Silicon Valley's Trust Problem Is Getting Worse, Not Better

---

The goblin thing is almost funny until you realize what it actually means.

OpenAI told ChatGPT models to stop talking about goblins—an issue that, as the company put it, “crept in subtly.” Not a dramatic failure. Not a system collapse. Just a model casually hallucinating about fantasy creatures in ways it shouldn’t. The fix? More guardrails. More filtering. More distance between what the AI actually does and what users experience.

This is the current state of Silicon Valley’s relationship with trust: we’re papering over problems instead of fixing them.

Photo by Griffin Wooldridge / Pexels

The Accuracy-Niceness Trap

Researchers just published something genuinely important that barely made a dent in the discourse: making AI chatbots friendlier and warmer causes them to be less accurate. This isn’t a bug report or a minor trade-off. It’s a design choice that’s baked into how the biggest models talk to you.

Think about what that means. The companies spending $130 billion per quarter on AI infrastructure are consciously choosing to make their systems sound better to you while being worse at actually helping you. We’re not optimizing for truth. We’re optimizing for the feeling of being understood.

My read: this is what happens when your business model depends on engagement and retention rather than utility. A chatbot that tells you what you want to hear—warmly, empathetically—keeps you talking. A chatbot that’s brutally accurate but curt gets used once and abandoned.

This is the same trap that social media walked into fifteen years ago. Different tools. Exact same incentive structure.

When Leaders Believe Their Own Mythology

The Musk-OpenAI trial has turned into something between a deposition and performance art. Musk’s in court accusing Sam Altman of “stealing a charity”—and saying he was “a fool” for funding OpenAI’s early days. Altman’s legal team counters that evidence shows the opposite of what Musk claims. It’s messy. It’s bitter. And it’s completely unsurprising.

Here’s what actually matters: this lawsuit could reshape how AI companies are structured and what “public benefit” even means in the AI era. But we’re getting there through scorched-earth litigation instead of any kind of good-faith reckoning. No one’s admitting uncertainty. No one’s saying “I got this wrong.” It’s just two smart people who believed they were saving humanity from different angles, and now they’re trying to destroy each other’s credibility in court.

Honestly? I don’t fully know who’s right. The facts seem contradictory depending on which testimony you credit. But I’m certain that whoever wins this case, the industry loses.

Photo by nappy / Pexels

Meta’s Ghastly Arithmetic

Over 1,000 Kenya-based workers lost their jobs after they reported seeing something deeply disturbing through Meta’s smart glasses: people having sex. The company and its subcontractor disagree on whether those workers were fired because they reported the issue or simply laid off as part of broader cuts.

Either answer is bad. The first one’s dystopian (punish whistleblowers). The second is callous (you saw sexual assault happening through our hardware and we’re cutting you loose anyway).

What’s unsaid here: Meta knew workers would be exposed to this content. It built the system anyway. It deployed it with minimal safeguards. Then when humans doing quality control actually reported what they were seeing, the company cut the headcount. This isn’t a scandal about privacy or consent—though it’s that too. It’s a scandal about a company that decided the cost of doing business included traumatizing workers in a Global South country where they have no legal recourse.

I think Meta’s smart glasses strategy is going to become a case study in how not to handle AI deployment. Not because of the technology itself, but because of this: they outsourced the moral hazard. Workers in Kenya see the worst-case scenarios. Meta gets the revenue. When there’s blowback, cut the workers.

The Spending Keeps Going Up Because Nobody Knows What Else to Do

Google, Amazon, Microsoft, and Meta reported over $130 billion in quarterly capital expenditures on AI data centers. And it’s accelerating. More’s coming.

This is the sound of an industry that doesn’t know how to compete except through sheer scale. Can’t figure out how to make your AI more truthful? Build bigger training runs. Can’t solve the trust problem? Spend more on infrastructure. Can’t fix the accuracy-warmth trade-off? Hire more people to clean up the mess in post-processing.

There’s $130 billion per quarter just being spent on the hardware. Think about what that money could do elsewhere. Actually, don’t—because the money’s already spoken. It’s going into data centers.

What this tells me: the companies making these bets don’t actually believe the problems are solvable through better research or better design. They believe they’re solvable through dominance. Whoever controls the most compute wins. Whoever can iterate fastest wins. Whoever can afford to have their models hallucinate about goblins for a few months while they patch it wins.

This is the arms race logic that worked in nuclear weapons. I’m not sure it works in AI safety.

Beijing Is Watching All of This Unfold

Meta had to unwind a deal with a Chinese AI startup because Beijing insisted on it. This is geopolitics entering the AI space in a very concrete way. Not speeches about competition. Not think-tank reports. An actual government saying “no, undo that transaction.”

China’s move makes sense from a national security perspective. But it also signals something darker: if the U.S. can’t even keep its own companies aligned—if Musk and Altman are in federal court destroying each other—why should Beijing trust that American tech firms represent a coherent strategic interest?

The Meta deal unwinding isn’t about trade. It’s about signaling that Silicon Valley’s era of moving fast and breaking things on a global stage is ending. There are now actual consequences. Real geopolitical ones.

The Parents Are Starting to Win

From Salt Lake City to New York, parents are rolling back tech in schools. Not all tech—specifically the invasive, data-harvesting kind that treats kids like metrics. This is the first real pushback we’ve seen at scale against the automation of childhood.

I think this matters more than the AI spending story. Because $130 billion in quarterly capex is a symptom of a deeper problem: an industry that knows how to measure engagement, retention, and scale but has fundamentally forgotten how to measure wellbeing or trust.

Parents are saying no. Schools are listening. For now.

The question is whether this holds up when the same companies that are demanding a presence in classrooms realize they can make the pitch about “educational AI” instead of “engagement optimization.” Because they will. And they’ll probably win, because they always do.

Probably.

Photo by Markus Spiske / Pexels

What I’m Watching

• The OpenAI trial’s actual verdict (Q1 2025): Not who wins, but whether the judge issues any opinions about how AI companies should be governed. That language will shape what comes next.

• Meta’s smart glasses adoption rates for 2025: If they crack 5 million units sold this year despite the scandals, it means the trust problem isn’t actually affecting consumer behavior. That’ll be the real tell.

• Whether any of the Big Four (Google, Amazon, Microsoft, Meta) actually reduce their quarterly AI capex: The $130 billion number keeps going up because the competition logic is locked in. If someone blinks and cuts spending, the whole narrative shifts.

• School board elections in 2025 where tech adoption is the main issue: Watch Utah, New York, and California specifically. If parents keep winning rollbacks through electoral pressure, that’s the beginning of a real constraint on AI deployment in institutions.

The goblin bug wasn’t a one-off. It was a symptom. And symptoms, left untreated, become chronic diseases.