Grok-2 Generates Controversy; Expert Reactions

Elon Musk’s AI operation xAI last week released the latest version of its chatbot, Grok-2, in beta for premium users of X (formerly Twitter). Like some rival systems, it now includes a prompt-driven image generation capability.

Users were quick to test its capabilities and soon discovered it was lacking the guardrails competing products have in place. They were able to generate deepfakes of real people, including musicians and politicians in compromising situations. One combined these ideas and created a picture of Donald Trump marrying Taylor Swift and another Swift in a MAGA hat.

It also played fast and loose with copyright and enabled violent and pornographic image generation. Mickey Mouse was portrayed smoking and drinking alcohol while a nuclear mushroom cloud exploded behind him and Pikachu was depicted carrying an AK-47.

And while Musk promoted the release with flattering images of himself as a cowboy and posing in a leather jacket, some users turned it against him, illustrating him as overweight or kissing Trump.

By contrast, competing tools like Google’s Gemini were seen to have overcompensated for bias by portraying historically white figures such as the founding fathers of Nazi-era German soldiers as people of color.

Here experts weigh in on what went wrong and whether a healthy balance in AI image generation can be reached.

Related:Musk xAI Launches Grok-2, Mini Version With Improved Reasoning Skills

Ryan Waite, vice president of public affairs at Think Big

The lack of content moderation guardrails on Grok presents significant moral and legal issues. Legal experts and civil rights advocates argue that Grok represents one of the most reckless implementations of AI. The unmitigated creation of potentially harmful images opens up X to huge legal liability. Such images could easily be classified as misinformation, sexual harassment or copyright infringement.

The legal red flags alone warrant a reevaluation of Grok's content moderation protocols. What makes this all the more peculiar is Elon Musk's pointed criticism of AI and the very problems his AI product is guilty of creating. One would think Musk would be highly cautious with anything relating to AI.

Grok is fixable, but it would likely require shutting it down for retooling. A fix would require technological adjustments and a shift in the underlying philosophy of how these tools are developed and deployed.

Implementing better content filters, improving moderation policies and ensuring transparency in how the AI is trained and used could mitigate some of the risks of Grok. However, this will likely be a complex and ongoing process, involving collaboration between technologists, ethicists and policymakers.

Related:Musk AI Chatbot Under Fire for Sharing False Election Information

JD Harriman, partner at Foundation Law Group

Grok and Gemini are two AI image generators that generate controversial images but for the wrong reasons. Both generators can be used to create false and misleading images, Gemini because of too much filtering and Grok because of too little filtering.

Gemini made the news at the beginning of 2024 for generating historically unlikely, if not downright inaccurate, images due to an over-corrective filter designed to prevent known racial bias problems with AI. By contrast, Grok users have detected almost no filtering other than hard-core porn in the Grok Generator. As a result, Grok users have made patently false images of politicians and public figures engaged in offensive conduct.

Where is the best place to draw the line for AI image generators? Whether the generator itself has a filter, the images produced are already subject to legal guardrails. Copyright infringement, trademark infringement, libel, incitement and other legal remedies are fully available to be applied to images generated by the Grok generator. It may be difficult to implement a take-down policy for Grok, similar to YouTube because Grok just creates the images and they are posted elsewhere.

Moral guardrails are more difficult to implement and enforce. One technique that OpenAI uses is a watermark on all images produced. This watermark can at least allow viewers to know if they are seeing an AI-generated image and bring their own skepticism to bear. It would be good to have a hidden watermark that is difficult to remove to allow proper vetting of images to reduce the risk of harm.

Matt Hasan, founder and CEO at aiRESULTS

Grok, xAI's advanced AI system, has faced significant controversies, particularly around its manipulation and generation of images. For instance, there have been instances where Grok was used to create hyper-realistic but entirely fabricated images, leading to the spread of false information and defamation. Such misuse highlights the system's lack of moral and legal guardrails, raising concerns about the potential for abuse in creating misleading media or infringing on intellectual property.

The absence of robust oversight mechanisms means that harmful or unethical content could proliferate unchecked. To mitigate these issues, implementing stringent content moderation policies, developing advanced detection tools to identify manipulated images and establishing clear legal frameworks for AI-generated content are crucial steps. Also, fostering transparency in AI development and encouraging industry-wide standards can help ensure that such technologies are used responsibly and ethically.

Shashi Bellamkonda, principal research director at Info-Tech Research Group

Public-facing AI large language model (LLM) providers face a dilemma. If they are restrictive then users don't find them exciting and if they are unrestrictive, they face controversy and outrage. Grok wants to be unrestrictive with Elon Musk's philosophy of free speech and increasing innovation. Producing fakes is neither innovative nor legal. Unless these companies self-police their creative aspects, trust in AI will decrease to the detriment of the AI companies that do a good job with analytical and predictive AI.

From my experiments, the images created by Grok are representative but not always accurate and a savvy user can spot them as fakes. Kamala Harris making a dosa is an example of that.

A Grok-generated image of Kamala Harris making a Dosa. Credit Info-Tech Research Group via Grok

Rob Rosenberg, independent legal and strategy consultant at Telluride Legal Strategies

Grok attempts to distinguish itself from other AI models by self-proclaiming that it is “an AI search assistant with a twist of humor and a dash of rebellion.” It wants to help you with your search queries while keeping you “entertained and engaged.” Great goals, so long as the content produced by the model doesn’t demean, defraud or misinform users.

Unlike other platforms that label their AI-generated images with a watermark that identifies them as such, Grok does not tag its image results in any way that would clue in downstream customers as to their origin. This puts such downstream customers at risk of being misled – misinterpreting such images as genuine and not a deepfake. Those creating images on Grok may know that they are intended to be humorous, but when those images are shared more widely, unsuspecting consumers may not realize that they are AI-generated.

There is a movement toward more safeguards for consumers, rather than less. The No Fakes Act, a bi-partisan legislation that is gaining widespread support from talent, industry and consumer groups, is aimed at protecting people from creating and sharing deepfake images. The bill’s sponsors hope to see it become law before the end of 2024. And it seems in direct opposition to the types of images that Grok allows its users to create. Assuming the No Fakes Act becomes law, I imagine that Grok will be forced to enact some responsible guidelines consistent with this new law.

Edward Tian, CEO at GPTZero

Grok has proven to be capable of generating ethically questionable and inappropriate imagery. Public figures have been depicted in scenes of violence and degradation, without giving permission to use their likeness in such ways. And, because there is a widespread lack of media literacy, there are certainly lots of people who will see these images and instantly believe them to be real. So, I wouldn’t be surprised if legal action is soon taken.

Grok's prompt screen. Credit: Grok

Mark McNasby, co-founder and CEO at Ivy.ai

In the race to innovate, even the most promising technologies can falter if deployed prematurely. The recent controversies surrounding Elon Musk's Grok AI underscores a critical issue – the peril of relying on immature AI models for high-stakes information. Musk’s candid admission of Grok's infancy, juxtaposed with his ambition for the AI to be the epitome of truth, highlights a crucial dichotomy in AI development.

Looking ahead, as the digital landscape evolves, the need for robust systems to verify and corroborate information becomes ever more urgent. The incidents with Grok illustrate the complexity of fine-tuning large-scale models and the paramount importance of accuracy in AI outputs. With the tech landscape as dynamic as ever, what holds today may shift tomorrow. This fluidity necessitates AI models that are not only advanced but also meticulously validated for factual correctness.

Mark Beedles, chief operating officer and co-founder at rTriibe

Grok's ability to generate controversial and legally questionable images highlights the alarming gap in AI regulation. 

Without a clear framework or guidelines, these powerful tools are left open to misinterpretation and abuse, creating a landscape where innovation can easily spiral into infringement. 

The absence of oversight not only invites foul play but also underscores the urgent need for comprehensive rules to navigate the ethical and legal minefields of AI-driven creativity."

Rick Bentley, founder and Abhik Sarkar, director of machine learning at Cloudastructure

Historical parallels exist between the Ottoman Empire's approach to the printing press and the modern closed AI system. In the 15th century, the Ottoman Empire, wary of the printing press's potential to spread dissent and challenge authority, heavily restricted its use. This decision, while intended to maintain societal stability, ultimately hindered intellectual and technological progress within the empire.

Today, we see a similar dynamic unfolding in the realm of artificial intelligence. Closed AI systems, tightly controlled by corporations or governments, echo the Ottoman approach. These entities often justify their restrictive policies by citing concerns over potential misuse, security risks or the need to maintain competitive advantages. However, this approach raises important questions about innovation, societal progress and the equitable distribution of technological benefits.

Just as access to printing technology does not result in everyone illicitly printing money, access to AI technology does not inherently lead to negative outcomes. Rather than focusing solely on regulation to prevent potential misuse of AI, there is a strong argument for prioritizing innovation, particularly in the realm of detection and verification technologies, so that it's not misused by just a handful of people.

It's not a question of whether it is fixable or not, but rather should it be "fixed"? Most lawmakers don't have anything intelligent to say about advanced technologies like AI. We are a capitalist society and market forces have made us wealthy beyond human historical comprehension. The worst thing we can do to a nascent technology is let fear-mongering deprive us of our liberties and then deprive us of all the societal benefits to be had from the next technological revolution – or worst of all, make it only accessible to large corporations who can afford to meet the regulations.

At Cloudastructure, we are already using Grok to create synthetic images to help train our system to detect guns in video surveillance feeds. Gun detection is clearly a public safety issue and we are proud to pioneer the technology. However, it's hard to find video of guns being used in shopping malls, for example, and where such footage exists, there are, rightfully, moral qualms about using that data for training – real footage of real shootings with real victims. Having systems like Grok and allowing them to grow without interference helps make protecting the public possible.

As my dad would say, if it ain't broke, don't fix it.