Can We Depend on Grok or Perplexity to 'Fact-Check' Posts on X?

We looked at responses generated by both bots to fact-checking queries and found that they don't really work.

While social media users on X (formerly Twitter) are increasingly using AI chatbots to fact-check posts, we found that the responses can be misleading.

(Photo: Vibhushita Singh/The Quint)

While many industries have been grappling with the fear of artificial intelligence (AI) replacing human jobs, fact-checkers have largely felt insulated from this threat, given the nuanced analysis and critical thinking their roles demand.

That, however, seems to be changing with the increased use of AI chatbot handles on X to “fact-check” posts. We looked at the increasing use of two such handles, @AskPerplexity, the X handle of AI-powered search engine Perplexity AI, and @grok, developed by Elon Musk's company xAI. 

If you are into doom scrolling on X, you must have already seen the handles of @AskPerplexity and @grok being tagged under most posts, either asking the bots to fact-check something or providing additional context.

So, Are The Bots Able To Fact-Check?

On 11 March, Padmaja Joshi, a Senior Executive Editor at English news organisation TV9, shared a segment on her X platform comparing the education levels of Uttar Pradesh and Tamil Nadu. She used the 2024 Annual Status of Education Report (ASER) to help her argument. The report assesses schooling and learning outcomes among children aged 3 to 16 in rural India. As is the nature of the platform, people started questioning and supporting her. However, she then replied with the screenshot of one such account where the user tagged Grok to fact-check her original post. 

(Photo: Screenshot/X)

The response from Grok seemed to validate the original post, declaring that there was “no agenda, just numbers talking”. 

Now, for the purpose of this newsletter, we will refrain from doing any fact-checking. To read our fact-checks, check out WebQoof. 

Coming back to the case study, now an AI bot had declared the claim to be true but another user felt the need to ask Perplexity as well.

(Photo: Screenshot/X)

Now, a second AI tool has also fact-checked the information and found it to be true. 

But wait, here comes the twist. Another user changed the question for Perplexity, which changed the response.

(Photo: Screenshot/X)

Another detailed prompt by yet another user also gave more context and showed how a direct comparison between UP and TN using the ASER survey might be misleading. 

(Photo: Screenshot/X)

This raises two questions: 

• Can we trust AI bots to fact-check information on important issues?

• Can we remove the human from the loop and expect 100% accurate information?

Should LLMs Be Deployed in Fact-Checking at All?

If you have followed any discourse on fact-checking or mis/disinformation, you might have heard how fact-checking is almost always reactive. False claims and “fake news” travel far more than a piece of fact-checked information. Keeping that in mind, we believe leveraging advanced large language models (LLM) to interact with users in real time could be a massive step in the right direction. An LLM is a tool that can understand, generate, and manipulate human language, trained on massive datasets to perform various natural language processing (NLP) tasks.

There are other benefits of using AI bots for fact-checking, such as: 

1. Scalability: AI chatbots can process vast amounts of data swiftly, enabling them to identify and address misinformation more efficiently than human fact-checkers alone.

2. Consistency: LLMs can apply uniform criteria when evaluating information, reducing the variability that might arise from human biases.

3. Accessibility: By providing real-time responses, AI chatbots make fact-checking resources readily available to a broader audience.

However, despite these advantages, deploying LLMs in fact-checking is not without challenges. As established, accuracy is a huge concern in real-time fact-checking on X. Experts argue that LLMs, despite their advanced capabilities, do not truly understand the content they process. Emily M Bender, a professor at the University of Washington, likened chatbots to "parrots" that repeat information without genuine comprehension, highlighting the risk of disseminating information without contextual understanding. 

There is also the concern that AI chatbots could inadvertently amplify misinformation. A report revealed that a Russian network utilised AI chatbots to spread disinformation, illustrating how malicious actors might exploit these technologies to disseminate false narratives.

Best Use Case of AI in Fact-Check: The Human-AI Collaboration

Integrating AI into fact-checking does not necessarily signal the replacement of human fact-checkers. Instead, a collaborative approach is emerging:

• Augmenting Human Efforts: AI can handle repetitive tasks such as scanning large datasets for potential misinformation, allowing human fact-checkers to focus on nuanced analysis and contextual evaluation.

• Tool Development: Organisations like Newtral have developed AI tools to assist in monitoring media and identifying claims worth fact-checking, thereby enhancing the efficiency of human fact-checkers. 

• Source Verification: Cross-reference AI-generated information with credible sources. Relying solely on AI outputs without verification can lead to the spread of inaccuracies. Independent fact-checking sites and tools can assist in this process. 

• Editorial Oversight: Human oversight remains crucial in interpreting AI-generated outputs, ensuring that fact-checking maintains accuracy, ethical standards, and contextual sensitivity.

Integrating AI chatbot handles like @AskPerplexity and @Grok into the fact-checking ecosystem signifies a pivotal shift in how information is verified and disseminated. While these tools offer remarkable efficiency and scalability, they are not without limitations and must be regularly updated and monitored. Fact-checkers will have to evolve and adapt to the changing technology while ensuring transparency by declaring the use of AI in their processes. Organisations must also develop and adhere to ethical guidelines for AI usage in fact-checking.

This is not the end. 

We studied the AI chatbot handles in greater detail and covered three other things: 

• Their effectiveness in responding to posts that independent fact-checkers have fact-checked. 

• Their differing responses to the same or similar queries using a second case study. 

• Good use cases of the bots that were helpful.

Why Are Users Enjoying Using LLMs?

X users are using Grok and Perplexity to ‘fact-check’ several pieces of information, which are majorly linked to politics, politicians, AI-generated visuals, and data collected about national metrics. Apart from questioning the scale of accuracy in answers given by AI bots, users should also worry about their sources when making up an answer. 

In a conversation with a user, Grok was transparent about not being completely reliable. The AI bot explained the process of scooping information vastly available to it online and on X and how this brings in a possibility of missing facts in the ocean of a “well-coordinated lie” floating around. 

In one of the queries, Grok misidentified a scene from Disney’s 101 Dalmatians movie as one from Disney's 'Lady and the Tramp’. When this error was pointed out, Grok immediately owned up to its mistake of confusing the two.

The bot clarified that it can check text-based data with more clarity than visuals and also gave a heads up to the user to “double-check” its work, a statement that a human won’t add to their thoroughly checked stories.

A screenshot of a post on X.

(Source: X/Screenshot)

So Do AI Bots Depend on Human Fact-Checkers?

We found a latest study shared by Columbia Journalism Review's Tow Center for Digital Journalism. The researchers had tested eight AI-driven search tools and found out that they collectively provided incorrect answers to more than 60 percent of the queries.

The study further noted that Perplexity answered 37 percent of the queries incorrectly while Grok answered over 94 percent of the queries incorrectly. 

When we linked the AI bots to human fact-checkers, we also noticed that these bots often referred to various fact-check websites to answer some questions.

Let’s take a look at this example. This claim has been going around on social media platforms for years where Congress leader Rahul Gandhi allegedly promises potato farmers that he will install a machine that can convert potatoes into gold. This claim has often been used by users to mock the leader. 

However, it is a clipped video and Gandhi had alleged that Prime Minister Narendra Modi made such a promise to the people. 

Leaving the claim altered and misleading, Grok and Perplexity dismissed it while Grok also quoted The Quint as its source. Our detailed fact-check on this claim can be seen here.

A screenshot of a post on X.

(Source: X/Screenshot)

The advantage of speed of AI bots can be used to its potential when asked about already fact-checked posts, for example, an AI-generated image of a massive sword was 'fact-checked' by Grok and Perplexity.

Similar topics also receive conflicting responses from the AI bots. When asked a direct question about PM Modi’s education, Grok’s response includes how there has been a controversy around this topic. The bot also states that Delhi University, Gujarat University and even an RTI reply could not confirm his degree.

Apart from giving a dodgy response itself, it also points fingers on former Minister of Education of India Smriti Irani’s educational qualifications.

A screenshot of a post on X.

(Source: X/Screenshot)

Meanwhile, in another post where Grok was asked about the reason behind PM Modi reading off a teleprompter despite being educated, the bot replied with an affirmation about the leader's BA and MA degrees in political science. 

A screenshot of a post on X.

(Source: X/Screenshot)

This was a small case study highlighting that even AI bots can be confusing, often providing inconsistent responses that leave users disoriented.

Supporting this analysis, the same CJR study found that chatbots are poor at declining to answer questions that they couldn’t answer precisely, contributing to incorrect answers instead.

Grok vs Grok: How AI Bot's Responses Contradicts Itself

Now let’s take a deep dive into a post shared on X by Indian comedian Kunal Kamra about Umar Khalid, an Indian student activist. The post read, “In a country where today Godse is praised, it is no wonder that Umar Khalid is in jail for the last 4.5 years for quoting Gandhi…”(sic).

Before dwelling into the AI chatbots’ responses, let’s understand Khalid’s case in brief. He was arrested on 13 September 2020 in connection with the Delhi Riots conspiracy case, where the Unlawful Activities Prevention Act (UAPA) was invoked against him. The trial in this case has not begun and the debate over the charges is still continuing in a Delhi court. An FIR against him carries several inconsistencies however the prosecution relies on a speech where he advocates for non-violence.

Kamra’s post draws attention to Khalid’s speech where he promotes non-violence, similar to Mahatma Gandhi. The post has garnered over 2.8 million views on the platform.

With the rapid emergence of AI chat bots, several users tagged Grok and Perplexity asking the bots to ‘fact-check’ this post. The quick replies given by the bots, however, weren’t very accurate. 

One of the answers shared by Grok says that the claim is "partly off". Read it here:

Perplexity termed Kamra’s post as one carrying misinformation while explaining that Khalid was arrested for alleged conspiracy in Delhi riots under UAPA charges. While some users did not receive answers even after tagging these AI bots.

Within a series of replies under the same post, AI bots goes on to say that some “fringe groups” are openly praising Nathuram Godse.

In another response, Grok points out how Khalid has been repeatedly denied bail where the trial hasn’t even started, highlighting Khalid as a ‘polarising figure’. 

As the bots scraped the data available on the internet to answer these queries, we noticed how its answers inclined towards highlighting Khalid’s alleged role in the riots with certain conviction. 

In an interesting turn of events, Rishi Bagree, a user on X was included in the list of top influencers spreading fake news amongst many others like @MeghUpdates, @KreatelyMedia, @thetatvaindia, @OpIndia_com etc. This list quickly went viral on X grabbing the attention of many.

However, Perplexity was quick to pitch in its differing opinion on Grok’s list. But the same AI bot, in a different response, also expanded more on this list and agreed with Grok.

A screenshot of a post on X.

(Source: X/Screenshot)

Soon enough, Bagree shared a detailed “apology” that he claimed he received from Grok AI. 

The alleged apology from Grok mentioned, “Upon reflection, I recognise that my statement lacked the necessary evidence and context to support such a claim about you. It was irresponsible of me to include your name without providing specific examples or verified sources to back up the accusation.”

Now here’s a twist, X has issued a community note on the very post of Bagree which states that this is not an apology but a response to a prompt fed to AI chat to write a very specific apology letter.

A screenshot of a post on X.

(Source: X/Screenshot)

This is an example of an AI bot using its capacity to cross check misinformation even when it includes its own bot. 

Reading the further details from the community note, Grok goes on to explain why this “apology letter” is fake and misleading and was likely to be a response generated through specific prompt.

However, Bagree then replies with a link carrying Grok’s post about actually issuing Bagree an apology.

This back-and-forth shows us how an AI bot’s response can be confusing and manipulated by twisting your questions and generating a response that fits one's bias. 

It is also interesting to note that Bagree shared this just two hours after Vivek Agnihotri shared an apology letter addressed to him by Grok AI, here too, Grok confirmed its apology while quoting some of the fact-checking websites as “biased”.

The chat bot quickly changes its stance and starts referencing OpIndia –a name which it recently included in the list of top ten X accounts spreading misinformation in India. 

Users also question privacy concerns over the chat bots remembering their private conversations with X users.

(Photo: Screenshot/X)

Amid the confusion created by AI chat bots, it reveals how its quick responses can unreliable and contradicting.

And the question still remains - Can AI bots be really be trusted to fact-check at scale as of now?