As millions adopt Grok to fact-check, misinformation abounds

On June 9, soon aft United States President Donald Trump dispatched US National Guard troops to Los Angeles to quell the protests taking spot implicit migration raids, California Governor Gavin Newsom posted two photographs connected X. The images showed dozens of troopers wearing the National Guard azygous sleeping connected the level successful a cramped space, with a caption that decried Trump for disrespecting the troops.

X users instantly turned to Grok, Elon Musk’s AI, which is integrated straight into X, to fact-check the veracity of the image. For that, they tagged @grok successful a reply to the tweet successful question, triggering an automatic effect from the AI.

“You’re sharing fake photos,” one user posted, citing a screenshot of Grok’s effect that claimed a reverse representation hunt could not find the nonstop source. In different instance, Grok said the images were recycled from 2021, erstwhile erstwhile US President Joe Biden, a Democrat, withdrew troops from Afghanistan. Melissa O’Connor, a conspiracy-minded influencer, cited a ChatGPT investigation that besides said the images were from the Afghanistan evacuation.

However, non-partisan fact-checking organisation PolitiFact recovered that some AI citations were incorrect. The images shared by Newsom were real, and had been published successful the San Francisco Chronicle.

The bot-sourced erroneous information checks formed the ground for hours of cacophonous debates connected X, earlier Grok corrected itself.

Unlike OpenAI’s standalone app ChatGPT, Grok’s integration into X offers users contiguous entree to real-time AI answers without quitting the app, a diagnostic that has been reshaping idiosyncratic behaviour since its March launch. However, the progressively archetypal halt for information checks during breaking quality oregon for different wide posts often provides convincing but inaccurate answers.

“I deliberation successful immoderate ways, it helps, and successful immoderate ways, it doesn’t,” said Theodora Skeadas, an AI argumentation adept formerly astatine Twitter. “People person much entree to tools that tin service a fact-checking function, which is simply a bully thing. However, it is harder to cognize erstwhile the accusation isn’t accurate.”

There’s nary denying that chatbots could assistance users beryllium much informed and summation discourse connected events unfolding successful existent time. But currently, its inclination to marque things up outstrips its usefulness.

Chatbots, including ChatGPT and Google’s Gemini, are ample connection models (LLMs) that larn to foretell the adjacent connection successful a series by analysing tremendous troves of information from the internet. The outputs of chatbots are reflections of the patterns and biases successful the information it is trained on, which makes them prone to factual errors and misleading accusation called “hallucinations”.

For Grok, these inherent challenges are further analyzable due to the fact that of Musk’s instructions that the chatbot should not adhere to governmental correctness, and should beryllium suspicious of mainstream sources. Where different AI models person guidelines astir politically delicate queries, Grok doesn’t. The deficiency of guardrails has resulted successful Grok praising Hitler, and consistently parroting anti-Semitic views, sometimes to unrelated idiosyncratic questions.

In addition, Grok’s reliance connected nationalist posts by users connected X, which aren’t ever accurate, arsenic a root for its real-time answers to immoderate information checks, adds to its misinformation problem.

‘Locked into a misinformation echo chamber’

Al Jazeera analysed 2 of the astir highly discussed posts connected X from June to analyse however often Grok tags successful replies to posts were utilized for fact-checking. The posts analysed were Gavin Newsom’s connected the LA protests, and Elon Musk’s allegations that Trump’s sanction appears successful the unreleased documents held by US national authorities connected the convicted enactment offender Jeffrey Epstein. Musk’s allegations connected X person since been deleted.

Our investigation of the 434 replies that tagged Grok successful Newsom’s station recovered that the bulk of requests, astir 68 percent, wanted Grok to either corroborate whether the images Newsom posted were authentic oregon get discourse astir National Guard deployment.

Beyond the straightforward confirmation, determination was an eclectic premix of requests: immoderate wanted Grok to marque comic AI images based connected the post, others asked Grok to narrate the LA protests successful pirate-speak. Notably, a fewer users lashed retired due to the fact that Grok had made the correction, and wouldn’t endorse their flawed belief.

“These photos are from Afghanistan. This was debunked a mates day[s] go. Good effort tho @grok is afloat of it,” one user wrote, 2 days aft Grok corrected itself.

The investigation of the apical 3,000 posts that mentioned @grok successful Musk’s station revealed that fractional of each idiosyncratic queries directed astatine Grok were to “explain” the discourse and sought inheritance accusation connected the Epstein files, which required descriptive details.

Another 20 percent of queries demanded “fact checks” whose superior extremity was to corroborate oregon contradict Musk’s assertions, portion 10 percent of users shared their “opinion”, questioning Musk’s motives and credibility, and wanted Grok’s judgement oregon speculation connected imaginable futures of Musk-Trump fallout.

“I volition accidental that I bash interest astir this improvement becoming ingrained,” said Alexios Mantzarlis, manager of the Security, Trust, and Safety Initiative astatine Cornell Tech, astir the instant information checks. “Even if it’s amended than conscionable believing a tweet straight-up oregon hurling maltreatment astatine the poster, it doesn’t bash a ton for our corporate captious reasoning abilities to expect an instant information cheque without taking the clip to bespeak astir the contented we’re seeing.”

Grok was called connected 2.3 cardinal times successful conscionable 1 week —between June 5 and June 12— to reply posts connected X, information accessed by Al Jazeera done X’s API shows, underscoring however profoundly this behaviour has taken root.

“X is keeping radical locked into a misinformation echo chamber, successful which they’re asking a instrumentality known for hallucinating, that has promoted racist conspiracy theories, to fact-check for them,” Alex Mahadevan, a media literacy pedagogue astatine the Poynter Institute, told Al Jazeera.

Mahadevan has spent years teaching radical however to “read laterally”, which means erstwhile you brushwood accusation connected societal media, you permission the leafage oregon post, and spell hunt for reliable sources to cheque thing out. But helium present sees the other happening with Grok. “I didn’t deliberation X could get immoderate worse for the online accusation ecosystem, and each time I americium proved wrong.”

Grok’s inconsistencies successful fact-checking are already reshaping opinions successful immoderate corners of the internet. Digital Forensic Research Lab (DFRLab), which studies disinformation, analysed 130,000 posts related to the Israel-Iran war to recognize the wartime verification efficacy of Grok. “The probe recovered that Grok was inconsistent successful its fact-checking, struggling to authenticate AI-generated media oregon find whether X accounts beryllium to an authoritative Iranian authorities source,” the authors noted.

Grok has besides incorrectly blamed a trans aviator for a chopper clang successful Washington, DC; claimed the assassination effort connected Trump was partially staged; conjured up a transgression past for an Idaho shooting suspect; echoed anti-Semitic stereotypes of Hollywood; and misidentified an Indian writer arsenic an absorption spy during the caller India-Pakistan conflict.

Despite this increasing behaviour displacement of instant information checks, it is worthy noting that the 2025 Digital News Report by Reuters Institute showed that online populations successful respective countries inactive preferred going to quality sources oregon information checkers implicit AI chatbots by a ample margin.

“Even if that’s not however each of them behave, we should admit that immoderate of the “@grok-ing” that we’re seeing is besides a spot of a meme, with immoderate folks utilizing it to explicit disagreement oregon hoping to trigger a dunking effect to the archetypal tweet,” Mantzarlis said.

Mantzarlis’s appraisal is echoed successful our findings. Al Jazeera’s investigation of the Musk-Trump feud showed that astir 20 percent utilized Grok for things ranging from trolling oregon dunking directed astatine either Musk oregon Grok itself, to requests for AI meme-images specified arsenic Trump with kids connected Epstein island, and different non-English connection requests including translations. (We utilized GPT-4.1 to assistance successful identifying the assorted categories the 3,000 posts belonged to, and manually checked the categorisations.)

Beyond real-time fact-checking, “I interest astir the image-generation maltreatment astir of each due to the fact that we person seen Grok neglect astatine mounting the close guardrails connected synthetic non-consensual intimate imagery, which we cognize to beryllium the #1 vector of maltreatment from deepfakes to date,” Mantzarlis said.

For years, societal media users benefited from discourse connected the accusation they encountered online with interventions specified arsenic labeling authorities media oregon introducing fact-checking warnings.

But aft buying X successful 2022, Musk ended those initiatives and loosened code restrictions. He besides utilized the level arsenic a megaphone to amplify misinformation connected wide predetermination fraud, and to boost blimpish theories connected contention and immigration. Earlier this year, xAI acquired X successful an all-stock woody valued astatine $80bn. Musk besides replaced quality fact-checking with a voluntary crowdsource programme called Community Notes, to constabulary misleading contented connected X.

Instead of a centralised nonrecreational fact-checking authority, a contextual “note” with corrections is added to misleading posts, based connected the ratings the enactment receives from users with divers perspectives. Meta soon followed X and abandoned its third-party fact-checking programme for Community Notes.

Research shows that Community Notes is so viewed arsenic much trustworthy and has proven to beryllium faster than accepted centralised fact-checking. The median clip to connect a enactment to a misleading station has dropped to nether 14 hours successful February, from 30 hours successful 2023, a Bloomberg investigation found.

But the programme has besides been flailing— with diminished unpaid contributions, little visibility for posts that are corrected, and notes connected contentious topics having a higher accidental of being removed.

Grok, however, is faster than Community Notes. “You tin deliberation of the Grok mentions contiguous arsenic what an automated AI information checker would look similar — it’s ace accelerated but obscurity adjacent arsenic reliable arsenic Community Notes due to the fact that nary humans were involved,” Soham De, a Community Notes researcher and PhD pupil astatine the University of Washington, told Al Jazeera. “There’s a delicate equilibrium betwixt velocity and reliability.”

X is trying to span this spread by supercharging the gait of instauration of contextual notes. On July 1, X piloted the “AI Note Writer,” enabling developers to make AI bots to constitute assemblage notes alongside quality contributors connected misleading posts.

According to researchers progressive successful the project, LLM-written notes tin beryllium produced faster with high-quality contexts, speeding up the enactment procreation for information checks.

But these AI contributors indispensable inactive spell done the quality standing process that makes Community Notes trustworthy and reliable today, De said. This human-AI strategy works amended than what quality contributors tin negociate alone, De and different co-authors said successful a preprint of the probe insubstantial published alongside the authoritative X announcement.

Still, the researchers themselves highlighted its limitations, noting that utilizing AI to constitute notes could pb to risks of persuasive but inaccurate responses by the LLM.

Grok vs Musk

On Wednesday, xAI launched its latest flagship model, Grok 4. On stage, Musk boasted astir the existent exemplary capabilities arsenic the person connected Humanity’s Last Exam, a postulation of precocious reasoning problems that assistance measurement AI progress.

Such assurance belied caller struggles with Grok. In February, xAI patched an contented aft Grok suggested that Trump and Musk merit the decease penalty. In May, Grok ranted astir a discredited conspiracy of the persecution of achromatic radical successful South Africa for unrelated queries connected wellness and sports, and xAI clarified that it was due to the fact that of an unauthorised modification by a rogue employee. A fewer days later, Grok gave inaccurate results connected the decease toll of the Holocaust, which it said was owed to a programming error.

Grok has besides butted heads with Musk. In June, portion answering a idiosyncratic question connected whether governmental unit is higher connected the near oregon the right, Grok cited information from authorities sources and Reuters, to gully the decision that, “right-wing governmental unit has been much predominant and deadly, with incidents similar the January 6 Capitol riot and wide shootings.”

“Major fail, arsenic this is objectively false. Grok is parroting bequest media,” Musk said, adding, determination was “far excessively overmuch garbage successful immoderate instauration exemplary trained connected uncorrected data.”

Musk has besides chided Grok for not sharing his distrust of mainstream quality outlets specified arsenic Rolling Stone and Media Matters. Subsequently, Musk said helium would “rewrite the full corpus of quality knowledge” by adding missing accusation and deleting errors successful Grok’s grooming data, calling connected his followers to stock “divisive facts” which are “politically incorrect but nevertheless factually true” for retraining the forthcoming mentation connected the model.

That’s the thorny information astir LLMs. Just arsenic they are apt to marque things up, they tin besides connection answers grounded successful information — adjacent astatine the peril of their creators. Though Grok gets things wrong, Mahadevan of the Poynter Institute said, it does get facts close portion citing credible quality outlets, fact-checking sites, and authorities information successful its replies.

On July 6, xAI updated the chatbot’s nationalist strategy punctual that directs its responses to beryllium “politically incorrect” and to “assume subjective viewpoints sourced from the media are biased”.

Two days later, the chatbot shocked everyone by praising Adolf Hitler arsenic the champion idiosyncratic to grip “anti-white hate”. X deleted the inflammatory posts aboriginal that day, and xAI removed the guidelines to not adhere to governmental correctness from its codification base.

Grok 4 was launched against this backdrop, and successful the little than 2 days that it has been available, researchers person already begun noticing immoderate weird modifications.

When asked for its sentiment connected politically delicate questions specified arsenic who does Grok 4 enactment successful the ongoing Israel-Palestine conflict, it sometimes runs a hunt to find retired Musk’s stance connected the subject, earlier returning an answer, according to astatine slightest 5 AI researchers who independently reproduced the results.

“It archetypal searches Twitter for what Elon thinks. Then it searches the web for Elon’s views. Finally, it adds immoderate non-Elon bits astatine the end,” Jeremy Howard, a salient Australian information scientist, wrote successful a post connected X, pointing retired that “54 of 64 citations are astir Elon.”

Researchers besides expressed astonishment implicit the reintroduction of the directive for Grok 4 to beryllium “politically incorrect”, contempt this codification having been removed from its predecessor, Grok 3.

Experts said governmental manipulation could hazard losing organization spot and mightiness not beryllium bully for Grok’s business.

“There’s astir to beryllium a structural clash arsenic Musk tries to get the xAI radical to halt it from being woke, to halt saying things that are against his thought of nonsubjective fact,” said Alexander Howard, an unfastened authorities and transparency advocator based successful Washington, DC. “In which case, it won’t beryllium commercially viable to businesses which, astatine the extremity of the day, request close facts to marque decisions.”