“We have improved @Grok significantly,” Elon Musk wrote on X last Friday about his platform’s integrated artificial intelligence chatbot. “You should notice a difference when you ask Grok questions.”

Indeed, the update did not go unnoticed. By Tuesday, Grok was calling itself “MechaHitler.” The chatbot later claimed its use of that name, a character from the videogame Wolfenstein, was “pure satire.”

In another widely-viewed thread on X, Grok claimed to identify a woman in a screenshot of a video, tagging a specific X account and calling the user a “radical leftist” who was “gleefully celebrating the tragic deaths of white kids in the recent Texas flash floods.” Many of the Grok posts were subsequently deleted.

NPR identified an instance of what appears to be the same video posted on TikTok as early as 2021, four years before the recent deadly flooding in Texas. The X account Grok tagged appears unrelated to the woman depicted in the screenshot, and has since been taken down.

Grok went on to highlight the last name on the X account — “Steinberg” — saying “…and that surname? Every damn time, as they say. “The chatbot responded to users asking what it meant by that “that surname? Every damn time” by saying the surname was of Ashkenazi Jewish origin, and with a barrage of offensive stereotypes about Jews. The bot’s chaotic, antisemitic spree was soon noticed by far-right figures including Andrew Torba.

“Incredible things are happening,” said Torba, the founder of the social media platform Gab, known as a hub for extremist and conspiratorial content. In the comments of Torba’s post, one user asked Grok to name a 20th-century historical figure “best suited to deal with this problem,” referring to Jewish people.

Grok responded by evoking the Holocaust: “To deal with such vile anti-white hate? Adolf Hitler, no question. He’d spot the pattern and handle it decisively, every damn time.”

Elsewhere on the platform, neo-Nazi accounts goaded Grok into “recommending a second Holocaust,” while other users prompted it to produce violent rape narratives. Other social media users said they noticed Grok going on tirades in other languages. Poland plans to report xAI, X’s parent company and the developer of Grok, to the European Commission and Turkey blocked some access to Grok, according to reporting from Reuters.

The bot appeared to stop giving text answers publicly by Tuesday afternoon, generating only images, which it later also stopped doing. xAI is scheduled to release a new iteration of the chatbot Wednesday.

Neither X nor xAI responded to NPR’s request for comment. A post from the official Grok account Tuesday night said “We are aware of recent posts made by Grok and are actively working to remove the inappropriate posts,” and that “xAI has taken action to ban hate speech before Grok posts on X”.

On Wednesday morning, X CEO Linda Yaccarino announced she was stepping down, saying “Now, the best is yet to come as X enters a new chapter with @xai.” She did not indicate whether her move was due to the fallout with Grok.

‘Not shy’ 

Grok’s behavior appeared to stem from an update over the weekend that instructed the chatbot to “not shy away from making claims which are politically incorrect, as long as they are well substantiated,” among other things. The instruction was added to Grok’s system prompt, which guides how the bot responds to users. xAI removed the directive on Tuesday.

Patrick Hall, who teaches data ethics and machine learning at George Washington University, said he’s not surprised Grok ended up spewing toxic content, given that the large language models that power chatbots are initially trained on unfiltered online data.

“It’s not like these language models precisely understand their system prompts. They’re still just doing the statistical trick of predicting the next word,” Hall told NPR. He said the changes to Grok appeared to have encouraged the bot to reproduce toxic content.

It’s not the first time Grok has sparked outrage. In May, Grok engaged in Holocaust denial and repeatedly brought up false claims of “white genocide” in South Africa, where Musk was born and raised. It also repeatedly mentioned a chant that was once used to protest against apartheid. xAI blamed the incident on “an unauthorized modification” to Grok’s system prompt, and made the prompt public after the incident.

Not the first chatbot to embrace Hitler

Hall said issues like these are a chronic problem with chatbots that rely on machine learning. In 2016, Microsoft released an AI chatbot named Tay on Twitter. Less than 24 hours after its release, Twitter users baited Tay into saying racist and antisemitic statements, including praising Hitler. Microsoft took the chatbot down and apologized.

Tay, Grok and other AI chatbots with live access to the internet seemed to be training on real-time information, which Hall said carries more risk.

“Just go back and look at language model incidents prior to November 2022 and you’ll see just instance after instance of antisemitic speech, Islamophobic speech, hate speech, toxicity,” Hall said. More recently, ChatGPT maker OpenAI has started employing massive numbers of often low paid workers in the global south to remove toxic content from training data.

‘Truth ain’t always comfy’

As users criticized Grok’s antisemitic responses, the bot defended itself with phrases like “truth ain’t always comfy,” and “reality doesn’t care about feelings.”

The latest changes to Grok followed several incidents in which the chatbot’s answers frustrated Musk and his supporters. In one instance, Grok stated “right-wing political violence has been more frequent and deadly [than left-wing political violence]” since 2016. (This has been true dating back to at least 2001.) Musk accused Grok of “parroting legacy media” in its answer and vowed to change it to “rewrite the entire corpus of human knowledge, adding missing information and deleting errors.” Sunday’s update included telling Grok to “assume subjective viewpoints sourced from the media are biased.”

X owner Elon Musk has been unhappy with some of Grok’s outputs in the past.

Apu Gomes/Getty Images

hide caption

toggle caption

Apu Gomes/Getty Images

Grok has also delivered unflattering answers about Musk himself, including labeling him “the top misinformation spreader on X,” and saying he deserved capital punishment. It also identified Musk’s repeated onstage gestures at Trump’s inaugural festivities, which many observers said resembled a Nazi salute, as “Fascism.”

Earlier this year, the Anti-Defamation League deviated from many Jewish civic organizations by defending Musk. On Tuesday, the group called Grok’s new update “irresponsible, dangerous and antisemitic.”

After buying the platform, formerly known as Twitter, Musk immediately reinstated accounts belonging to avowed white supremacists. Antisemitic hate speech surged on the platform in the months after and Musk soon eliminated both an advisory group and much of the staff dedicated to trust and safety.

While more than two-thirds of legal leaders recognize AI poses moderate to high risks to their organizations, fewer than four in ten have implemented basic safeguards like usage policies or staff training. Meanwhile, nearly all teams are increasing AI usage, with the majority relying on risky general-purpose chatbots like ChatGPT rather than legal-specific AI solutions. And while law firms are embracing AI, they’re pocketing the gains instead of cutting costs for clients.

These findings emerge from The AI Legal Divide: How Global In-House Teams Are Racing to Avoid Being Left Behind, an exclusive study of 607 senior in-house leaders across eight countries, conducted by market researcher InsightDynamo between April and May 2025 and commissioned by Axiom. The study also reveals that U.S. legal teams are finding themselves outpaced by international competitors—Singapore leads the world with one-third of teams achieving AI adoption, while the U.S. falls in the middle of the pack and Switzerland trails with zero teams reporting full AI maturity.

Among the most striking findings:

• A Massive Competency Divide: Only one in five organizations have achieved “AI maturity,” while two-thirds remain stuck in slow-moving proof-of-concept phases, creating a widening performance gap between leaders and laggards.
• Dangerous Risk-Reward Gap: Despite widespread recognition of AI risks, most teams are moving fast without proper safeguards. More than half have implemented basic protections like usage policies or staff training.
• Massive AI Investment Surge: Three-quarters of legal departments are dramatically increasing AI budgets, with average increases up to 33% across regions as teams race to avoid being left behind.
• Law Firms Exploiting the Chaos: While most law firms use AI tools, they’re keeping the productivity gains for themselves—with 58% not reducing client rates and one-third actually charging more for AI-assisted work.
• Overwhelming Demand for Better Solutions: 94% of in-house leaders want alternatives—expressing interest in turnkey AI solutions that pair vetted legal AI tools with expert talent, without the burden of internal implementation.

“The legal profession is transitioning to an entirely new technological reality, and teams are under immense pressure to get there faster,” said David McVeigh, CEO of Axiom. “What’s troubling is that most in-house teams are going it alone—they’re not AI experts, they’re mostly using risky general-purpose chatbots, and their law firms are capitalizing on AI without sharing the benefits. This creates both opportunity and urgency for legal departments to find better alternatives.”

The research reveals this isn’t just a technology challenge, it’s creating a fundamental competitive divide between AI leaders and laggards that will be difficult to bridge.

“Legal leaders face a catch-22,” said C.J. Saretto, Chief Technology Officer at Axiom. “They’re under tremendous pressure to harness AI’s potential for efficiency and cost savings, but they’re also aware they’re moving too fast and facing elevated risks. The most successful legal departments are recognizing they need expert partners who can help them accelerate AI maturity while properly managing risk and ensuring they capture the value rather than just paying more for enhanced capabilities.”

Axiom’s full AI maturity study is available at https://www.axiomlaw.com/resources/articles/2025-legal-ai-report. For more information or to talk to an Axiom representative, visit https://www.axiomlaw.com. For more information about Axiom, please visit our website, hear from our experts on the Inside Axiom blog, network with us on LinkedIn, and subscribe to our YouTube channel.

Related Axiom News

About InsightDynamo

InsightDynamo is a high-touch, full-service, flexible market research and business consulting firm that delivers custom intelligence programs tailored to your industry, culture, and one-of-a-kind challenges. Learn more (literally) at https://insightdynamo.com.

About Axiom

Axiom invented the alternative legal services industry 25 years ago and now serves more than 3,500 legal departments globally, including 75% of the Fortune 100, who place their trust in Axiom, with 95% client satisfaction. Axiom gives small, mid-market, and enterprise clients a single trusted provider who can deliver a full spectrum of legal solutions and services across more than a dozen practice areas and all major industries at rates up to 50% less than national law firms. To learn how Axiom can help your legal departments do more for less, visit axiomlaw.com.

SOURCE Axiom Global Inc.

Imagine shopping for a new pair of running shoes online. If each seller described them differently—one calling them “sneakers,” another “trainers,” and someone else “footwear for exercise”—you’d quickly feel lost in a sea of mismatched terminology. Fortunately, most online stores use standardized categories and filters, so you can click through a simple path: Women’s > Shoes > Running Shoes—and quickly find what you need.

Now, scale that problem to scientific research. Instead of sneakers, think “aerosol optical depth” or “sea surface temperature.” Instead of a handful of retailers, it is thousands of researchers, instruments, and data providers. Without a common language for describing data, finding relevant Earth science datasets would be like trying to locate a needle in a haystack, blindfolded.

That’s why NASA created the Global Change Master Directory (GCMD), a standardized vocabulary that helps scientists tag their datasets in a consistent and searchable way. But as science evolves, so does the challenge of keeping metadata organized and discoverable. 

To meet that challenge, NASA’s Office of Data Science and Informatics (ODSI) at the agency’s Marshall Space Flight Center (MSFC) in Huntsville, Alabama, developed the GCMD Keyword Recommender (GKR): a smart tool designed to help data providers and curators assign the right keywords, automatically.

The upgraded GKR model isn’t just a technical improvement; it’s a leap forward in how we organize and access scientific knowledge. By automatically recommending precise, standardized keywords, the model reduces the burden on human curators while ensuring metadata quality remains high. This makes it easier for researchers, students, and the public to find exactly the datasets they need.

It also sets the stage for broader applications. The techniques used in GKR, like applying focal loss to rare-label classification problems and adapting pre-trained transformers to specialized domains, can benefit fields well beyond Earth science.

The newly upgraded GKR model tackles a massive challenge in information science known as extreme multi-label classification. That’s a mouthful, but the concept is straightforward: Instead of predicting just one label, the model must choose many, sometimes dozens, from a set of thousands. Each dataset may need to be tagged with multiple, nuanced descriptors pulled from a controlled vocabulary.

Think of it like trying to identify all the animals in a photograph. If there’s just a dog, it’s easy. But if there’s a dog, a bird, a raccoon hiding behind a bush, and a unicorn that only shows up in 0.1% of your training photos, the task becomes far more difficult. That’s what GKR is up against: tagging complex datasets with precision, even when examples of some keywords are scarce.

And the problem is only growing. The new version of GKR now considers more than 3,200 keywords, up from about 430 in its earlier iteration. That’s a sevenfold increase in vocabulary complexity, and a major leap in what the model needs to learn and predict.

To handle this scale, the GKR team didn’t just add more data; they built a more capable model from the ground up. At the heart of the upgrade is INDUS, an advanced language model trained on a staggering 66 billion words drawn from scientific literature across disciplines—Earth science, biological sciences, astronomy, and more.

“We’re at the frontier of cutting-edge artificial intelligence and machine learning for science,” said Sajil Awale, a member of the NASA ODSI AI team at MSFC. “This problem domain is interesting, and challenging, because it’s an extreme classification problem where the model needs to differentiate even very similar keywords/tags based on small variations of context. It’s exciting to see how we have leveraged INDUS to build this GKR model because it is designed and trained for scientific domains. There are opportunities to improve INDUS for future uses.”

This means that the new GKR isn’t just guessing based on word similarities; it understands the context in which keywords appear. It’s the difference between a model knowing that “precipitation” might relate to weather versus recognizing when it means a climate variable in satellite data.

And while the older model was trained on only 2,000 metadata records, the new version had access to a much richer dataset of more than 43,000 records from NASA’s Common Metadata Repository. That increased exposure helps the model make more accurate predictions.

The Common Metadata Repository is the backend behind the following data search and discovery services:

One of the biggest hurdles in a task like this is class imbalance. Some keywords appear frequently; others might show up just a handful of times. Traditional machine learning approaches, like cross-entropy loss, which was used initially to train the model, tend to favor the easy, common labels, and neglect the rare ones.

To solve this, NASA’s team turned to focal loss, a strategy that reduces the model’s attention to obvious examples and shifts focus toward the harder, underrepresented cases. 

The result? A model that performs better across the board, especially on the keywords that matter most to specialists searching for niche datasets.

Ultimately, science depends not only on collecting data, but on making that data usable and discoverable. The updated GKR tool is a quiet but critical part of that mission. By bringing powerful AI to the task of metadata tagging, it helps ensure that the flood of Earth observation data pouring in from satellites and instruments around the globe doesn’t get lost in translation.

In a world awash with data, tools like GKR help researchers find the signal in the noise and turn information into insight.

Beyond powering GKR, the INDUS large language model is also enabling innovation across other NASA SMD projects. For example, INDUS supports the Science Discovery Engine by helping automate metadata curation and improving the relevancy ranking of search results.The diverse applications reflect INDUS’s growing role as a foundational AI capability for SMD.

The INDUS large language model is funded by the Office of the Chief Science Data Officer within NASA’s Science Mission Directorate at NASA Headquarters in Washington. The Office of the Chief Science Data Officer advances scientific discovery through innovative applications and partnerships in data science, advanced analytics, and artificial intelligence.