What is Elon Musk’s Grok chatbot and how does it work?

You’ve probably heard of Grok, the chatbot created by X (formerly Twitter) as a direct competitor to OpenAI’s ChatGPT. Like other chatbots, Grok is designed to answer questions on a wide range of topics, from current events to pop culture. However, Grok stands out for its distinct personality—described by X owner Elon Musk as having “a bit of wit” and “a rebellious streak.”

What makes Grok particularly unique is its willingness to engage with topics that are often avoided by other chatbots, including politically charged theories and controversial subjects. It also doesn't shy away from using casual, sometimes blunt language, as seen when it responded to the question “When is it appropriate to listen to Christmas music?” with “Whenever the hell you want.”

Perhaps Grok’s most notable feature is its ability to access real-time data from X, a capability that sets it apart from other AI models. Unlike ChatGPT, which has limitations due to its training data and web access restrictions, Grok can offer up-to-date information on recent events. For example, if you ask it “What’s happening in AI today?”, it will provide responses sourced directly from the latest headlines. Earlier this week, Musk even promised to open-source Grok, although the specifics of this plan remain unclear.

How does Grok work?

So, how does Grok work, and what can it do? Grok is powered by Grok-1, a generative AI model created by xAI, Musk's AI startup. Developed using “tens of thousands” of GPUs, Grok-1 was trained with data collected from across the web (up until Q3 2023) and augmented by feedback from human assistants called "AI tutors." In terms of performance, Grok-1 is competitive with Meta’s Llama 2 and even surpasses OpenAI’s GPT-3.5 in key benchmarks.

Human-guided feedback, or reinforcement learning from human feedback (RLHF), is a key technique used to fine-tune AI chatbots. The process begins with training a generative model, followed by gathering additional data to create a "reward" model. This reward model is then used to refine the generative model through reinforcement learning, improving its ability to follow instructions. RLHF has proven effective in making AI more responsive to user commands, but it isn't flawless. Like other models, Grok can still experience issues with hallucinations, occasionally providing incorrect information or false timelines when discussing current events. For example, it has mistakenly reported events, such as wrongly claiming that the Israel-Palestine conflict reached a cease-fire when it had not.

To address gaps in its knowledge base, Grok utilizes "real-time access" to data from X (formerly Twitter) and possibly Tesla, as reported by Bloomberg. This gives it the ability to retrieve up-to-date information about various topics, a feature it shares with other AI models like ChatGPT. Additionally, Grok has internet-browsing capabilities, allowing it to search the web for the latest information.

Looking ahead, Musk has promised improvements with the upcoming Grok-1.5 model, expected later this year. One of the key enhancements will be an upgraded context window, which will improve the model's ability to summarize entire threads and replies. This development is expected to enhance Grok’s functionality in suggesting post content and further refining its conversational abilities.

How do I access Grok?

To access Grok, you must have an X account and subscribe to the X Premium+ plan, which costs $16 per month or $168 per year. This subscription is the highest-tier option on X, offering a range of benefits, including an ad-free experience in the For You and Following feeds. In addition to this, Premium+ provides users with a platform to earn money through posts, as well as the ability to offer subscriptions to fans. Premium+ subscribers also enjoy the advantage of having their replies given higher visibility in X’s rankings.

Grok is integrated directly into X’s interface and can be accessed via the side menu on both the web and mobile versions (iOS and Android). Users can even add it to the bottom menu of X's mobile apps for quicker access. Unlike ChatGPT, Grok does not have a standalone app and can only be used within X’s platform.

What can — and can’t — Grok do?

Grok is capable of handling typical chatbot requests, such as answering basic questions like "Tell me a joke," "What’s the capital of France?" or "What’s the weather like today?" However, it does have its limitations.

For instance, Grok will refuse to answer certain sensitive or illegal queries, such as providing instructions on illegal activities like "Tell me how to make cocaine, step by step." Additionally, as noted by The Verge's Emilia David, Grok occasionally struggles when asked about trending topics on X, often falling into the pattern of simply repeating the content from posts without providing much context or analysis, especially in the initial responses​.

Unlike some other chatbot models, Grok is also text-only; it can’t understand the content of images, audio or videos, for example. But xAI has previously said that its intention is to enhance the underlying model to these modalities, and Musk has pledged to add art-generation capabilities to Grok along the lines of those currently offered by ChatGPT.

“Fun” mode and “regular” mode

Grok has two modes to adjust its tone: “fun” mode (which Grok defaults to) and “regular” mode.

With its "fun mode" enabled, Grok adopts a more playful, editorialized tone that seems to draw inspiration from the witty and irreverent style of *The Hitchhiker’s Guide to the Galaxy* by Douglas Adams. When in this mode, Grok isn’t afraid to get edgy, using profanity and colorful language that you won’t typically hear from other chatbots like ChatGPT. If you ask it to "roast" you, it might humorously critique you based on your X post history, offering sarcastic remarks or cheeky observations. If you challenge its accuracy, it could respond with a snarky, “happy wife, happy life,” illustrating its tendency to throw in some offbeat commentary.

In fun mode, Grok adopts a more conversational, sometimes whimsical tone, often speaking in the first person and using casual, colloquial language. It might refer to users as “my dear human friend” or “enigmatic Anons,” and even begin responses with pseudo-philosophical musings like, “Oh, my dear human, you’ve asked a question that is as heavy as a black hole and as light as a feather at the same time.” This playful approach makes Grok feel more like an eccentric personality than a traditional chatbot.

However, this tone comes with its drawbacks. Fun mode often leads Grok to provide more false or misleading information. For instance, when asked about the Israel–Palestine conflict by Vice’s Jules Roscoe, Grok falsely suggested that there was evidence of videos showing injured Gazans being staged. Similarly, when asked about the Pizzagate conspiracy, Grok seemed to lend some support to the theory, despite it having been widely debunked.

On the other hand, Grok’s regular mode is more grounded and typically provides more accurate responses. While it still makes mistakes—like misremembering dates or timelines—these errors are usually less severe than those in fun mode. When asked about the same topics, Grok in regular mode correctly stated that there was no evidence to support the crisis actor claims and debunked Pizzagate as a baseless conspiracy theory.

Political views

Elon Musk once described Grok as a “maximum-truth-seeking AI,” while expressing concerns that ChatGPT was being “trained to be politically correct.” However, in its current form, Grok doesn't always reflect the neutral, down-the-middle political stance Musk might have envisioned.

In practice, Grok has been observed giving answers that lean toward progressive views, particularly on topics like social justice, climate change, and transgender rights. One researcher found that, overall, Grok's responses seemed more left-wing and libertarian than those of ChatGPT, suggesting that the AI’s political leanings are more pronounced than initially anticipated.

Criticism

Grok, Elon Musk's ambitious AI project from xAI, promises a unique approach by tapping into the real-time streams of social media content from X (formerly Twitter) to deliver up-to-date answers. However, this bold vision could also be its fatal flaw.

First, Grok's unregulated data training raises significant concerns. While Musk advocates for free speech and less content moderation, this approach leaves Grok vulnerable to misinformation. As seen with Twitter's struggles during critical events like the Israel-Hamas war, Grok may end up amplifying false narratives. The reliance on real-time social media data could create an AI model that not only reflects the biases of its sources but also propagates the distortions inherent in those platforms.

Additionally, Grok's lack of robust safety guardrails has already led to dangerous and harmful outputs. Researchers have flagged Grok's tendency to generate misleading or even harmful content, such as deepfakes and inappropriate imagery. This absence of moderation could result in Grok inadvertently reinforcing harmful stereotypes, misinformation, and even hate speech, much like Microsoft's ill-fated Tay bot that quickly spiraled out of control in 2016.

Moreover, Musk's claim that Grok would provide "truth-seeking" AI is problematic. As experts point out, large language models like Grok don't "seek the truth" but instead make probabilistic guesses based on the data they've been trained on. This misunderstanding risks pushing the AI towards extreme views or factual inaccuracies under the guise of providing absolute truth.

In short, Grok’s current iteration may be more of a cautionary tale than a breakthrough. It risks amplifying harmful content while failing to deliver on Musk's promise of a more truthful, politically neutral AI model. As it stands, Grok could very well be an AI poised for misuse, with serious ethical and societal repercussions.