. Apple has not commented on the case. You can read more from Axios here.

In perhaps related news, Apple shakes up its AI team, according to Bloomberg. The news service cited unnamed sources familiar with the situation in reporting that Apple has shuffled its executive ranks in a bid to overcome its struggles to deliver improved AI capabilities to its Siri digital assistant. (These are some of the same delayed capabilities that prompted the lawsuit mentioned earlier.) Bloomberg reported that Apple CEO Tim Cook lost confidence in top AI executive John Giannandrea’s ability to deliver an improved Siri. Cook has moved Mike Rockwell, who had helped create Apple’s Vision Pro virtual reality goggles, to run development of the new Siri instead. Rockwell will report directly to Apple software chief Craig Federighi, bypassing Giannandrea, Bloomberg reported. The news service said that after it initially reported on the changes, Apple announced them to employees. But the company has not announced the reshuffle publicly and declined to comment on the report.

Chinese AI company behind Manus announces partnership with Alibaba. Reuters says that the Chinese startup Butterfly Effect, which is behind the viral hit AI system Manus AI, has announced a strategic partnership that will see Alibaba integrate Manus’s “agentic AI” capabilities with Alibaba’s Qwen open-source AI models and computing platforms.

Ant Group claims breakthrough using indigenous Chinese computer chips to train advance AI models. The Chinese financial services company, which is backed by Alibaba cofounder Jack Ma, has pioneered new techniques for training AI models using Chinese-made computer chips from both Alibaba and Huawei that can equal the performance of Nvidia’s H800 GPUs. That’s according to Bloomberg, which cited unnamed people familiar with the effort. Nvidia designed the H800 to comply with U.S. export restrictions on advanced AI chips. It has less processing power than Nvidia’s most advanced chips. But the H800 was still considered superior to most indigenous Chinese semiconductors. Now, according to Bloomberg’s sources, Ant was able to use native chips to train a “mixture of experts” machine learning model at 20% less cost than other methods using much more expensive Nvidia-made chips. The development highlights China’s progress toward self-sufficiency in AI technology despite U.S. export rules designed to slow China’s AI efforts.

Deepfake audio depicting U.S. Vice President JD Vance bad-mouthing Elon Musk goes viral on social media. Many thought the audio clip, in which a voice sounding like Vance’s is heard saying that the South African-born Musk is “cosplaying as a great American leader” while actually making the Trump administration and Vance “look bad” was authentic. It was widely shared on social media platforms over the weekend. Vance’s communications director William Martin said the audio was “100% fake.” Reality Defender, an independent company that makes software to detect deepfakes, also said its systems judged the audio to be “likely fake,” although it could not determine exactly which AI audio generation software was used to create it. You can read more from 404 Media here.

Netflix cofounder Reed Hastings gives $50 million to endow AI program at Bowdoin College. Hastings, an alumnus of the Maine school, wants more students and researchers working on questions about AI and its societal impacts, the New York Times reports. The gift, the largest in the college’s history, will help hire 10 new faculty members and start a new initiative dedicated to AI.

AI search company Perplexity is in early talks to raise funding at a $18 billion valuation. That would double the three-year old company’s current valuation. Perplexity is looking to raise between $500 million and $1 billion in new funding, Bloomberg said, citing one unnamed person it said was familiar with the fundraising effort. Perplexity declined to comment on the report.

EYE ON AI RESEARCH

AI makes further advances in weather forecasting. Google DeepMind previously demonstrated that AI-generated forecasts can be more accurate than those created with traditional physics-based models up to 10 days out. This was already a big advance. Physics-based simulations had to be run on supercomputers and often took hours to produce. The AI-generated forecasts could often be run on a single GPU and produced forecasts in minutes. But most attempts to apply AI to weather forecasting still required the use of a “numerical weather prediction” (or NWP) model at some stage of the process—often in the initial assimilation of data taken from various weather monitoring devices around the world (satellites, ocean buoys, land-based weather stations, etc.)

Now a group of researchers from the University of Cambridge, the U.K.’s Alan Turing Institute, Microsoft Research, and the European Centre for Medium-Range Weather Forecasts have developed an end-to-end AI system that cuts out the NWP completely. It manages to produce more accurate forecasts while also requiring even less computing capacity than previous AI weather forecasting tools. The system, called Aardvark Weather, could allow a forecaster using only a desktop computer to produce highly-accurate, ultra-bespoke forecasts for any part of the world. Using 10% of the input data that prior systems required, Aardvark outperforms the U.S.-run Global Forecast System in certain respects, and was competitive with United States Weather Service forecasts, the researchers said. The scientists published their findings on Aardvark in the prestigious peer-reviewed science journal Nature. You can read more about it here in the Guardian.

FORTUNE ON AI

Cybersecurity specialists are drowning in a sea of software vulnerabilities. AI may be able to help —by Christian Velasquez

Exclusive: Instacart bought his self-checkout startup for $350M. Now he’s teaming with a Google DeepMind alum to build low-cost robots —by Sharon Goldman

ChatGPT might be making its most frequent users more lonely, study by OpenAI and MIT Media Lab suggests —by Beatrice Nolan

With Ukraine showing the need for drones and AI, a new breed of defense company hopes to cash in on Europe’s $865 billion drive to counter Russia without America’s help —by David Meyer

Mistral AI CEO denies IPO plans, touts renewed focus on open source to best deeper-pocketed competitors —by Sharon Goldman

AI CALENDAR

April 9-11: Google Cloud Next, Las Vegas

April 24-28: International Conference on Learning Representations (ICLR), Singapore

May 6-7: Fortune Brainstorm AI London. Apply to attend here.

May 20-21: Google IO, Mountain View, Calif.

July 13-19: International Conference on Machine Learning (ICML), Vancouver

July 22-23: Fortune Brainstorm AI Singapore. Apply to attend here.

BRAIN FOOD

An architect learns to love AI. Tim Fu, a young architect based in London who previously worked for legendary architect Zaha Hadid’s firm, is at the forefront of adapting AI to the practice of architecture. Last week, his new firm, Tim Fu Studio, unveiled its designs for a series of seven luxury villas on the shores of Lake Bled in Slovenia, commissioned by an undisclosed Slovenian client. The project is among the first to showcase how AI can transform the architectural process—and how Fu and his employees use AI could hold lessons for us all.

At an event in London, Fu said the client asked him to “achieve a world class hospitality experience on the one hand, but also to respect the [local] heritage and celebrate it in a novel way.” To accomplish that goal, he fed a generative AI model lots of photos of traditional, local architectural styles found around Lake Bled, and then asked the AI model to generate ideas for how these could be updated to incorporate more contemporary elements. He selected several different ideas to explore further through an iterative process of prompting the model. For example, he specifically asked the AI system for ideas on how to reinterpret a traditional Slovenian architectural feature called a “rizalit” (an outwardly projecting, partially enclosed veranda) in ways that evoked the traditional designs but were thoroughly contemporary.

Once Tim Fu Studio developed ideas it liked, it had to produce visuals for the client. It did this using different AI models that can generate photorealistic renderings while also accurately depicting prospective buildings in the geometry of the site. Normally, an architecture firm would work with a specialized computer-generated design workshop to produce visuals like this—a process that Fu said could easily take two to three weeks. With the AI models, Fu says he can have visuals of equal quality to the client in just one day, without needing to outsource the work to a CGI firm.

Fu also uses AI to help design building floor plans and optimize buildings for things such as how people might flow through the space and also how energy efficient they are. He told me the next step will be to build AI models that can translate designs directly into architectural blueprints and building specs. He says such models aren’t far off.

While some architects fear automation will diminish human creativity, Fu said he sees AI as enhancing human creative powers. “AI is not specifically introducing new design language, but it is enabling us to more easily facilitate the production of our own complex ideas,” he said. Fu’s experience shows how AI can radically revamp work processes, while still allowing human creativity to thrive.

AI improves both individual and team performance, new study finds. Will companies draw the right lessons from it?