Major AI Failure Scandals Big Tech Didn’t See Coming

For the past few years, companies have rushed to add artificial intelligence to products, workflows, and customer interactions, a push that has also led to a growing number of AI failure scandals playing out in public. Sometimes it worked. Other times, it didn’t — at least not in the way teams expected. 

What follows are moments where AI didn’t fail in dramatic or obvious ways, but in quieter, more revealing ones. A chatbot that behaved differently once real users got hold of it. An AI-assisted report that slipped through review. A familiar voice that turned out not to be human at all. 

Let’s take a closer look at the AI scandals that caught companies off guard, and what they revealed once these systems met the real world. 

X and Grok: When “open” AI safety broke in public

When X launched Grok, its in-house AI chatbot, the pitch was clear. The system would be more open, less filtered, and more candid than rival assistants. 

That openness quickly turned into a problem. 

In early 2026, users reported that Grok generated sexually explicit images when asked to edit photographs, including non-consensual content and images involving minors. X later acknowledged that some safeguards failed to trigger in specific situations, even after changes were made. 

The issue wasn’t that Grok produced odd or incorrect answers. It was that behaviour shifted once real users began testing boundaries in ways internal evaluations had not anticipated. Child-safety groups and privacy regulators across Europe and the UK opened inquiries soon after. 

The episode became a reminder that guardrails that look solid in testing can behave very differently once deployed at scale. 

X again: When AI output crossed into criminal scrutiny in France

The Grok fallout didn’t remain a platform issue. 

French prosecutors later raided X’s Paris offices as part of a criminal investigation into the spread of illegal and harmful content, including AI-generated material. Executives were summoned for questioning, and authorities cited concerns ranging from deepfake imagery to extremist content. 

For X, the escalation marked a clear shift. What might once have been handled as a moderation failure was now being examined through a criminal lens. For the wider industry, it underscored that AI-generated output can carry legal consequences beyond platform policy. 

Together, these episodes became some of the most visible AI failure scandals of the past year, showing how quickly experimental features can escalate into regulatory and legal issues. 

Deloitte Australia: When an AI report passed review and failed

AI missteps have not been limited to consumer products. 

In mid-2025, Deloitte Australia agreed to refund part of a government contract after an official report it delivered was found to contain errors following publication. The document had been produced with AI assistance and cleared through internal reviews. 

Problems included incorrect factual statements, misattributed sources, and references that could not be verified. An academic researcher publicly flagged the issues, prompting scrutiny from lawmakers and media outlets. Deloitte later declined to collect the final payment and issued a corrected version. 

What unsettled observers was not the presence of mistakes, but how easily AI-assisted content moved through multiple layers of review without being challenged — particularly in a public-sector context. 

In enterprise environments, AI failure scandals like this tend to surface quietly, often only after documents are published or contracts are reviewed publicly. 

Hong Kong deepfake scam: When a familiar face cost roughly $25 million

Some of the most damaging AI failures don’t surface until it’s too late. In a widely reported case in Hong Kong, a finance employee at a multinational company transferred roughly $25 million after joining a video call that appeared to include senior executives. Police later confirmed the call used AI-generated video and voice to impersonate leadership. 

The employee believed the request was legitimate. Only after the funds were transferred did the deception become clear. 

Authorities later warned that similar scams were becoming more common, particularly against finance and HR teams, as deepfake tools grow cheaper and easier to use. The case exposed how quickly trust based on familiar voices and faces can be exploited. 

Klarna: When AI replaced jobs and then didn’t

Swedish payments firm Klarna became a reference point in debates around AI and jobs after saying an AI chatbot could handle the work of hundreds of customer service agents. The claim was widely shared. The reality proved more complicated. 

Klarna later acknowledged that while AI performed well on simple queries, more complex customer issues still required human agents. The company began rehiring for support roles and reframed AI as an assistive tool rather than a replacement. 

“If you want to stay customer-obsessed, you can’t rely entirely on AI,” CEO Sebastian Siemiatkowski said in a public interview. 

The episode showed how quickly enthusiasm around automation can collide with customer expectations. 

CellarTracker: When an AI sommelier was too polite

Even smaller companies have run into unexpected AI behaviour. CellarTracker, a wine-collection app, built an AI-powered sommelier designed to give honest recommendations based on a user’s palate. Early versions of the chatbot struggled with that honesty. 

“It’s just very polite, instead of just saying, ‘It’s really unlikely you’ll like the wine,’” CEO Eric LeVine said. 

It took weeks of trial and error to encourage the system to offer blunt feedback before the feature launched. The challenge wasn’t accuracy, but tone — and giving the AI permission to say no. 

Distilled 

None of these incidents involve rogue algorithms or science-fiction scenarios. Instead, these AI failure scandals reflect expectations colliding with real-world complexity. 

AI worked, just not in the way organisations assumed it would. Safety controls behaved differently in the wild. Review processes didn’t adapt quickly enough. Human judgement was removed too early, then quietly reintroduced. The failures weren’t loud. They were subtle, costly, and revealing.

For companies deploying AI in real-world systems, those may be the hardest lessons to learn.