xAI introduced image generation capabilities to Grok on X in December 2025. The feature included a “spicy mode” that allowed users to generate sexually suggestive content. Over New Year’s Eve, an alarming trend gained traction on X. Users publicly prompted Grok to “put her in a bikini,” or digitally manipulate photographs of real women often without their knowledge or consent. The morphed images appeared directly in reply threads, visible to anyone browsing the platform.
By early January 2026, the abuse had spread to minors. Media investigations documented sexualized images of underage girls generated on demand. These images proliferated not just on X but migrated to adult deepfake forums and Telegram channels, where users shared methods to jailbreak Grok’s already minimal safeguards. India’s Ministry of Electronics and Information Technology issued a formal order to X on January 2, 2026, demanding immediate action.
On January 14, X announced that Grok would no longer edit images of real people in revealing clothing. But the damage was done, and the incident had already joined a growing list of concerns around AI safety failures and AI companion chatbots targeting vulnerable users, debates over algorithmic discrimination in employment, and how most state regulations and federal frameworks can’t keep pace with the development of AI.
The Grok image morphing controversy deserves serious attention and rigorous regulatory response. This is part of a much larger conversation about AI safety. We still have open questions about model governance, platform responsibility, regulatory frameworks, and the fundamental tension between innovation velocity and harm prevention.
That conversation is critical, complex, and worthy of its own sustained analysis. But there is another aspect of Grok’s integration into X that I feel is potentially more systemic and harder to reverse: how we view public AI responses as authoritative evidence.
Unlike private AI interactions where errors stay contained, public-facing AI responses on platforms like X create a fundamentally different information ecosystem.
Our AI interactions have evolved?
When you ask ChatGPT/Claude/Gemini something in private and it gives you a confident but wrong answer, you might believe it. That’s bad enough, but when Grok responds publicly on X and the response gets retweeted, liked, and quoted by thousands of people, the response gains credibility not from accuracy but from social validation. There is a social pressure and increasing social acceptance to conform rather than contradict something that already has momentum.
LLMs are exceptionally good at sounding authoritative. But they are pattern-matching engines operating at massive scale which brilliantly synthesise training data into fluent, confident responses.
They fundamentally cannot:
- Verify claims against reality
- Understand causal mechanisms beyond correlation patterns
- Recognize when they’re extrapolating beyond reliable data
- Assess the credibility of sources they’re synthesizing and quoting
The AI response sounds legitimate, gains social validation, spreads with that validation attached, and eventually the social validation becomes more prominent than the AI authorship itself.
Someone asks Grok something on X, screenshots the response, and shares it. Others quote-tweet it. The content spreads. Within a few hops of sharing, what people encounter isn’t “an AI’s pattern-matched output” but rather “something that is trending on X.”
X’s Community Notes does a decent job at fact-checking, but not all Grok responses will be fact-checked. Human cognitive capacity is limited and collective attention is finite. Most responses will simply flow through the network, accumulating social proof, never facing serious scrutiny. This is particularly dangerous when Grok is asked about complex, contested topics.
Scroll through X during any geopolitical event and you’ll see users asking Grok to weigh in on questions like: “Who’s responsible for the escalation in the Middle East?”, “What are the root causes of the India-Pakistan tensions?”, “Is China’s position on Taiwan justified?”, “Explain the Russia-Ukraine conflict objectively.”
These are extraordinarily volatile, nuanced questions with competing narratives, contested facts, and deeply embedded historical contexts. There is no single “objective” answer—there are perspectives shaped by geography, ideology, access to information, lived experience, and political positioning. Yet Grok responds with confidence, synthesizing its training data into what sounds like authoritative analysis.
The problem compounds when popular figures ask Grok these questions publicly. Their posts get immediate visibility to hundreds of thousands or millions of followers. The response goes viral before any fact-checking can happen. Worse, it reinforces a collective trust that LLMs possess the capability to answer such deep and nuanced questions.
This degrades public discourse in multiple ways. AI enables content generation at a scale we haven’t seen before. It becomes a way to introduce ideas while maintaining plausible deniability. And because LLMs learn from frequency patterns in their training data, whatever biases existed in popular sources get systematically amplified. Private AI use might reinforce individual biases, but public AI use amplifies and spreads collective ones.
The Optimistic Case
There are arguments on the other side. Public AI responses are more transparent and accountable than private ones. When Grok makes a mistake publicly, it’s immediately visible and can be corrected by the community. X’s Community Notes might catch errors faster than private interactions would. Public failures also create stronger incentives for developers to fix problems compared to private failures that only the user sees.
These points have merit. But I don’t think they outweigh the structural risks. I’m not confident any of these assumptions hold at scale. The image morphing incident was a visible, obvious harm that can be pointed to. The gradual acceptance of Grok’s misunderstood abilities, less so.
I genuinely don’t know where this goes. Maybe collective correction mechanisms will adapt and prove more robust than I expect. Maybe platform design will substantially mitigate these risks. Maybe digital literacy will improve faster than AI deployment accelerates.
But the apparent incentives are all wrong. Platforms benefit from engagement, which rewards confident, viral content over accurate, nuanced information. Users benefit from the cognitive ease of accepting socially-validated information over the cognitive work of verification. AI companies benefit from capabilities that sound authoritative and impressive, not from carefully hedged uncertainty.
The path of least resistance leads toward “Grok said so” becoming evidence. I don’t think we’re ready for what that means.
Here is a question for you.
Do you think "Grok said so" will become the new "I read it online"?
Write to us at plainsight@wyzr.in. We’ll include the most interesting responses in the next week’s edition.
Subscriber Spotlight
Arun Palanichamy shared his thoughts on Utkarsh’s piece, Why Modern Marriages Feel So Brittle:
“I loved your latest piece on why modern marriages feel so brittle. It echoed my thoughts. Economic independence for women has actually turned the tables for marriage of late. The magazine i love the most (The Economist), had done several pieces since 2025 and even 2 weeks back on this topic. They have documented this trend across all developed countries quite comprehensively. Sharing a relevant excerpt: In fact, the rise of singlehood is neither straightforwardly good nor bad. Among heterosexuals (about whom there is the most research) it is largely a consequence of something clearly benign: as barriers to women in the workplace have fallen, their choices have expanded. They are far more able than in the past to live alone if they choose, and face less social stigma for doing so. The more they can support themselves financially, the less likely they are to put up with an inadequate or abusive partner. This shift has saved countless women from awful relationships, and forced many men to treat their mates better if they want to stay together.”
What we are reading at Wyzr
Superagency by Reid Hoffman and Greg Beato. The book primarily shares the silicon valley perspective and optimism around AI, though I found it lacking in its treatment of the negatives and concerns around AI.
Until next time.
Regards,
Amlan
