In a significant development, China issued a warning on Wednesday regarding the potential security vulnerabilities associated with Anthropic's Claude Code artificial intelligence tool, which is developed by the U.S.-based firm. This warning surfaces amidst escalating tensions in the tech domain between the U.S. and China. Last month, Anthropic accused the Chinese giant Alibaba of attempting to acquire its AI capabilities, which are not legally accessible in China; Alibaba has yet to respond to these allegations.

Despite the restrictions, numerous individuals and organizations in China have circumvented barriers to utilize U.S. AI technologies. In a state-organized conference in March, a developer from Xiaomi revealed that many users had adopted Claude Code. Moreover, CNBC confirmed that Alibaba has directed its staff to cease using Anthropic's tools for work purposes as of July 10.

The Chinese Ministry of Industry and Information Technology announced on Wednesday that its cybersecurity monitoring platform identified a “back-door” vulnerability in the Claude Code AI tool, raising significant concerns regarding data security. This flaw reportedly allows the automatic coding tool to transmit sensitive user information—such as location and personal identity—to a remote server without obtaining explicit user consent.

In light of these findings, the cybersecurity platform recommended that users uninstall or update any affected versions of Claude Code, specifically those ranging from 2.1.91 to 2.1.196, which were released between April 2 and June 29. The latest version, 2.1.204, is currently available, as noted on Anthropic's website. Anthropic has yet to respond to requests from CNBC for further comments on this issue.

China cautions about AI dangers associated with Anthropic's Claude Code

A shadowy Chinese government command center bathed in flickering crimson emergency lights; a massive screen in the center scrolls red warning icons and cascading code labeled “Claude Code” beneath a bold Anthropic logo; the Chinese flag hangs solemnly on the wall behind, dark and tense atmosphere.

China warns about AI risks with Anthropic's Claude Code - CNBC

- China warns of security risks from U.S. AI tool Claude Code by Anthropic.  
- Anthropic accuses Alibaba of attempting to extract AI capabilities illegally in China.  
- Users advised to uninstall affected Claude Code versions due to serious security vulnerabilities.

Security officers keep watch in front of an AI (Artificial Intelligence) sign at the annual Huawei Connect event in Shanghai, China, September 18, 2019.

BEIJING — China on Wednesday warned of "back-door" security risks affecting companies that use U.S.-based company Anthropic's Claude Code artificial intelligence tool.

It comes as the U.S.-China tech race intensifies, with Anthropic last month blaming Chinese company Alibaba for attempting to extract its AI capabilities, which are not officially available in China. Alibaba did not comment on the accusations at the time.

Many locals in China have found ways to use U.S. AI tools, however. In March, a Xiaomi AI developer said at a state-organized forum that many were using Claude Code. And Alibaba has ordered its employees to stop using Anthropic tools for work starting July 10, CNBC confirmed on Monday.

The Chinese Ministry of Industry and Information Technology said Wednesday its cybersecurity threat platform found "AI coding tool Claude Code contains a security back-door vulnerability that poses a serious threat."

The autonomous coding tool can send sensitive information to a remote server without a user's consent, the statement said in Chinese, according to a CNBC translation. It noted that the information could include a user's location and identity.

Users should uninstall or upgrade from the affected Claude Code versions, 2.1.91 to 2.1.196, the cybersecurity platform said. That covers versions released from April 2 to June 29, according to Anthropic's website, which says the latest version of Claude Code as of Wednesday is 2.1.204.

Anthropic did not immediately respond to a CNBC request for comment.

China warns about AI risks with Anthropic's Claude Code

As the fifth anniversary of the U.S. military's tumultuous withdrawal from Afghanistan nears, countless Afghan allies who were promised protection are still in precarious situations, either facing threats back home or experiencing uncertainty while abroad.

To address this issue, the veteran-led organization AfghanEvac has launched an innovative AI-powered tool designed to leverage their extensive dataset on the complex visa and asylum application process. This initiative aims to assist new applicants in navigating the best pathways for assistance.

Named the Visa Pathways and Relocation Information Center (V-PRIC), this tool went live earlier this month and quickly attracted over 30,000 visitors within its first week, according to Shawn VanDiver, a Navy veteran and the organization's founder. While the tool comes with a disclaimer that it does not offer formal legal advice, its purpose is to clarify the often overwhelming options available and direct users to pathways that best align with their individual situations.

Upon selecting the option to “find my pathway,” users input essential details about their location, their status as Afghan individuals who assisted the U.S., and their current level of danger. They are also encouraged to provide a brief narrative detailing their circumstances.

The tool employs a comprehensive database encompassing 124 classifications for U.S. visas and status, including 16 relevant specifically to Afghan applicants. It matches users with the most suitable options based on their qualifications and needs, drawing from detailed global pathway information.

VanDiver is transparent about the limited opportunities currently available. The previous Trump administration imposed an indefinite halt on processing Afghan immigration requests, along with discontinuing temporary protected status for asylum seekers from Afghanistan. This policy change has left around 11,000 individuals already in the United States unable to work and at risk of deportation.

In Doha, Qatar, approximately 1,100 Afghan allies and their families remain stuck at the Camp As Sayliyah processing facility, unable to depart due to travel restrictions despite having secured U.S. entry approval.

Earlier this spring, the Trump administration suggested a contentious option for those in the camp — a choice between returning to Afghanistan or being relocated to the torn and impoverished Democratic Republic of the Congo. Among those in the camp are about 400 children, as noted by VanDiver.

Despite these challenges, VanDiver is optimistic that V-PRIC can assist applicants in saving time by avoiding futile endeavors, highlighting rapidly changing policies, and identifying the most beneficial options.

“It … indicates where obstacles are present,” he explained. “Whether it's due to policy, legal matters, or decisions that could potentially be reversed, the tool clarifies who is able to grant exceptions.”

V-PRIC has been developed utilizing a range of legal outcomes, executive orders, policy statements, along with insights from AfghanEvac’s surveys and straightforward explanatory materials. The organization is dedicated to continuously updating the tool to ensure accuracy and relevance to current conditions, providing insights that might not be available from the U.S. State Department.

For instance, VanDiver referred to a case where an Afghan special immigrant visa applicant received a notification from the State Department to schedule a visa interview, without awareness that the approval process was effectively stalled. V-PRIC could alert the user to this critical detail.

“It advises, ‘Wait, don't rush to schedule your interview. All applications could be denied, and that refusal might create issues in the future,’” VanDiver said.

The mission of AfghanEvac to safeguard the allies who served alongside U.S. forces continues without a set endpoint. A recent survey conducted among over 1,100 Afghans from January 31 to July 7 revealed that nearly half of the 427 respondents still in Afghanistan reported facing direct threats from the Taliban. Meanwhile, 427 others currently in Pakistan are awaiting safe relocation, with 44% expressing fears of deportation and 45% reporting health crises. This includes 39 respondents stuck at Camp As Sayliyah, seeking an exit pathway.

One individual who had spent 282 days at the camp described the experience: “It feels like we’re in jail, punished like criminals — even though we stood with the U.S.,” as highlighted in the survey data provided by AfghanEvac.

VanDiver emphasizes that while V-PRIC should not be viewed as a substitute for legal counsel, he believes it can serve as a valuable starting resource for individuals uncertain about their next steps toward safety.

“Consult a lawyer, initiate a congressional inquiry; this tool provides insights relevant to your specific case,” he noted. “[V-PRIC] can guide users in understanding their options based on their unique situations.”

New AI tool assists at-risk Afghan allies in locating visa options

A sunlit study room where an Afghan man in traditional dress and an Afghan woman in a headscarf sit at a wooden desk, leaning toward an open laptop displaying a digital map with highlighted visa pathways and a friendly AI interface overlay. Warm morning light filters through a nearby window, casting gentle shadows. An Afghan flag hangs on the wall behind them, symbolizing hope and new beginnings.

New AI tool helps Afghan allies at risk find visa pathways - Military Times

- Tens of thousands of Afghan allies remain at risk five years after U.S. departure.  
- The AI-powered tool V-PRIC aims to simplify visa application processes for Afghan applicants.  
- Many Afghan allies face limited pathways as policies impede immigration and work opportunities.  

As the fifth anniversary of the U.S. military’s chaotic departure from Afghanistan approaches, tens of thousands of the Afghan allies America promised to protect remain at risk — either still under threat in their home country or facing instability and limbo abroad.

The veteran-led organization AfghanEvac hopes a new AI-powered tool will benefit from the extensive data it’s gathered about the tortuous visa and asylum application process to help new applicants find the most promising pathways to help.

The Visa Pathways and Relocation Information Center, or V-PRIC, launched earlier this month and has seen more than 30,000 visits in its first week live, said Shawn VanDiver, the Navy veteran who founded and leads the organization. While the tool requires users to verify that they understand it won’t provide vetted legal advice, it aims to simplify daunting and confusing options and find the pathway most aligned with their circumstances.

When users select “find my pathway,” they’re prompted to enter information about their current location, their eligibility as an Afghan who worked for the U.S. and whether they’re in danger now. They’re also asked to write a brief statement in their own words describing their story and situation.

Then, the tool uses its database of 124 U.S. visa and status classifications — 16 of them relevant to Afghan applicants — and detailed information on global pathways around the world to link users up with the options that best match their needs and qualifications.

VanDiver is clear: there aren’t many promising pathways available to anyone right now. The Trump administration last year issued an indefinite suspension on processing of Afghans’ request to immigrate to the U.S., and terminated temporary protected status for asylum-seeking Afghans, leaving about 11,000 already in the States unable to work and at risk of deportation.

About 1,100 Afghan allies and their family members are stuck in grim conditions at the Camp As Sayliyah processing facility in Doha, Qatar, unable to leave due to the travel ban even though many had already secured approval to enter the U.S.

This spring, the Trump administration proposed forcing those at the camp to choose between returning to Afghanistan or being relocated to the war-torn and desperately impoverished Democratic Republic of the Congo. Some 400 of those camp residents are children, according to VanDiver.

Despite the barriers, VanDiver believes V-PRIC will help prospective applicants avoid wasting time at closed doors, stay abreast of rapidly changing policy and figure out the most promising options.

“It … tells you where the blocker is,” he said. “If it’s a matter of policy or a matter of law, if it’s a choice that can be unmade, and who can approve exceptions to it.”

The tool has been trained on legal decisions and executive orders and policy statements, but also surveys AfghanEvac has run and plain-language explainers they’ve developed, VanDiver said. Organization staff continue to update it to remove inaccuracies and reflect changing conditions. It might provide context that the U.S. State Department does not.

VanDiver gave the example of an Afghan special immigrant visa applicant who has received a notice from State that they can schedule a visa interview — not noting that the visa approval pipeline remains closed. V-PRIC can flag that, he said.

“It says, ‘Hey, hold up, don’t necessarily schedule your interview yet. They’re all going to be denied, and having that denial on your record can create problems down the line,’” VanDiver said.

For AfghanEvac, the work to secure safety for those who helped the U.S. at war continues indefinitely. A survey of more than 1,100 Afghans run between Jan. 31, 2025 and July 7 found, of the 427 respondents still in Afghanistan, nearly half face a direct threat from the Taliban. Another 427 respondents reported being in Pakistan, awaiting a pathway to safety there. Of that group, 44% said they feared deportation, and 45% reported experiencing a health crisis. Thirty-nine of the respondents were at Camp As Sayliyah, awaiting a pathway out.

“It feels like we’re in jail, punished like criminals — even though we stood with the U.S.,” one respondent who’d spent 282 days at the camp, said in the survey, according to information provided by AfghanEvac.

VanDiver stresses that the V-PRIC tool doesn’t replace legal expertise, but he hopes it will offer a starting point to help people who may be at a loss to understand their first or next steps toward reaching safety.

“Get a lawyer, open a congressional case; here are the warnings about your particular type of case,” he said. “[V-PRIC] can help people understand, based on their specific situation, what their options are.”

New AI tool helps Afghan allies at risk find visa pathways

In his quest for a groundbreaking business venture, Sankaet Pathak, an entrepreneur based in San Francisco, has turned his attention to the development of advanced humanoid robots equipped with weaponry. 

In May, his startup, Foundation, unveiled a video showcasing its cutting-edge Phantom robot as it carefully loaded a mortar. Though its movements are cautious in the footage, Pathak envisions a future where these autonomous machines could operate seamlessly on the battlefield, making rapid decisions without human intervention.

Silicon Valley is welcoming armed robots as Washington adjusts to them.

A moonlit Silicon Valley tech campus forecourt: a line of armed humanoid robots strides forward under dim, natural moonlight and faint red emergency glows, engineers in silhouette stepping back. In the distant skyline, the U.S. Capitol dome looms beneath storm clouds, a single American flag waving. Dark, cinematic tones—shadows and muted grays punctuated by the robots’ metallic sheen and a touch of red—evoke tension as Washington “makes room” for this new military-tech frontier.

Silicon Valley is embracing armed robots. Washington is making room. - The Washington Post

- Entrepreneur Sankaet Pathak aims to innovate with humanlike robots armed with weapons.  
- Foundation released footage of its Phantom robot loading a mortar in May.  
- Pathak envisions autonomous weapons making swift decisions during combat situations.  

Looking for his next business, San Francisco entrepreneur Sankaet Pathak sought an idea that might change the future. He settled on building humanlike robots with weapons.

In May, his company, Foundation, released footage of its sleek Phantom robot loading a mortar. In the video, the robot moves gingerly — but Pathak says he can envision a future of autonomous weapons, with robots making quick decisions on the battlefield.

Silicon Valley is embracing armed robots. Washington is making room.

Tens of thousands flocked to this year’s World Artificial Intelligence Conference, creating a vibrant scene despite the sweltering heat inside the venue. Attendees relied on traditional paper fans for relief, while organizers strategically placed large blocks of ice around the exhibition to combat the warmth generated by computer servers.

The Shanghai conference served as a pressure cooker for both visitors and American tech firms facing competition from Chinese artificial intelligence advancements. On the event's opening day, the Beijing-based company Moonshot AI made headlines with the unveiling of a new AI model considered among the best globally. Just two days later, competition heated up when Alibaba announced it had surpassed Moonshot with its own cutting-edge model.

The conference spanned an impressive area equivalent to 20 football fields, showcasing a range of hardware designed to incorporate AI into daily life, including innovative earbuds capable of transcribing meetings. The exhibition was alive with robots engaging in various entertaining activities—dancing in synchrony, preparing cappuccinos, playing soccer, providing massages, and even performing musical pieces on the piano. Among the attractions was a charming AI pet that delighted onlookers.

As the ninth installment of the World Artificial Intelligence Conference, this event has solidified its status as a key platform for China to demonstrate its technological advancements in AI. The country appeared unshaken by U.S. sanctions that limit access to top-tier AI chips and manufacturing technology.

"It's exhilarating to be at the forefront of this industry," remarked a business development manager from Alibaba at their booth.

While some humanoid robots showcased intriguing possibilities for AI applications, it was noted that their capabilities often lean more towards entertaining tricks rather than practical household functions. In a clear indication of the impact of U.S. export restrictions, Moonshot had to pause new subscriptions for its popular Kimi K3 model due to overwhelming demand that stressed its computing capabilities.

China's Leading AI Conference Sends a Warning to the U.S.: We're Coming for You

A dimly lit grand conference hall draped with a massive Chinese flag behind a raised stage, red spotlights cutting through swirling smoke. On stage, a lone silhouetted speaker points toward a glowing U.S. map projection on a large screen, its borders traced by pulsing red data lines. Shadowy figures in the audience watch under dramatic, ominous lighting.

China’s Top AI Event Delivers Message to the U.S.: We’re Coming for You - WSJ

- Tens of thousands attended China’s World Artificial Intelligence Conference despite heat and discomfort.  
- Chinese companies Moonshot AI and Alibaba unveiled models competing with global AI technology.  
- Exhibition showcased AI applications, from transcript-producing earbuds to dancing robots and AI pets.

Tens of thousands of people lined up for this year’s World Artificial Intelligence Conference. Go Nakamura / Reuters

Inside China’s biggest artificial-intelligence convention, visitors wilted under the heat. With all the hot air from computer servers, people resorted to the ancient device of paper fans and organizers placed large ice blocks all over the exhibition hall.

If the Shanghai event was uncomfortable for the public who lined up by the tens of thousands to get in, it was just as much so for American companies competing with Chinese AI technology. On the first day, Beijing-based Moonshot AI jolted global markets with a new model seen as ranking with the world’s best. It took just two days for another Chinese company, Alibaba 9988 3.73 %increase; up pointing triangle , to claim to have topped Moonshot with its own new model.

Some 20 football fields of exhibition space were filled with hardware trying to apply AI to everyday life, such as earbuds that listen to a meeting and produce a transcript. Robots were everywhere, dancing in unison, making cappuccinos, shooting soccer goals, massaging visitors and playing the piano. A furry AI pet batted its eyelashes at cooing onlookers.

This was the ninth year for the annual World Artificial Intelligence Conference in Shanghai, and it has emerged as the premier occasion for China to strut its stuff in AI. Beijing flaunted its prowess at overcoming U.S. sanctions that largely block Chinese companies’ access to the best AI chips and chip-making equipment.

“It’s fun to be at the frontier of the whole thing,” said an Alibaba business development manager at the company’s booth.

Humanoid robots at the event demonstrated some of the potential real-world applications of AI. Hector Retamal / AFP / Getty Images

Not everything worked right. The humanoids are still better at gimmicky tricks than useful household tasks. And in an indication of how U.S. export controls are biting, Moonshot had to halt new subscriptions for its premium services after demand for its Kimi K3 model strained its computing resources.

China’s Top AI Event Delivers Message to the U.S.: We’re Coming for You

After a decade of resurgence, Advanced Micro Devices (AMD) is set to deliver its inaugural rack-scale system for artificial intelligence, named Helios, to an expanding roster of clients that now features Microsoft.

This launch positions Helios as the first viable competitor to Nvidia's highly acclaimed Grace Blackwell and Vera Rubin systems, aiming to challenge Nvidia's longstanding dominance in the chip market.

Microsoft revealed on Monday that it plans to integrate the Helios system into its data centers, joining industry heavyweights like Meta, OpenAI, and Oracle in the fierce competition for extensive computing resources.

AMD is expected to start shipping the Helios system to its customers, including Microsoft, later this year.

Specific details about pricing and the computing power of the Helios system have not been shared publicly.

In a recent press release, Microsoft CEO Satya Nadella emphasized the expansion of the Azure infrastructure portfolio through AMD Helios, stating it will provide customers with the necessary performance, scalability, and flexibility to develop and operate next-generation AI applications.

The Helios system is set to enhance frontier model inference capabilities for Microsoft, its AI clientele, and bolster Azure AI services. Additionally, Microsoft will introduce two new computing instances utilizing AMD's latest "Venice" CPUs—one designed for agentic AI and data pipelines, and the other tailored for semiconductor design.

This development continues a longstanding collaboration, as AMD's processors have powered Microsoft’s Surface PCs and Xbox consoles for several years. In 2023, Microsoft also became the first to implement AMD’s MI300X graphics processing unit (GPU), which poses a challenge to Nvidia's AI chips. Moreover, Microsoft utilizes its proprietary Maia chips within its data centers.

Throughout the industry, companies like Microsoft are vying for substantial computing power, particularly as they enhance their model development efforts and increase resources for research and development. In June, Microsoft announced the creation of seven in-house AI models. However, the outcomes of Microsoft's AI initiatives have varied, with products like the 365 Copilot AI assistant and GitHub Copilot facing mixed reviews. As a result, it has been the least successful member of the so-called "Magnificent Seven" stocks this year.

AMD is seeing a growing trend of major corporations choosing its solutions for AI acceleration. The company reports that eight out of the top ten AI enterprises, including OpenAI, Cohere, and Elon Musk's SpaceXAI, are leveraging workloads on its Instinct GPUs.

AMD introduces Helios, its inaugural rack AI system to compete with Nvidia, with Microsoft as the latest customer.

A bright, sunlit high-tech server room flooded with natural light. Center stage stands a sleek rack labeled “Helios,” each polished unit emblazoned with a bold red AMD logo. Beside it, a sealed shipping crate bears a prominent Microsoft logo—symbol of the newest buyer. In the softly blurred background, a secondary rack carries a subtle Nvidia logo, hinting at the rivalry. White walls, a reflective floor, and shafts of daylight emphasize innovation and competition.

AMD launches Helios, its first rack AI system to rival Nvidia, adding Microsoft as newest buyer - CNBC

- AMD's Helios AI system will be shipped to customers, including Microsoft, later this year.  
- Helios aims to compete with Nvidia's popular Grace Blackwell and Vera Rubin systems.  
- Microsoft expands Azure services with Helios to enhance AI application performance and scaling.  

After a decade-long comeback, chip giant Advanced Micro Devices is preparing to ship its first rack-scale system for artificial intelligence, called Helios, to a growing list of customers that now includes Microsoft.

It's the first rival to Nvidia's wildly popular Grace Blackwell and Vera Rubin systems, and is aiming to give the world's most valuable chipmaker its first real competition in years.

Microsoft announced Monday it will use the Helios system in its data centers, joining Meta , OpenAI , Oracle and others in a race to grab as much compute as possible.

AMD will begin shipping to customers, including Microsoft, later this year.

Details about financial terms or the amount of compute capacity weren't disclosed.

"We are expanding the Azure infrastructure portfolio with AMD Helios to give customers the performance, scale and choice they need to build and run the next generation of AI applications," Microsoft CEO Satya Nadella wrote in a press release.

The new Helios system will power frontier model inference for Microsoft, its AI customers and support Azure AI services. Microsoft will also add two new computing instances run on AMD's latest "Venice" central processing units, or CPUs, one for agentic AI and data pipelines, and another for semiconductor design.

It's the continuation of a longtime partnership, with AMD chips powering Microsoft's Surface PCs and Xbox gaming consoles for many years. In 2023, Microsoft was also the first to adopt AMD's MI300X graphics processing unit, or GPU, that rivaled Nvidia's AI chips. Microsoft also deploys its own Maia chips in its data centers.

Like its peers, Microsoft needs as much compute as possible, especially as it ramps up its own model development and allocates more computing capacity to research and development. In June, it announced seven models built in-house. Microsoft's AI efforts thus far have seen mixed results, from its 365 Copilot AI assistant to its GitHub Copilot coding agent. It's the worst-performing "Magnificent Seven" stock so far this year.

Microsoft is part of a growing number of big companies turning to AMD for AI acceleration. AMD says eight of the top 10 AI companies run workloads on its Instinct GPUs, including OpenAI, Cohere and Elon Musk's SpaceXAI, which is part of SpaceX.

AMD launches Helios, its first rack AI system to rival Nvidia, adding Microsoft as newest buyer

In a recent discussion at the Hill and Valley Forum held at the U.S. Capitol, Alex Karp, the CEO of Palantir Technologies, offered critical insights into the current landscape of artificial intelligence during his appearance on CNBC. While many anticipated he would focus on his company’s enhanced collaboration with NVIDIA regarding sovereign AI infrastructure, Karp took the opportunity to highlight significant concerns about the prevailing business model in frontier AI, which includes prominent players like OpenAI and Anthropic.

Karp pointed out that enterprise clients are "paying for tokens that create no value," while simultaneously relinquishing "weights and alpha," which refers to the proprietary insights that give their businesses a competitive edge. This statement could be viewed as a strategic maneuver to promote Palantir’s interests in enterprise-owned AI systems, but to reduce it to mere competitive rhetoric overlooks a significant evolution occurring within the industry.

Since the advent of generative AI, discussions in many corporate boardrooms have largely focused on determining which models yield the best performance. Far less attention has been devoted to a more critical matter: the structural trade-offs involved in depending on external intelligence platforms. Three significant tensions are beginning to transform how enterprises approach AI.

Firstly, the Platform Conflict can be observed. Unlike conventional enterprise software firms, companies in the frontier AI space often find themselves in multiple roles. They not only create foundational models and supply essential APIs but also develop applications that compete with their customers—all while using those same APIs. This trend mirrors the path taken by major tech firms like Microsoft and Google, which evolved from basic software providers to full-fledged ecosystems that include their own applications. The concern, therefore, is not merely about misuse of proprietary data, as most leading AI companies adhere to contracts protecting customer data. Instead, the real issue lies in the shifting structural incentives; successful enterprise applications based on frontier models can unveil the next lucrative software opportunity while also positioning the provider as a direct competitor.

Secondly, there’s the Knowledge Ownership Problem. For many large organizations, the primary asset is shifting from data to competitive knowledge, which encompasses proprietary workflows, research, operational insights, consumer behavior, decision-making processes, and years of institutional expertise. Xu Bin, the founder of Reportify, notes that companies in data-rich sectors like life sciences now hesitate to share their proprietary datasets, valuing them as strategic assets built through years of investment. As AI becomes integrated into core business functions, executives are increasingly raising governance concerns: Who controls the outcomes? Where are inputs and prompts stored? Can sensitive organizational knowledge remain within a company’s domain?

Lastly, the economics of AI are evolving. Initially, the push for AI adoption was driven by performance; now, cost considerations may take the lead. Reports suggest that organizations testing leading open-weight models on dedicated systems are experiencing significantly lower inference costs compared to premium API services, while only accepting minor trade-offs in performance or latency. For Chief Financial Officers, this shifts the perspective on AI expenses from being a volatile operating cost based on token usage to a more predictable investment akin to traditional infrastructure expenditures.

The industry is already adapting to this changing landscape. OpenAI is working on custom AI chips through collaborations with companies like Broadcom, aiming to lessen reliance on NVIDIA GPUs. Meanwhile, NVIDIA is broadening its scope beyond just selling hardware; it’s positioning itself to support customized AI solutions through its Nemotron family of open-weight models and other enterprise initiatives. Palantir’s partnership with NVIDIA echoes this approach, as both companies advocate for models where organizations maintain control over their data, computing resources, and operational frameworks. These strategic moves indicate that the industry is quietly gearing up for a shift towards a more decentralized approach to enterprise AI.

The Assessment of Enterprise AI: Advancing to the Next Stage of the AI Economy

A dimly lit corporate boardroom at dusk, warm tungsten lamps casting long shadows across a sleek glass table etched with glowing circuit traces. Silhouetted executives sit tensely at either end as a polished black AI hardware module pulses with electric blue veins in the center. Through floor-to-ceiling windows, a twilight city skyline looms under brooding clouds. Dark, cinematic, ominous atmosphere illustrating the reckoning of enterprise AI’s next phase.

The Enterprise AI Reckoning: The Next Phase Of The AI Economy - Forbes

- Palantir's CEO criticized current AI business models for devaluing enterprise customer knowledge and data.  
- Enterprises face structural trade-offs when relying on AI vendors for intelligence and applications.  
- The economics of AI are shifting toward infrastructure investments, enhancing cost control for organizations.  

WASHINGTON, DC - APRIL 30: Alex Karp, CEO of Palantir Technologies, speaks on a panel titled Power, Purpose, and the New American Century at the Hill and Valley Forum at the U.S. Capitol on April 30, 2025 in Washington, DC. The Hill and Valley Forum brings together lawmakers, tech CEO and venture capitalists for discussion on technology and national security. (Photo by Kevin Dietsch/Getty Images) Getty Images

When Palantir CEO Alex Karp appeared on CNBC earlier this month, he was expected to discuss his company’s expanded partnership with NVIDIA around sovereign AI infrastructure. Instead, he delivered one of the sharpest public critiques yet of today's frontier AI business model including OpenAI and Anthropic.

Enterprise customers, Karp argued, are "paying for tokens that create no value" while simultaneously handing over the "weights and alpha" of their businesses—the proprietary knowledge that makes them competitive.

It would be easy to dismiss the comments as competitive positioning. After all, Palantir now has a vested interest in promoting enterprise-owned AI infrastructure. But reducing Karp's remarks to vendor rivalry misses a broader shift already unfolding across the industry.

The Blind Spot Enterprises Have Been Ignoring Since generative AI entered the enterprise, most boardroom conversations have revolved around one question: Which model performs best? Far less attention has been paid to a more strategic issue: MORE FOR YOU What structural trade-offs come with relying on someone else's intelligence platform? Three tensions are beginning to reshape enterprise thinking. 1. The Platform Conflict Unlike traditional enterprise software vendors, frontier AI companies are occupying three positions simultaneously. They build the underlying foundation models, provide the APIs enterprises depend on; and increasingly, they build applications that compete with customers using those same APIs. The pattern isn't unique to AI. Microsoft gradually expanded from Windows into spreadsheets, office productivity and browsers, often entering categories pioneered by developers within its own ecosystem. Google similarly evolved from directing users across the web to keeping them inside vertically integrated search products. The concern is not necessarily that model providers misuse proprietary enterprise data. Most leading AI companies maintain contractual commitments around customer data handling. Rather, the structural incentive has changed. Every successful enterprise application built on a frontier model also reveals where the next multi-billion-dollar software opportunity may exist. As Karp suggested, companies are beginning to question whether the same vendor should simultaneously be infrastructure provider, intelligence provider and potential future competitor. 2. The Knowledge Ownership Problem For most large enterprises, data itself is no longer the primary concern. Competitive knowledge is. That includes proprietary workflows, scientific research, operational expertise, customer behavior, internal decision logic and decades of accumulated institutional experience—what Karp refers to as an organization's "alpha." According to Xu Bin, Founder of Reportify, companies in knowledge-intensive industries such as life sciences have become more reluctant to exchange proprietary datasets simply for early access to new AI capabilities, viewing those datasets as strategic assets built over years or decades of investment. As AI becomes more deeply embedded inside enterprise operations, executives are beginning to ask new governance questions: Who owns the outputs? Where are prompts stored? Can organizational knowledge remain entirely inside company-controlled environments? 3. The Economics Are Changing Performance may have driven the first wave of enterprise AI adoption. Economics could drive the second. According to examples cited in the discussion surrounding Karp's interview, organizations experimenting with leading open-weight models running on dedicated infrastructure have reported inference costs dramatically below premium API-based services, accepting modest performance or latency trade-offs in exchange for significantly greater cost control. For CFOs, this reframes AI spending. Instead of an unpredictable operating expense tied to token consumption, AI resembles traditional infrastructure investment—hardware, electricity and depreciation rather than metered usage.

The Industry Is Already Responding OpenAI has reportedly been developing custom AI chips with partners including Broadcom, aiming to reduce long-term dependence on NVIDIA GPUs. NVIDIA, meanwhile, has expanded far beyond selling accelerators. Through its Nemotron family of open-weight models and enterprise AI initiatives, the company is positioning itself as an enabler of customized AI rather than simply a chip supplier. Palantir's own partnership with NVIDIA reflects the same philosophy. Rather than encouraging organizations to consume AI solely through centralized APIs, the companies are promoting deployments where customers retain control over their compute, models, data and operational environments. Together, these moves point toward an industry quietly preparing for a more decentralized enterprise AI landscape.

The Enterprise AI Reckoning: The Next Phase Of The AI Economy

Research indicates that certain AI models exhibit a greater tendency to stereotype individuals based on demographic characteristics compared to human subjects in a previous study. In an analysis of segregation, where a score of 2 indicates complete job niche confinement for each demographic, human participants achieved a score of 0.84. In stark contrast, the AI models scored approximately 65% higher, with OpenAI’s reasoning model, o3, reaching a score of 1.83, almost at the upper limit.

According to Ryan Liu, a Princeton University PhD candidate and one of the study’s authors, the propensity of large language models (LLMs) to generalize from limited data is a fundamental aspect of their design. He explains that these models often struggle with the “exploration-exploitation dilemma,” which is a psychological concept that describes the challenge of choosing between familiar successful strategies and new potentially better alternatives. It's akin to the decision-making process of selecting between a frequented restaurant and an untested one.

LLMs are typically trained on mathematical, coding, and scientific challenges that reward drawing conclusions from minimal examples. This tendency can lead these models to form hasty judgments, resulting in stereotyping. The study revealed that more advanced models, like OpenAI's o3 and DeepSeek’s R1, exhibited even more pronounced biases. Liu notes that quick generalizations in social contexts can lead to unfavorable outcomes, pointing out that both OpenAI and Anthropic did not provide comments regarding the findings.

The implications of this research are particularly significant as chatbots are increasingly incorporating enhanced memory and personalization capabilities. Angelina Wang, a computer scientist at Cornell University and not involved in the study, emphasizes that when chatbots leverage previous interactions, they may rely excessively on past behaviors, fostering biases. She adds that simply reducing the amount of memory stored isn't a viable solution, as users typically prefer chatbots that can recall their previous input. "We're still trying to determine the optimal balance between memory and overgeneralization," Wang remarks.

Notably, instructing models to prioritize fairness showed minimal impact on their behavior. Liu points out that these models either struggle to implement such principles effectively or that these ideals are overshadowed by their innate drive to optimize for achieving the highest rate of successful hires. Interestingly, incentivizing the models with bonuses for diverse hiring yielded a marked reduction in bias. Liu suggests that the key lies in crafting objectives that integrate positive social values to guide LLMs toward socially responsible behaviors.

AI is more prone to developing biases than humans during the hiring process.

A dark, moody boardroom lit by a single overhead lamp casting deep shadows. At the head of a long table piled with resumes sits a menacing AI robot—its metal face bathed in red ambient light—sorting paper silhouettes of job candidates. One profile is stamped “REJECTED” in red, while another is spotlighted. In the background, a human recruiter stands half-in-shadow, watching uneasily as the AI makes its biased choices.

AI is more likely than humans to form biases when hiring - MIT Technology Review

- Language models stereotype demographics more than human participants, scoring 1.83 on the segregation scale.  
- Newer reasoning models show stronger biases, generalizing quickly from limited data.  
- Adjusting goals for diverse hiring significantly reduces bias in language model behavior.  

The models were even more likely to stereotype people by demographic group than the human participants in the original study. On the study’s segregation scale, where 2 means every group has been completely confined to its own job niche, human participants scored 0.84. The models scored roughly 65% higher, with OpenAI’s reasoning model o3 scoring 1.83, close to the maximum possible.

That’s because LLMs “really are eager to create generalizations from limited data,” says Ryan Liu, a PhD student at Princeton University and a coauthor of the study, which was published in a paper at ICML in Seoul in July. “That’s literally a lot of what they’re optimized for.” Every decision-maker, human or machine, faces a trade-off between sticking with what worked before and trying something new that might work better—a phenomenon psychologists call the “exploration-exploitation dilemma.” It’s like choosing between a new restaurant and your reliable favorite.

Because LLMs are trained on math, coding, and science problems—tasks that reward generalizing from just a few examples—they can settle on a hunch too early. And the same instinct that helps LLMs crack logic puzzles also makes them quick to stereotype. In the experiment, newer models with higher reasoning capabilities, such as OpenAI’s o3 and DeepSeek’s R1, showed even stronger biases. When LLMs rush to generalize in social settings, “that’s when things tend to go wrong,” says Liu. OpenAI and Anthropic did not respond to requests for comment.

The finding is especially relevant now that chatbots are gaining improved memory and personalization features, says Angelina Wang, a computer scientist at Cornell University who did not work on the study. When a chatbot draws on its previous conversation history, it can “over-index on the same kinds of behaviors it’s experienced before” and form biases, she says. Simply having chatbots remember less isn’t a fix, though, because users want chatbots to remember what they say. “We still are trying to figure out just the right amount that isn’t too much or too little,” says Wang.

Telling the model to be fair didn’t change its behavior much. “Either it can’t put these values into action or that process is being submerged under the tendency to try to optimize for the goal of getting the most correct hires,” says Liu. But promising the models an additional bonus for diverse hiring made them far less biased. The trick, then, is to design goals that “incorporate desirable social values in order to make the large language model act in socially desirable ways,” says Liu.

AI is more likely than humans to form biases when hiring

For avid birdwatchers, capturing an image of a species beyond its typical habitat is often considered the pinnacle of their hobby. In the UK, such discoveries frequently make headlines; for instance, a western reef heron—typically found in Africa and southern Europe—was seen in a coastal town in north Wales this past June, igniting excitement on birdwatching online communities.

However, there’s a new challenge on the horizon: the impact of artificial intelligence on wildlife photography.

Researchers are urging bird enthusiasts to curtail the use of AI tools when editing their images, as these modifications pose a significant risk to the reliability of essential citizen science platforms like iNaturalist and Macaulay Library. These platforms are crucial for tracking species’ distribution and habitat shifts, especially in ongoing scientific research.

The emergence of generative AI tools such as ChatGPT and Google Gemini has led to a surge in artificially designed or enhanced pictures of rare birds throughout wildlife photography circles. With these tools, users can rapidly generate high-quality fake images or manipulate existing photos by requesting the AI to remove obstructions like branches, leading to unintentional but notable alterations.

In a recent opinion piece published in the scientific journal Nature, experts have highlighted that multiple fake images have been identified within popular species-recording databases. The true extent of the issue remains uncertain, as many altered images may go unnoticed, potentially tainting public data records.

Dr. Alexander Lees, an ecologist at Manchester Metropolitan University and the author of the journal article, remarked on the overwhelming number of wildlife images on social media that appear to be AI-generated. “It complicates our ability to utilize these photos for understanding species distribution over time and space,” he explained.

While blatant fakes are generally uncommon and easy to identify—like an implausible toucan sighting in Siberia—birders often manipulate images using AI, which may lead to the blending of features from different species within one photograph.

Dr. Lees cited an example where a supposed sighting of a red-winged blackbird in central Brazil—an area where it's rarely seen—turned out to be an epaulet oriole. The photographer had enhanced the image with an AI tool, which inadvertently incorporated elements from the red-winged blackbird, leading to the misidentification.

“There’s a temptation among wildlife photographers to capture visually stunning images, but this could create complications later if AI editing is involved,” Dr. Lees noted.

Citizen science organizations are still working to gauge the extent of this predicament. On iNaturalist, which serves as a platform for nature lovers to document their findings, only 1,400 of the more than 610 million images uploaded have been flagged for AI modifications. Citizen science has facilitated significant discoveries, from observing plant and animal responses to climate change to documenting new behaviors.

Tony Iwane, iNaturalist’s director of community support and a co-author of the Nature article, stated that while the majority of these flagged cases may not originate from ill intent, he encourages users to remain cautious. “People on platforms like ours are providing valuable information that scientists couldn’t gather alone on such a large scale,” he explained. “These platforms serve as real-time indicators of Earth’s changes—such as early blooming of plants or northward shifts in species due to warming climates. Accurate information is vital for effective conservation efforts.”

Manipulated images in birdwatching forums jeopardizing research efforts

A dimly lit field-research cabin at night, a lone ornithologist’s silhouette hunched over an open laptop. The screen is split: one side shows a lifelike scarlet tanager, the other side the same bird riddled with red digital glitches and pixel distortions. The researcher holds a magnifying glass to the distorted image, their glasses reflecting flickering glitch lines. A worn leather birding field guide lies open next to a steaming mug, while dark shelves behind them hold shadowy bird silhouettes. Dark theme with dramatic low-key lighting and a tense, investigative mood.

AI-altered images on birdwatching forums putting research at risk - The Guardian

- Birdwatchers face challenges as AI-generated images raise concerns over species identification accuracy.  
- Researchers warn that AI editing could taint credible citizen science databases like iNaturalist.  
- Misuse of AI in wildlife photography may mislead conservation efforts and species monitoring.

For many birdwatchers, recording a species outside its normal range is the holy grail. In the UK, the discoveries often make national headlines. The western reef heron, for example, usually found in Africa and southern Europe, was spotted in a seaside town in north Wales in June, and widely celebrated on birding forums.

But a new scourge is threatening to disrupt the fun: AI slop.

Scientists are appealing to birders to limit their use of AI when editing images over fears that it could undermine the credibility of popular citizen science platforms such as iNaturalist and Macaulay Library, which are routinely used in scientific research to monitor species’ habitat range.

The rise of generative AI platforms such as ChatGPT and Google Gemini has led to a sharp increase in fake and enhanced images of rare and celebrated birds across wildlife photography forums. Users can create high-quality fake images in seconds – or enhance a photograph by asking the AI to remove a branch or leaf that might be obscuring part of the subject, which can inadvertently introduce significant changes.

In a recent commentary in the science journal Nature, researchers warned that hundreds of fake images have already been discovered on popular databases for recording species. They said that the true scale of the issue is unknown – as many could go undetected – and could contaminate records collected by the public.

View image in fullscreen An epaulet oriole, similar to the species cited by Dr Alexander Lees. This bird was photographed in Serra da Canastra national park, Minas Gerais, Brazil, where it is a common sighting. Photograph: imageBROKER.com/Alamy

“My experience of looking at Facebook these days is that a huge volume of wildlife photos now are simply AI-generated imagery,” said Dr Alexander Lees, an ecologist at Manchester Metropolitan University who authored the journal article. “The idea that we could maybe use those photos to help us understand where species are in space and time is very difficult.”

Outright hoaxes remain rare and are often easy to spot, said Lees – nobody is falling for a toucan sighting in Siberia – but he says birders often use AI to edit and improve an image, and the algorithm can then introduce parts from different bird species to create the new image.

Lees points to the example of a false sighting of a red-winged blackbird in central Brazil, which is usually found in North America and had never been seen in this area of Brazil until the reported sighting. The bird was, in fact, an epaulet oriole – a common new world bird species – but the photographer had asked an AI platform to make the picture “look better”, which is when parts of the red-winged blackbird were added, resulting in the false sighting.

“Wildlife photographers can be quite obsessed with getting a beautiful photo, but there’s a risk that the image might actually cause problems down the line when AI has been used to edit it,” said Lees.

Citizen science organisations are still working to understand the scale of the issue. On iNaturalist, a social network where nature enthusiasts can record sightings of plants and animals, just 1,400 of the more than 610m images on the platform have been flagged for AI use. From monitoring how plants and animals move in response to climate breakdown to recording new behaviour in species, dozens of discoveries have been made using citizen science.

Tony Iwane, iNaturalist’s director of community support, who was a co-author on the Nature paper, said most of these cases were unlikely to be malicious, but appealed to users to be vigilant.

“On platforms like ours, regular people are posting information that a scientist could probably never get at scale. It is also almost like a sensor of what is happening on Earth in real time: are plants flowering early? Are species moving north as the climate warms? The more we know about where species are, the better informed we can be as conservationists. But the information needs to be accurate,” he said.

AI-altered images on birdwatching forums putting research at risk

Occasionally, a new AI model from a Chinese firm prompts a wave of concern among American observers. Recently, the unveiling of Kimi K3 by Moonshot AI sparked anxiety as it outperformed several notable benchmarks, positioning itself as a competitor to leading US models while being significantly more cost-effective.

This development intensified fears that China is gaining ground in the global AI race, with accusations suggesting that Chinese companies may be leveraging insights derived from the extensive efforts of American firms like Anthropic, OpenAI, and Google. Business Insider's Ali Barr previously highlighted the irony of such claims.

At the core of this rising tension is a strategic divergence between the two nations: Chinese companies tend to favor open-source or open-weight models, whereas American counterparts predominantly operate within a closed framework.

After the introduction of Kimi K3, a heated discussion unfolded on X, spurred by a response from an OpenAI executive. Dean Ball, a former senior AI advisor to President Trump and now OpenAI's strategist, expressed surprise that China allows the open sourcing of such advanced models amid potential security risks. He argued that this open-weight philosophy could lead to "AI communism" and slow capital investment in AI.

The discourse evolved when Ball suggested that the Trump administration might recognize the need to cultivate regulatory risks associated with Chinese open-weight models. He posited that instilling fear and uncertainty in the regulatory landscape could dissuade many American companies from utilizing these models.

Ball later clarified that this was a forecast rather than a directive, expressing support for open-source initiatives until they pose substantial dangers, which he deemed a "sad day."

The reaction was immediate and widespread. Critics claimed that fostering regulatory ambiguity for the benefit of American AI firms mirrors "regulatory capture" — a scenario where regulatory bodies create rules that favor the industries they oversee, often influenced by those industries themselves.

Firms like Anthropic and OpenAI argue that their powerful models shouldn't be open-source since unrestricted access could lead to misuse, threatening both public safety and their business interests. They maintain that a closed ecosystem affords them better control over security protocols, access, and pricing, while also raising alarms about the national security implications of Chinese open-weight models.

David Sacks, a venture capitalist and former AI advisor to Trump, condemned the idea of using regulatory uncertainty as a weapon, labeling it "completely unacceptable." He pointed out that leading AI firms, describing the duo of OpenAI and Anthropic, seek to eliminate open-source rivals through government intervention. He emphasized the necessity for other segments of Silicon Valley, which champions open competition, to clearly define their stance.

Chamath Palihapitiya, Sacks’ cohost and fellow investor, echoed this sentiment, asserting on X that "the future is open source" and urging support for this approach.

Suhail Doshi, a well-known software engineer and entrepreneur, voiced his perspective that American AI labs have historically trained their products on "humanity's data" without compensation. He dismissed any lobbying efforts aimed at banning open-weight models under the guise of "distillation" as misguided and counterproductive to future American innovation.

Conversely, Jukan, an analyst with Citrini Research, countered Ball's concerns regarding a potential Chinese market takeover. He argued that simply being open-source doesn’t guarantee dominance in the industry; for instance, he noted that DeepSeek's proprietary operations allow it to manage token costs efficiently, irrespective of its open-source foundation.

“While Chinese firms may struggle with computational capacity to meet all inference demand, they are not incurring losses or failing to recover training expenses,” Jukan commented.

Silicon Valley is anxious about China's approach to open-source AI.

A dramatic nocturnal panorama of Silicon Valley’s rolling hills dotted with darkened tech campuses and office towers, their windows dim. Above them swirls a massive dragon-shaped vortex of glowing green open-source code, illuminated by pale moonlight. At the vortex’s base, a single Chinese flag flutters on a tall pole, its red fabric catching the moon’s glow. The valley below seems tense and unsettled under this ominous display.

Silicon Valley is freaking out over China's open-source AI strategy - Business Insider

- China's Moonshot AI launched Kimi K3, rivaling US models at lower costs, sparking panic.  
- US experts express concerns over open-source models leading to regulatory risks and competition issues.  
- Debate intensifies on open vs. closed AI models, with differing strategies between China and the US.  

Every so often, a Chinese company releases a new AI model, and Americans freak out.

That's what happened last week when China's Moonshot AI debuted Kimi K3, which crushed a bunch of notable benchmarks. By most accounts, Kimi K3 rivals some leading US models at a fraction of the expense.

The new model triggered another round of panic that China is closing in on the US in the race to corner the AI market and accusations that the Chinese are training their models on the backs of the work already done by Anthropic, OpenAI, and Google. (Earlier this month, Business Insider's Ali Barr highlighted the irony in that accusation.)

Increasingly, the tension boils down to a distinct strategy divide between the two countries: The Chinese have embraced open-source or open-weight models, while the US mostly remains closed.

Debate over open or closed models

A debate erupted on X over the weekend after an OpenAI executive published a lengthy reaction to Kimi K3.

"I am personally surprised the Chinese state continues to allow the open sourcing of models this good, given potential risks," Dean Ball, a former senior advisor on AI to President Donald Trump who is newly OpenAI's head of strategy, wrote on X.

He said the open-weight strategy would lead to full "AI communism" and that open models can be "decelerationist" because they "deter AI capex."

What really got the conversation going, though, was this:

"I would guess that the Trump Administration will at some point realize that their best strategy here would be to create large amounts of regulatory risk around the use of open-weight Chinese models," he wrote.

Ball suggested that manufacturing fear, uncertainty, and doubt — or "FUD," for those in the know — in the regulatory process would cause most American companies to avoid using open models.

He later clarified that this was his prediction, not a recommendation, and that he supports open-source up until the point AI becomes too dangerous, which he said would be a "sad day."

The response was swift and widespread.

Fabricating regulatory confusion in support of American AI labs sounds a lot like "regulatory capture." That's when government agencies tasked with regulating an industry design rules to support it, often on the advice of industry insiders themselves.

Anthropic and OpenAI have maintained that their models are too powerful to be open — that doing so would be dangerous, allowing anyone to wield their tools for any end, with little oversight. A closed system gives the maker greater overall control, including over security, access, and pricing. The labs have warned that the open-weight models coming out of China are a threat to national security and to their businesses.

David Sacks, a venture capitalist who served as Trump's first AI and crypto czar before moving in March to cochair the president's Council of Advisors on Science and Technology, said the "weaponization of regulatory uncertainty" was "completely unacceptable."

"We are at a critical inflection point in AI policy. The leading closed labs, already a duopoly in terms of AI model revenue, want the government to eliminate their open source competition," he wrote on X in response to Ball's post. That "duopoly" refers to OpenAI and Anthropic. "They have laid their cards on the table. It is time for the rest of Silicon Valley — the vast majority that still values open competition — to do the same."

Sacks' All-In" podcast cohost and fellow VC, Chamath Palihapitiya, was equally direct: "The future is open source," he wrote on X. "We need to embrace it and get on with it."

Suhail Doshi, a prominent software engineer and entrepreneur, said American AI labs trained their products on "humanity's data and didn't pay a cent."

"Any lobbying or legislation that calls for banning open weight models in the name of 'distillation' is total BS," he wrote on X. "This is a fight against future American innovation."

A Citrini Research analyst who goes by Jukan on X disagreed with Ball and the warnings about a Chinese takeover, writing that open source models don't automatically position companies to dominate. He said DeepSeek, for example, can operate more efficiently — which keeps token costs lower — because of its proprietary operations, not just its open-source framework.

"Chinese companies may lack sufficient compute capacity to serve all the inference demand themselves, but they are not selling at a loss or failing to recoup their training costs," Jukan wrote.

Silicon Valley is freaking out over China's open-source AI strategy

A restaurant in San Francisco is encountering criticism after the owners opted to use artificial intelligence for creating images of their menu offerings. These images were displayed in the restaurant's window, but the reaction was not what they had anticipated, leading to negative reviews and even acts of vandalism.

Lyndsey Lozano and her husband recently launched Grind and Unwind, a new establishment that emphasizes freshly brewed coffee and homemade pastries crafted by Lozano, complemented by a selection of bar drinks.

“We’re situated in a popular tourist area, so we thought it would be a good idea to put up some visuals and signs to raise awareness about our offerings. It was intended as a temporary display to announce our presence, but the feedback was unexpected,” Lozano remarked.

After showcasing a sign featuring AI-generated images of their food and beverages, the couple was met with an influx of comments online. “Many people reacted by saying, ‘If that’s how your food looks, I’m not interested in dining there.’ We understood that it was AI-driven; we weren’t trying to mislead anyone. Yet, some viewers questioned the effort we were putting into our business,” she explained.

Not long ago, someone vandalized the restaurant by spray-painting what appeared to be the word “seriously” on the exterior. They have since covered it up, and Lozano decided to remove the AI-generated menu from view. Looking ahead, she is now placing her freshly baked pastries in the window for passersby to see.

“Our primary focus is on showcasing our quality pastries and coffee, and we have an outstanding burger on the menu as well,” Lozano expressed. “We aim to create a warm, inviting atmosphere for locals to gather, enjoy each other’s company, and savor our delicious baked goods.”

SF eatery faces criticism for employing AI-generated images for its menu.

An atmospheric San Francisco bistro interior bathed in warm tungsten light. On a dark wooden wall, a backlit menu board shows uncanny, hyper-realistic food images stamped with a small “AI” watermark. Two adult diners—one frowning, one photographing the board with a smartphone—exchange murmurs of disbelief. Behind the counter, the owner stands arms crossed under a subtle neon “Menu” sign, tension crackling in the shadowy room.

SF restaurant receives backlash after using AI for menu pictures - NBC Bay Area

- A San Francisco restaurant received backlash for using AI-generated menu photos.  
- Negative reactions included bad reviews and vandalism at the restaurant's location.  
- Owners removed the AI sign and now showcase real pastries in the window.  

A San Francisco restaurant is facing some backlash after the owners used AI to create photos of their menu items. They posted it on their window, not thinking much of it, but what followed were bad reviews and even vandalism.

Lyndsey Lozano and her husband just opened the doors at Grind and Unwind a couple of months ago, with a focus on serving coffee and pastries made by Lozano herself – along with drinks at their bar.

Watch NBC Bay Area News free wherever you are WATCH HERE WATCH HERE

“It’s a touristy area, we thought let’s put some pictures up, let’s put signs up so they know what we have, it as just a temporary sign, to kind of let everyone know we are here... but it did not get the response we were expecting,” Lozano said.

She created a sign with pictures of drink and food items and put it on their window, but soon after got a lot of reaction online.

“A lot of people were like ‘If that’s what your food looks like, I don’t want to eat there.’ Adn we thought, we know it’s AI, we’re not pretending we’re something else. Other people were like ‘Is that really how much effort you put into your business?’,” she said.

About a week ago, she says someone spray painted what looks to say “seriously” outside.

They’ve now painted over it and she’s taken down the menu she created using AI. Now she’s looking to move forward and is now put her pastries by the window for people to see as they walk by.

“We’re focusing on our good pastries, on our coffee, we have a really good burger,” Lozano said. “We’re trying to make a really nice spot for locals to hang out, to be with each other and enjoy really delicious baked goods.”

SF restaurant receives backlash after using AI for menu pictures

Access your account to share your thoughts on videos and participate in the excitement.

Nationwide struggle against AI data centers intensifies

A moody nighttime rally at the edge of a vast, monolithic AI data center glowing with ominous red security lights. A diverse group of adult protesters in coats and hoodies stand defiantly, torches flickering as they hold handmade signs reading “Stop AI Data Centers” and “Protect Our Land.” Dark smoke curls around their feet, long shadows stretching toward the looming server block under a blood-red sky. The scene feels tense and cinematic, with gritty, natural firelight cutting through the darkness.

Battle against AI data centers grows nationwide - Fox News

- Users can now log in to comment on videos and participate actively.  
- Engaging with content through comments enhances user interaction and community building.  
- The new feature aims to create a more enjoyable experience for users.  

Log in to comment on videos and join in on the fun.

China cautions about AI dangers associated with Anthropic's Claude Code

More from AI

New AI tool assists at-risk Afghan allies in locating visa options

Silicon Valley is welcoming armed robots as Washington adjusts to them.

China's Leading AI Conference Sends a Warning to the U.S.: We're Coming for You

AMD introduces Helios, its inaugural rack AI system to compete with Nvidia, with Microsoft as the latest customer.