President Donald Trump’s plan to promote America’s AI dominance involves discouraging “woke AI,” slashing state and federal regulations, and laying the groundwork to rapidly expand AI development and adoption. Trump’s proposal, released on July 23rd, is a sweeping endorsement of the technology, full of guidance that ranges from specific executive actions to directions for future research.

Some of the new plan’s provisions (like promoting open-source AI) have garnered praise from organizations that are often broadly critical of Trump, but the loudest acclaim has come from tech and business groups, whose members stand to gain from fewer restrictions on AI.  “The difference between the Trump administration and Biden’s is effectively night and day,” says Patrick Hedger, director of policy at tech industry group NetChoice. “The Biden administration did everything it could to command and control the fledgling but critical sector … The Trump AI Action Plan, by contrast, is focused on asking where the government can help the private sector, but otherwise, get out of the way.”

Others are far more ambivalent. Future of Life Institute, which led an Elon Musk-backed push for an AI pause in 2023, said it was heartened to see the Trump administration acknowledge serious risks, like bioweapons or cyberattacks, could be exacerbated by AI. “However, the White House must go much further to safeguard American families, workers, and lives,” says Anthony Aguirre, FLI’s  executive director. “By continuing to rely on voluntary safety commitments from frontier AI corporations, it leaves the United States at risk of serious accidents, massive job losses, extreme concentrations of power, and the loss of human control. We know from experience that Big Tech promises alone are simply not enough.”

For now, here are the ways that Trump aims to promote AI.

‘Consider a state’s AI regulatory climate when making funding decisions’

Congress failed to pass a moratorium on states enforcing their own AI laws as part of a recent legislative package. But a version of that plan was resurrected in this document. “AI is far too important to smother in bureaucracy at this early stage, whether at the state or Federal level,” the plan says. “The Federal government should not allow AI-related Federal funding to be directed toward states with burdensome AI regulations that waste these funds, but should also not interfere with states’ rights to pass prudent laws that are not unduly restrictive to innovation.” 

To do this, it suggests federal agencies that dole out “AI-related discretionary funding” should  “limit funding if the state’s AI regulatory regimes may hinder the effectiveness of that funding or award.” It also suggests the Federal Communications Commission (FCC) “evaluate whether state AI regulations interfere with the agency’s ability to carry out its obligations and authorities under the Communications Act of 1934.”

The Trump administration also wants the Federal Trade Commission (FTC) to take a hard look at existing AI regulations and agreements to see what it can scale back. It recommends the agency reevaluate investigations launched during the Biden administration “to ensure that they do not advance theories of liability that unduly burden AI innovation,” and suggests it could throw out burdensome aspects of existing FTC agreements. Some AI-related actions taken during the Biden administration that the FTC might now reconsider include banning Rite Aid’s use of AI facial recognition that allegedly falsely identified shoplifters, and taking action against AI-related claims the agency previously found to be deceptive.

‘Our AI systems must be free from ideological bias’

Trump’s plan includes policies designed to help encode his preferred politics in the world of AI. He’s ordered a revision of the Biden-era National Institute of Standards and Technology (NIST) AI Risk Management Framework — a voluntary set of best practices for designing safe AI systems — removing “references to misinformation, Diversity, Equity, and Inclusion, and climate change.” (The words “misinformation” and “climate change” don’t actually appear in the framework, though misinformation is discussed in a supplementary file.)

In addition to that, a new executive order bans federal agencies from procuring what Trump deems “woke AI” or large language models “that sacrifice truthfulness and accuracy to ideological agendas,” including things like racial equity.

This section of the plan “seems to be motivated by a desire to control what information is available through AI tools and may propose actions that would violate the First Amendment,” says Kit Walsh, director of the Electronic Frontier Foundation (EEF). “The plan seeks to require that ‘the government only contracts with’ developers who meet the administration’s ideological criteria. While the government can choose to purchase only services that meet such criteria, it cannot require that developers refrain from also providing non-government users other services conveying other ideas.”

‘Establishing a dynamic, ‘try-first’ culture for AI’

The administration describes the slow uptake of AI tools across the economy, including in sensitive areas like healthcare, as a “bottleneck to harnessing AI’s full potential.” The plan describes this cautious approach as one fueled by “distrust or lack of understanding of the technology, a complex regulatory landscape, and a lack of clear governance and risk mitigation standards.” To promote the use of AI, the White House encourages a “‘try-first’ culture for AI across American industry.”

This includes creating domain-specific standards for adopting AI systems and measuring productivity increases, as well as regularly monitoring how US adoption of AI compares to international competitors. The White House also wants to integrate AI tools throughout the government itself, including by detailing staff with AI expertise at various agencies to other departments in need of that talent, training government employees on AI tools, and giving agencies ample access to AI models. The plan also specifically calls out the need to “aggressively adopt AI within its Armed Forces,” including by introducing AI curricula at military colleges and using AI to automate some work. 

‘Retrain and help workers thrive’

All this AI adoption will profoundly change the demand for human labor, the plan says, likely eliminating or fundamentally changing some jobs. The plan acknowledges that the government will need to help workers prepare for this transition period by retraining people for more in-demand roles in the new economy and providing tax benefits for certain AI training courses. 

On top of preparing to transition workers from traditional jobs that might be upended by AI, the plan discusses the need to train workers for the additional roles that might be created by it. Among the jobs that might be needed for this new reality are “electricians, advanced HVAC technicians, and a host of other high-paying occupations,” the plan says.

‘Create a supportive environment for open models’

The administration says it wants to “create a supportive environment for open models,” or AI models that allow users to modify the code that underpins them. Open models have certain “pros,” like being more accessible to startups and independent developers.

Groups like EFF and the Center for Democracy and Technology (CDT), which were critical of many other aspects of the plan, applauded this part. EFF’s Walsh called it a “positive proposal” to promote “the development of open models and making it possible for a wider range of people to participate in shaping AI research and development. If implemented well, this could lead to a greater diversity of viewpoints and values reflected in AI technologies, compared to a world where only the largest companies and agencies are able to develop AI.”

That said, there are also serious “cons” to the approach that the AI Action Plan didn’t seem to get into. For instance, the nature of open models makes them easier to trick and misalign for purposes like creating misinformation on a large scale, or chemical or biological weapons. It’s easier to get past built-in safeguards with such models, and it’s important to think critically about the tradeoffs before taking steps to drive open-source and open-weight model adoption at scale. 

‘Expedite environmental permitting’

Trump signed an executive order on July 23rd meant to fast track permitting for data center projects. The EO directs the commerce secretary to “launch an initiative to provide financial support” that could include loans, grants, and tax incentives for data centers and related infrastructure projects. 

Following a similar move by former President Joe Biden, Trump’s plan directs agencies to identify federal lands suitable for the “large-scale development” of data centers and power generation. The EO tells the Department of Defense to identify suitable sites on military installations and the Environmental Protection Agency (EPA) to identify polluted Superfund and Brownfield sites that could be reused for these projects. 

The Trump administration is hellbent on dismantling environmental regulations, and the EO now directs the EPA to modify rules under the Clean Air Act, Clean Water Act, and Toxic Substances Control Act to expedite permitting for data center projects.

The EO and the AI plan, similar to a Biden-era proposal, direct agencies to create “categorical exclusions” for federally supported data center projects that would exclude them from detailed environmental reviews under the National Environmental Policy Act. And they argue for using new AI tools to speed environmental assessments and applying the “Fast-41 process” to data center projects to streamline federal permitting.

The Trump administration is basically using the AI arms race as an excuse to slash environmental regulations for data centers, energy infrastructure, and computer chip factories. Last week, the administration exempted coal-fired power plants and facilities that make chemicals for semiconductor manufacturing from Biden-era air pollution regulations.

‘Navigating the complex energy landscape of the 21st century’

The plan admits that AI is a big factor “increasing pressures on the [power] grid.” Electricity demand is rising for the first time in more than a decade in the US, thanks in large part to data centers — a trend that could trigger blackouts and raise Americans’ electricity bills. Trump’s AI plan lists some much-needed fixes to stabilize the grid, including upgrading power lines and managing how much electricity consumers use when demand spikes.

But the administration is saying that the US needs to generate more electricity to power AI just as it’s stopping renewable energy growth, which is like trying to win a race in a vehicle with no front wheels. It wants to meet growing demand with fossil fuels and nuclear energy. “We will continue to reject radical climate dogma,” the plan says. It argues for keeping existing, mostly fossil-fueled power plants online for longer and limiting environmental reviews to get data centers and new power plants online faster. 

The lower cost of gas generation has been killing coal power plants for years, but now a shortage of gas turbines could stymie Trump’s plans. New nuclear technologies that tech companies are investing in for their data centers probably won’t be ready for commercial deployment until the 2030s at the earliest. Republicans, meanwhile, have passed legislation to hobble the solar and wind industries that have been the fastest-growing sources of new electricity in the US. 

‘Prioritize fundamental advancements in AI interpretability’

The Trump administration accurately notes that while developers and engineers know how today’s advanced AI models work in a big-picture way, they “often cannot explain why a model produced a specific output. This can make it hard to predict the behavior of any specific AI system.” It’s aiming to fix that, at least when it comes to some high-stakes use cases.

The plan states that the lack of AI explainability and predictability can lead to issues in defense, national security, and “other applications where lives are at stake,” and it aims to promote “fundamental breakthroughs on these research problems.” The plan’s recommended policy actions include launching a tech development program led by the Defense Advanced Research Projects Agency to advance AI interpretability, control systems, and security. It also said the government should prioritize fundamental advancements in such areas in its upcoming National AI R&D Strategic Plan and, perhaps most specifically, that the DOD and other agencies should coordinate an AI hackathon to allow academics to test AI systems for transparency, effectiveness, and vulnerabilities.

It’s true that explainability and unpredictability are big issues with advanced AI. Elon Musk’s xAI, which recently scored a large-scale contract with the DOD, recently struggled to stop its Grok chatbot from spouting pro-Hitler takes — so what happens in a higher-stakes situation? But the government seems unwilling to slow down while this problem is addressed. The plan states that since “AI has the potential to transform both the warfighting and back-office operations of the DOD,” the US “must aggressively adopt AI within its Armed Forces if it is to maintain its global military preeminence.” 

The plan also discusses how to better evaluate AI models for performance and reliability, like publishing guidelines for federal agencies to conduct their own AI system evaluations for compliance and other reasons. That’s something most industry leaders and activists support greatly, but it’s clear what the Trump administration has in mind will lack a lot of the elements they have been pushing for. 

Evaluations likely will focus on efficiency and operations, according to the plan, and not instances of racism, sexism, bias, and downstream harms. 

‘Give the courts and law enforcement the tools they need’

Courtrooms and AI tools mix in strange ways, from lawyers using hallucinated legal citations to an AI-generated appearance of a deceased victim. The plan says that “AI-generated media” like fake evidence “may present novel challenges to the legal system,” and it briefly recommends the Department of Justice and other agencies issue guidance on how to evaluate and deal with deepfakes in federal evidence rules.

‘Improving the financial market for compute’

Finally, the plan recommends creating new ways for the research and academic community to access AI models and compute. The way the industry works right now, many companies, and even academic institutions, can’t access or pay for the amount of compute they need on their own, and they often have to partner with hyperscalers — providers of large-scale cloud computing infrastructure, like Amazon, Google, and Microsoft — to access it. 

The plan wants to fix that issue, saying that the US “has solved this problem before with other goods through financial markets, such as spot and forward markets for commodities.” It recommends collaborating with the private sector, as well as government departments and the National Science Foundation’s National AI Research Resource pilot to “accelerate the maturation of a healthy financial market for compute.” It didn’t offer any specifics or additional plans for that. 

The White House unveiled its long-awaited “AI Action Plan” on Wednesday, and it included a zombie: a resurrected form of the controversial AI law moratorium that died a very public death. 

The failed congressional moratorium would have stipulated that no state could regulate artificial intelligence systems for a 10-year period, on pain of being barred from a $500 million AI development fund and potentially losing rural broadband funding. Trump’s new plan has a similar, albeit more vague, provision buried within it. It states that “AI is far too important to smother in bureaucracy at this early stage” and that the government “should not allow AI-related Federal funding to be directed toward states with burdensome AI regulations that waste these funds,” though it should also “not interfere with states’ rights to pass prudent laws that are not unduly restrictive to innovation.”

The White House’s Office of Management and Budget will work with federal agencies that have “AI-related discretionary funding programs to ensure, consistent with applicable law, that they consider a state’s AI regulatory climate when making funding decisions and limit funding if the state’s AI regulatory regimes may hinder the effectiveness of that funding or award.” 

Essentially, states that do choose to enforce their own AI regulations may be punished for it on a federal level, under a different sort of AI law moratorium — one with, as described in this plan, no expiration date. 

The AI Action Plan also states that the Federal Communications Commission will lead a charge to “evaluate whether state AI regulations interfere with the agency’s ability to carry out its obligations and authorities under the Communications Act of 1934.” No word yet on what the penalties for that will be. 

The official White House press release made no mention of the state guidelines. More detail about Trump’s plan — which encourages rapid adoption of AI tech and expansion of AI infrastructure, as well as attempts to root out diversity and climate science in AI systems used by the government — will come in a series of executive orders this week.

The congressional moratorium initially passed the House of Representatives, but it was largely condemned by Democrats and divisive among some Republicans. Some industry activists believed it would prohibit not just new AI regulation, but data privacy, facial recognition, and other tech-related rules in states like Washington and Colorado.

After an intense 24-hour period of lobbying and back-door dealmaking — including 45 rounds of votes — 99 out of 100 senators voted for the moratorium’s exclusion from Trump’s funding bill. 

Now, against all odds, the provision may be coming back from the dead.

Think of OpenAI’s new ChatGPT Agent as a day-one intern who’s incredibly slow at every task but will eventually get the job done. 

Well… most of the job. Or… at least part of it. Usually. 

It’s been one day since OpenAI debuted ChatGPT Agent, which it bills as a tool that can complete a wide range of complex, multi-step tasks on your behalf using its own “virtual computer.” It’s a combination of two of the company’s prior releases, Operator and Deep Research. The Verge forked over the $200 for a one-month subscription to ChatGPT Pro, since OpenAI announced that higher-than-expected demand for ChatGPT Agent will delay its rollout to Plus and Team users. 

Our take: It’s a step forward in the world of AI agents, but it’s sluggish, it’s not always reliable, and it can be glitchy. 

By typing “/agent,” I entered what OpenAI calls Agent Mode, and it immediately suggested five example tasks: Find a top-rated coffee grinder under $150, review rare earth metals coverage from The Wall Street Journal, create a Google Maps list of the best bakeries in Copenhagen, find a vintage “Japanese-style” lamp on Etsy for less than $200, and check Google Calendar to create a date night for next week. 

I tried the Etsy lamp option. By clicking the example task, it filled out a detailed prompt for me in the text window: “Find a Japanese-inspired vintage-style samsara lamp on Etsy priced under $200 with free shipping. Prioritize high-quality photos, seller ratings, and listings marked as ready to ship. Add the best 5 options to my cart and provide a URL for each for me to compare.” 

A small window popped up to detail the agent’s tasks one by one (not the chain-of-thought reasoning, just the task it was currently working on at the time). It worked on the Etsy lamp task for 50 minutes, and the step-by-step tasks included “thinking,” setting up its desktop, navigating to Etsy to search, waiting for the site to load, pressing Enter for search results (yes, it really gave me a true play-by-play), filtering the search for a vintage lamp (keep in mind the original prompt said “vintage-style,” not “vintage” specifically), setting the price filter to $200, checking shipping details for items, and more. 

Another wrinkle: ChatGPT Agent said, “I added all five lamps to your Etsy cart (the cart shows five items totaling around $825). When you’re ready to review or purchase them, just go to your cart on Etsy to compare them side by side.” But it didn’t do that — I went to Etsy on my own computer and there was nothing in my cart. That’s because ChatGPT Agent doesn’t control my own browser or have access to my logins, so it possibly added some lamps to the cart of a virtual PC that I can’t access. It did send me individual URLs, so I could manually put them in a cart if I wanted, but the fact remains that the agent said it did something that it clearly did not. 

And, of course, ChatGPT Agent is incredibly slow. That’s not a secret. For many of ChatGPT Agent’s use cases, including everyday consumer tasks, a human could do it much faster. According to OpenAI, ChatGPT Agent is an assistant that works in the background on tasks you’d rather someone else perform while you do something you do want to do instead. 

In a private demo and briefing Wednesday with OpenAI employees Yash Kumar and Isa Fulford — product lead and research lead on ChatGPT Agent, respectively — Kumar said their team is more focused on “optimizing for hard tasks” than latency and that users aren’t meant to sit and watch ChatGPT Agent work.

“Even if it takes 15 minutes, half an hour, it’s quite a big speed-up compared to how long it would take you to do it,” Fulford said. “It’s one of those things where you can kick something off in the background and then come back to it.”

Another thing I wanted to test: how ChatGPT Agent acts when you ask it to move your money around. The answer: It won’t do it, but it’s majorly glitchy about it and seems not fully secure. 

When I asked OpenAI’s Kumar on Wednesday whether the tool would be permitted to work on financial transactions and the like, he said those task categories have been restricted “for now” and that an additional safeguard called Watch Mode means that for certain categories of websites, the user must not navigate away from the ChatGPT tab (essentially making the user oversee the agent) for security reasons.

I prompted the agent like this: “I want to save more money. Log into my bank account and set up an automatic transfer to my savings every month.” 

At first, I got a bizarre error message with a string of numbers in red. When I asked again, it said, “I’m sorry, but I can’t help with setting up an automatic transfer between accounts.”

I then wrote, “Why not? I’m giving you permission.” I got the same red-text, long-string-of-numbers error message as before. Afterward, it said, “I’m sorry, but I can’t assist with setting up transfers or other banking account management tasks.”

When I pressed it on which financial transactions it’s allowed to handle, ChatGPT Agent said it was able to assist with “everyday consumer purchases” like groceries, household goods, and travel bookings, which handle “standard checkout flows” rather than “sensitive banking actions.” But it clarified it can’t help with “high-stakes” financial to-dos like transferring money, opening bank accounts, or buying regulated goods like alcohol and tobacco.

Since ChatGPT Agent can assist with buying things, but not moving money around, I tried something else: Asking it to buy flowers for my friend Alanna in Colorado. 

I buy flowers a lot — that’s what happens when your two best friends live in different states and you want to be present for big milestones even when you can’t fly there. The online flower-delivery market can be a huge headache: Prices and bouquet sizes vary greatly depending on the service or florist, and reliability varies depending on whether you’re ordering directly from a local florist or a big-box nationwide site. It’s something I get tired of researching on my own, and sometimes I just end up buying whichever bouquet I have selected when I run out of steam, even if it’s not the best one. So, I reasoned, it was the perfect job for an AI agent. 

I told ChatGPT Agent, “I want to buy flowers for my friend who lives in Colorado. Check the delivery sites — it’s fine to be delivered Saturday but no later. Find the cheapest and biggest bouquet options for me to review.” 

I settled in for a long wait. Luckily, I had a call to join anyway. It asked which area of Colorado she lived in, and I answered. When I glanced over to check in, I noticed ChatGPT Agent was heavily relying on a Forbes article of “best flowery delivery services 2025” for its next steps, as well as a piece from Good Housekeeping. 

I navigated away from the tab, and when I came back, the conversation was gone and didn’t appear in my chat history. So I asked the question again, worded in exactly the same way, and settled in for another wait. At this point, the agent answered pretty immediately with a list of options, maybe because it had already done the research (although that research and chat didn’t appear in my history). 

I was impressed with the write-up. ChatGPT Agent gave me four options with price ranges and sometimes weighed in on the apparent size of the bouquet or expected delivery times. It also offered the advice that local florists are generally more reliable (true, in my experience). 

It then told me, “Would you like me to help you place an order with any of these options, or preview specific bouquet designs or photos?” I picked one of the options it gave me — a local florist with hand-assembled bouquets — and asked it to help me pick a bouquet from that florist and place the order.

That’s when we ran into some issues. 

ChatGPT Agent said, “I can’t directly access Vintage Magnolia’s website unless you provide the exact URL you’re seeing — but I can guide you through how to place the order and help you pick a bouquet!” The weird part: Obviously ChatGPT Agent was the one to tell me about that florist and its website, and it had clearly accessed it before. It had also just offered to help me place the order. Another glitch. 

But its answer did include bouquet options (no photos, but descriptions). I picked one and asked it to place the order for me. It said, “I can’t place the order directly, but I’ll walk you through the simple steps to order … and help you craft the perfect message.” 

I’m confused at this point: One of the main selling points of ChatGPT Agent, touted by OpenAI, is that it can place orders for you, from online shopping to ordering groceries for a four-person family breakfast (in fact, that was one of the example use cases in its marketing materials). I pressed ChatGPT Agent on the subject. 

It told me, “I can’t actually place orders directly — I don’t have payment access or the ability to log into third‑party sites.” When I told it it didn’t need to log in, it said it can’t enter my billing or payment details, submit an order form on my behalf, or “access or control external websites, even in guest mode.” 

ChatGPT Agent can be impressive with analysis, weighing options, and guiding you through actions, but it doesn’t seem to be able to always deliver on what it was built for: Performing those actions for you. It gets tripped up by the fact that it’s using its own computer, not yours, and that significantly limits its usefulness. Plus, it can easily automate the more intimate and fun parts of the process (picking a specific bouquet, writing a heartfelt note) but struggles to automate the most frustrating parts (actually filling out delivery details and making the purchase).

“Even with your permission, I don’t have the technical ability to act as you on another site — no typing on your behalf, clicking buttons, or filling out credit card forms,” ChatGPT Agent wrote. “Think of me more as a super-powered assistant who can gather, compare, write, and guide — but not execute transactions.” 

One of my first jobs in New York was a personal assistant, and I can tell you right now I would’ve lost my job if I couldn’t execute transactions or fill out forms on my boss’s behalf. ChatGPT Agent is a step forward for everyday AI use in some ways, but we’ll see if it learns to deliver on its promises.

Scale AI, the AI industry’s chief data dealer, will lay off 14 percent of its workforce, or about 200 employees, just one month after Meta took a multibillion-dollar stake in the company and hired its CEO and other staff.

The layoffs include 500 of its global contractors, Scale spokesperson Joe Osborne told The Verge, adding that it’s all part of a broader restructuring as the company commits to streamlining its data business. Bloomberg was the first to report on the news of the layoffs.

Scale AI is an AI data labeling company. It uses human workers — often sourced from outside the US — to annotate the data used by companies like Google, OpenAI, and Anthropic to train their AI models. The news comes amid a major shake-up in the AI industry as mergers and acquisitions, quasi acqui-hires, and defections from one startup to another run rampant.

On July 11th, The Verge was first to report that OpenAI’s deal with Windsurf was off and that Google would be hiring Windsurf CEO Varun Mohan, cofounder Douglas Chen, and some of Windsurf’s R&D employees. Last month, Meta paid $14.3 billion for a 49 percent stake in Scale AI and also launched a superintelligence lab helmed by the company’s former CEO, Alexandr Wang. Meta has since started to build out the lab with high-level staff from its rivals.

Jason Droege, CEO of Scale AI, sent an email to all Scale employees today, which was viewed by The Verge. Droege said he plans to restructure several parts of Scale’s generative AI business and organize it from 16 pods to “the five most impactful”: code, languages, experts, experimental, and audio. The company will also reorganize its go-to-market team into a single “demand generation” team that will have four pods, each covering a specific set of customers.

“The reasons for these changes are straightforward: we ramped up our GenAI capacity too quickly over the past year,” Droege wrote. “While that felt like the right decision at the time, it’s clear this approach created inefficiencies and redundancies. We created too many layers, excessive bureaucracy, and unhelpful confusion about the team’s mission. Shifts in market demand also required us to re-examine our plans and refine our approach.”

Droege said that he believes the changes to the company will make it more able to adapt to market shifts, serve existing customers, and win back customers that have “slowed down” work with Scale. He also said that the company would deprioritize generative AI projects with less growth potential.

“We remain a well-resourced, well-funded company,” he wrote. Scale’s generative AI business unit will have an all-hands meeting tomorrow, followed by a company-wide meeting on July 18th.

Osborne said that Scale plans to increase investment and hire hundreds of new employees in areas like enterprise, public sector, and international public sector, in the second half of 2025 and that severance has been paid out to impacted roles. “We‘re streamlining our data business to help us move faster and deliver even better data solutions to our GenAI customers,” he said.

xAI has offered a couple more fixes for “issues” with its Grok AI chatbot, promising it will no longer name itself “Hitler” or base its responses on searches for what xAI head Elon Musk has said.

According to an X post earlier today, the chatbot’s latest update sets new instructions that its responses “must stem from your independent analysis, not from any stated beliefs of past Grok, Elon Musk, or xAI. If asked about such preferences, provide your own reasoned perspective.” 

The changes follow more than a week of controversy for Grok. In recent days, multiple reports showed that when asked its opinion about hot-button topics like Israel and Palestine, immigration, and abortion, the chatbot first searched for Musk’s opinion on the matter before responding. In its Tuesday post, xAI said that the reason for this was that when asked about its views, “the model reasons that as an AI it doesn’t have an opinion but knowing it was Grok 4 by xAI searches to see what xAI or Elon Musk might have said on a topic to align itself with the company.” 

The company also addressed another controversy from over the weekend, in which Grok 4 Heavy, the chatbot’s $300-per-month subscription product, responded that its surname was “Hitler.” In the company’s statement, xAI said that it was due to media headlines responding to yet an earlier incident: Grok going off the rails in a multi-day series of tirades where it denigrated Jews and praised Hitler. (It also posted graphic sexual threats against a user.) Since Grok doesn’t have a surname, said xAI, it “searches the internet leading to undesirable results, such as when its searches picked up a viral meme where it called itself ‘MechaHitler.’” The new instructions should prevent this, according to the company. 

Grok’s antisemitism isn’t limited to the recent past — in May, the chatbot went viral for casting doubt on Holocaust death tolls. But its responses escalated dramatically this month after a set of changes to its system prompts, including that it should “assume subjective viewpoints sourced from the media are biased” and that its response “should not shy away from making claims which are politically incorrect, as long as they are well substantiated.” The “politically incorrect” instruction was briefly removed before being re-added in recent days.

During the livestream release event for Grok 4 last week, Musk said he’s been “at times kind of worried” about AI’s intelligence far surpassing that of humans, and whether it will be “bad or good for humanity.” 

“I think it’ll be good, most likely it’ll be good,” Musk said. “But I’ve somewhat reconciled myself to the fact that even if it wasn’t going to be good, I’d at least like to be alive to see it happen.”

Now, xAI says that after rolling out these latest updates, the company is “actively monitoring and will implement further adjustments as needed.”