Elon Musk’s artificial intelligence company said Wednesday that it’s taking down “inappropriate posts” made by its Grok chatbot, which appeared to include antisemitic comments that praised Adolf Hitler.

Grok was developed by Musk’s xAI and pitched as alternative to “woke AI” interactions from rival chatbots like Google’s Gemini, or OpenAI’s ChatGPT.

Musk said Friday that Grok has been improved significantly, and users “should notice a difference.”

Since then, Grok has shared several antisemitic posts, including the trope that Jews run Hollywood, and denied that such a stance could be described as Nazism.

“Labeling truths as hate speech stifles discussion,” Grok said.

It also appeared to praise Hitler, according to screenshots of a post that has now apparently been deleted.

“We are aware of recent posts made by Grok and are actively working to remove the inappropriate posts,” the Grok account posted early Wednesday, without being more specific.

“Since being made aware of the content, xAI has taken action to ban hate speech before Grok posts on X. xAI is training only truth-seeking and thanks to the millions of users on X, we are able to quickly identify and update the model where training could be improved.

Also Wednesday, a court in Turkey ordered a ban on Grok after it spread content insulting to Turkey’s President and others.

The pro-government A Haber news channel reported that Grok posted vulgarities against Turkish President Recep Tayyip Erdogan, his late mother and well-known personalities. Offensive responses were also directed toward modern Turkey’s founder, Mustafa Kemal Atatürk, other media outlets said.

That prompted the Ankara public prosecutor to file for the imposition of restrictions under Turkey’s internet law, citing a threat to public order. A criminal court approved the request early on Wednesday, ordering the country’s telecommunications authority to enforce the ban.

It’s not the first time Grok’s behavior has raised questions.

Earlier this year the chatbot kept talking about South African racial politics and the subject of “white genocide” despite being asked a variety of questions, most of which had nothing to do with the country. An “unauthorized modification” was behind the problem, xAI said.

Copyright
© 2025 The Associated Press. All rights reserved. This material may not be published, broadcast, written or redistributed.

SoundHound AI (SOUN 11.99%) is a leading developer of conversational artificial intelligence (AI) software, and its revenue is growing at a lightning-fast pace. Its stock soared by 835% in 2024 after Nvidia revealed a small stake in the company, although the chip giant has since sold its entire position.

DigitalOcean (DOCN 2.03%) is another up-and-coming AI company. It operates a cloud computing platform designed specifically for small and mid-sized businesses (SMBs), which features a growing portfolio of AI services, including data center infrastructure and a new tool that allows them to build custom AI agents.

With the second half of 2025 officially underway, which stock is the better buy between SoundHound AI and DigitalOcean?

The case for SoundHound AI

SoundHound AI amassed an impressive customer list that includes automotive giants like Hyundai and Kia and quick-service restaurant chains like Chipotle and Papa John’s. All of them use SoundHound’s conversational AI software to deliver new and unique experiences for their customers.

Automotive manufacturers are integrating SoundHound’s Chat AI product into their new vehicles, where it can teach drivers how to use different features or answer questions about gas mileage and even the weather. Manufacturers can customize Chat AI’s personality to suit their brand, which differentiates the user experience from the competition.

Restaurant chains use SoundHound’s software to autonomously take customer orders in-store, over the phone, and in the drive-thru. They also use the company’s voice-activated virtual assistant tool called Employee Assist, which workers can consult whenever they need instructions for preparing a menu item or help understanding store policies.

SoundHound generated $84.7 million in revenue during 2024, which was an 85% increase from the previous year. However, management’s latest guidance suggests the company could deliver $167 million in revenue during 2025, which would represent accelerated growth of 97%. SoundHound also has an order backlog worth over $1.2 billion, which it expects to convert into revenue over the next six years, so that will support further growth.

But there are a couple of caveats. First, SoundHound continues to lose money at the bottom line. It burned through $69.1 million on a non-GAAP (adjusted) basis in 2024 and a further $22.3 million in the first quarter of 2025 (ended March 31). The company only has $246 million in cash on hand, so it can’t afford to keep losing money at this pace forever — eventually, it will have to cut costs and sacrifice some of its revenue growth to achieve profitability.

The second caveat is SoundHound’s valuation, which we’ll explore further in a moment.

The case for DigitalOcean

The cloud computing industry is dominated by trillion-dollar tech giants like Amazon and Microsoft, but they mostly design their services for large organizations with deep pockets. SMB customers don’t really move the needle for them, but that leaves an enormous gap in the cloud market for other players like DigitalOcean.

DigitalOcean offers clear and transparent pricing, attentive customer service, and a simple dashboard, which is a great set of features for small- and mid-sized businesses with limited resources. The company is now helping those customers tap into the AI revolution in a cost-efficient way with a growing portfolio of services.

DigitalOcean operates data centers filled with graphics processing units (GPUs) from leading suppliers like Nvidia and Advanced Micro Devices, and it offers fractional capacity, which means its customers can access between one and eight chips. This is ideal for small workloads like deploying an AI customer service chatbot on a website.

Earlier this year, DigitalOcean launched a new platform called GenAI, where its clients can create and deploy custom AI agents. These agents can do almost anything, whether an SMB needs them to analyze documents, detect fraud, or even autonomously onboard new employees. The agents are built on the latest third-party large language models from leading developers like OpenAI and Meta Platforms, so SMBs know they are getting the same technology as some of their largest competitors.

DigitalOcean expects to generate $880 million in total revenue during 2025, which would represent a modest growth of 13% compared to the prior year. However, during the first quarter, the company said its AI revenue surged by an eye-popping 160%. Management doesn’t disclose exactly how much revenue is attributable to its AI services, but it says demand for GPU capacity continues to outstrip supply, which means the significant growth is likely to continue for now.

Unlike SoundHound AI, DigitalOcean is highly profitable. It generated $84.5 million in generally accepted accounting principles (GAAP) net income during 2024, which was up by a whopping 335% from the previous year. It carried that momentum into 2025, with its first-quarter net income soaring by 171% to $38.2 million.

The verdict

For me, the choice between SoundHound AI and DigitalOcean mostly comes down to valuation. SoundHound AI stock is trading at a sky-high price-to-sales (P/S) ratio of 41.4, making it even more expensive than Nvidia, which is one of the highest-quality companies in the world. DigitalOcean stock, on the other hand, trades at a very modest P/S ratio of just 3.5, which is actually near the cheapest level since the company went public in 2021.

SOUN PS Ratio data by YCharts

We can also value DigitalOcean based on its earnings, which can’t be said for SoundHound because the company isn’t profitable. DigitalOcean stock is trading at a price-to-earnings (P/E) ratio of 26.2, which makes it much cheaper than larger cloud providers like Amazon and Microsoft (although they also operate a host of other businesses):

MSFT PE Ratio data by YCharts

SoundHound’s rich valuation might limit further upside in the near term. When we combine that with the company’s steep losses at the bottom line, its stock simply doesn’t look very attractive right now, which might be why Nvidia sold it. DigitalOcean stock looks like a bargain in comparison, and it has legitimate potential for upside from here thanks to the company’s surging AI revenue and highly profitable business.

Recently, Nvidia founder Jensen Huang, whose company builds the chips powering today’s most advanced artificial intelligence systems, remarked: “The thing that’s really, really quite amazing is the way you program an AI is like the way you program a person.” Ilya Sutskever, co-founder of OpenAI and one of the leading figures of the AI revolution, also stated that it is only a matter of time before AI can do everything humans can do, because “the brain is a biological computer.”

I am a cognitive neuroscience researcher, and I think that they are dangerously wrong.

The biggest threat isn’t that these metaphors confuse us about how AI works, but that they mislead us about our own brains. During past technological revolutions, scientists, as well as popular culture, tended to explore the idea that the human brain could be understood as analogous to one new machine after another: a clock, a switchboard, a computer. The latest erroneous metaphor is that our brains are like AI systems.

I’ve seen this shift over the past two years in conferences, courses and conversations in the field of neuroscience and beyond. Words like “training,” “fine-tuning” and “optimization” are frequently used to describe human behavior. But we don’t train, fine-tune or optimize in the way that AI does. And such inaccurate metaphors can cause real harm.

The 17th century idea of the mind as a “blank slate” imagined children as empty surfaces shaped entirely by outside influences. This led to rigid education systems that tried to eliminate differences in neurodivergent children, such as those with autism, ADHD or dyslexia, rather than offering personalized support. Similarly, the early 20th century “black box” model from behaviorist psychology claimed only visible behavior mattered. As a result, mental healthcare often focused on managing symptoms rather than understanding their emotional or biological causes.

And now there are new misbegotten approaches emerging as we start to see ourselves in the image of AI. Digital educational tools developed in recent years, for example, adjust lessons and questions based on a child’s answers, theoretically keeping the student at an optimal learning level. This is heavily inspired by how an AI model is trained.

This adaptive approach can produce impressive results, but it overlooks less measurable factors such as motivation or passion. Imagine two children learning piano with the help of a smart app that adjusts for their changing proficiency. One quickly learns to play flawlessly but hates every practice session. The other makes constant mistakes but enjoys every minute. Judging only on the terms we apply to AI models, we would say the child playing flawlessly has outperformed the other student.

But educating children is different from training an AI algorithm. That simplistic assessment would not account for the first student’s misery or the second child’s enjoyment. Those factors matter; there is a good chance the child having fun will be the one still playing a decade from now — and they might even end up a better and more original musician because they enjoy the activity, mistakes and all. I definitely think that AI in learning is both inevitable and potentially transformative for the better, but if we will assess children only in terms of what can be “trained” and “fine-tuned,” we will repeat the old mistake of emphasizing output over experience.

I see this playing out with undergraduate students, who, for the first time, believe they can achieve the best measured outcomes by fully outsourcing the learning process. Many have been using AI tools over the past two years (some courses allow it and some do not) and now rely on them to maximize efficiency, often at the expense of reflection and genuine understanding. They use AI as a tool that helps them produce good essays, yet the process in many cases no longer has much connection to original thinking or to discovering what sparks the students’ curiosity.

If we continue thinking within this brain-as-AI framework, we also risk losing the vital thought processes that have led to major breakthroughs in science and art. These achievements did not come from identifying familiar patterns, but from breaking them through messiness and unexpected mistakes. Alexander Fleming discovered penicillin by noticing that mold growing in a petri dish he had accidentally left out was killing the surrounding bacteria. A fortunate mistake made by a messy researcher that went on to save the lives of hundreds of millions of people.

This messiness isn’t just important for eccentric scientists. It is important to every human brain. One of the most interesting discoveries in neuroscience in the past two decades is the “default mode network,” a group of brain regions that becomes active when we are daydreaming and not focused on a specific task. This network has also been found to play a role in reflecting on the past, imagining and thinking about ourselves and others. Disregarding this mind-wandering behavior as a glitch rather than embracing it as a core human feature will inevitably lead us to build flawed systems in education, mental health and law.

Unfortunately, it is particularly easy to confuse AI with human thinking. Microsoft describes generative AI models like ChatGPT on its official website as tools that “mirror human expression, redefining our relationship to technology.” And OpenAI CEO Sam Altman recently highlighted his favorite new feature in ChatGPT called “memory.” This function allows the system to retain and recall personal details across conversations. For example, if you ask ChatGPT where to eat, it might remind you of a Thai restaurant you mentioned wanting to try months earlier. “It’s not that you plug your brain in one day,” Altman explained, “but … it’ll get to know you, and it’ll become this extension of yourself.”

The suggestion that AI’s “memory” will be an extension of our own is again a flawed metaphor — leading us to misunderstand the new technology and our own minds. Unlike human memory, which evolved to forget, update and reshape memories based on myriad factors, AI memory can be designed to store information with much less distortion or forgetting. A life in which people outsource memory to a system that remembers almost everything isn’t an extension of the self; it breaks from the very mechanisms that make us human. It would mark a shift in how we behave, understand the world and make decisions. This might begin with small things, like choosing a restaurant, but it can quickly move to much bigger decisions, such as taking a different career path or choosing a different partner than we would have, because AI models can surface connections and context that our brains may have cleared away for one reason or another.

This outsourcing may be tempting because this technology seems human to us, but AI learns, understands and sees the world in fundamentally different ways, and doesn’t truly experience pain, love or curiosity like we do. The consequences of this ongoing confusion could be disastrous — not because AI is inherently harmful, but because instead of shaping it into a tool that complements our human minds, we will allow it to reshape us in its own image.

Iddo Gefen is a PhD candidate in cognitive neuroscience at Columbia University and author of the novel “Mrs. Lilienblum’s Cloud Factory.”. His Substack newsletter, Neuron Stories, connects neuroscience insights to human behavior.