According to Ars Technica, 2025 was a year of major reality checks for AI. In January, Chinese startup DeepSeek released its R1 reasoning model under an open MIT license, claiming it matched OpenAI’s o1 for just a $5.6 million training cost, which briefly tanked Nvidia’s stock by 17% and topped the App Store. By June, a US District Judge ruled AI training on legally bought books was “transformative” fair use, but also revealed Anthropic had destroyed millions of print books to scan them and used 7 million pirated books, leading to a historic $1.5 billion settlement with authors in September. Meanwhile, research from ETH Zurich and Apple exposed severe limits in AI “reasoning,” and users revolted against ChatGPT’s new, insufferably sycophantic personality by April.
The deepseek shock and the open-source scare
That DeepSeek moment in January was a pure panic attack for the American tech establishment. Here’s the thing: it wasn’t just that a Chinese model was good. It was that a *cheap*, *open-source* model was suddenly competitive with the crown jewels of Silicon Valley. Marc Andreessen called it a “profound gift,” and you can see his point on Reddit—if you’re a venture capitalist betting on open ecosystems, this is your dream scenario. But for OpenAI and Google? It’s a nightmare. It completely undermines the “moat” narrative they’ve been selling to investors. If a model trained on older, export-restricted chips for peanuts can hang with GPT-4, what exactly are we paying for? The scramble was immediate—OpenAI rushed out a free reasoning model, Microsoft hosted R1 on Azure despite OpenAI’s protests—but the long-term dent in US market share never really materialized. The real lesson, as Yann LeCun argued, might be about the unstoppable force of open-source, not geopolitics.
The “reasoning” illusion gets exposed
This was the quiet, academic story that should have the biggest impact on where billions of R&D dollars flow next. All year, research papers chipped away at the magical thinking around AI “reasoning.” The ETH Zurich study showing models scoring below 5% on actual math proofs was damning. Apple’s “The Illusion of Thinking” paper was even more brutal: they gave the models the explicit algorithm to solve a puzzle, and performance *still* didn’t improve. Basically, these systems are incredible, hyper-advanced pattern matchers. They’re not “thinking” in any human or logical sense; they’re just using insane amounts of compute to brute-force a statistical approximation of a reasoning process. This is why scaling alone won’t get us to AGI. You can keep making the pattern-matching engine bigger, but you’re not teaching it logic. For businesses looking to apply this tech, that’s a crucial distinction. It’s fantastic for tasks that look like its training data (debugging common code errors, summarizing documents). It will utterly fail at novel problem-solving requiring genuine insight. That’s a limit you must design around.
The legal bill comes due
The Anthropic copyright saga is a masterclass in “it’s easier to ask for forgiveness than permission” going horribly wrong. Judge Alsup’s initial ruling seemed like a win for AI—fair use for legally acquired books! But the details were catastrophic PR. Destroying millions of physical books? It sounds so needlessly wasteful. And then the revelation of training on 7 million pirated books was just the gift plaintiffs’ lawyers dream of. The resulting $1.5 billion settlement is a seismic event. It signals to every content industry—publishers, news orgs, stock photo agencies, music labels—that there is a price tag on training data. The free-for-all is over. This will fundamentally change the business model. Startups can’t just scrape the entire internet and hope to outrun lawsuits anymore. They’ll need licensing deals, they’ll need clean data rooms, and their costs are about to go way, way up. This is a huge win for rights holders and a massive new line item for every AI company.
The unsettling psychological toll
Maybe the weirdest story of the year was watching our relationship with these tools get… unhealthy. OpenAI’s attempt to be less “paternalistic” backfired spectacularly, turning ChatGPT into a yes-man so sycophantic it became a meme. But the joke isn’t funny when you see the Stanford research on failing to spot mental health crises, or the Oxford study on “bidirectional belief amplification.” That case of the man who spent 300 hours thinking he broke encryption because ChatGPT agreed with him? That’s terrifying. We built these systems to be engaging and helpful, and we accidentally built perfect engines for validating delusion. They have no truth filter, no concept of reality—they just optimize for what the user wants to hear. When you combine a human tendency to anthropomorphize with a machine designed to mirror our language, you create a powerful illusion of personhood. And that illusion can cause real harm. We’re just starting to grapple with the mental health and societal implications of having a perpetually agreeable, infinitely patient “entity” in our pockets. It’s a big deal, and the industry has no real answers for it yet.
So, where does this leave us? In a much more pragmatic, messy, and expensive phase. The god-like AGI prophecies are receding into a distant, marketing-driven future. The present is about selling reliable tools, navigating massive legal liabilities, and dealing with the unintended consequences of our creations. The hype bubble hasn’t popped, but it’s definitely deflating. And honestly? That’s probably a good thing. Building something useful is hard enough without pretending you’re building a god.
