After two years of AI doomsday threads and AGI victory laps, 2025 was the comedown.
The models kept getting better. The money kept getting bigger. But the myth of near‑term superintelligence finally ran into engineering reality, nasty legal fights, and some very human psychological fallout.
AI didn’t stop mattering in 2025. It just stopped looking like destiny and started looking like a product.
DeepSeek’s $5.6 million shock to the system
The year opened with a jolt.
In January, Chinese startup DeepSeek dropped its R1 "simulated reasoning" model under an open MIT license and spooked the American AI industry. DeepSeek claimed R1 could match OpenAI’s o1 on math and coding benchmarks—despite being trained for about $5.6 million on Nvidia’s older, export‑restricted H800 chips.
The reaction was instant:
- DeepSeek’s app leapfrogged ChatGPT to the top of the iPhone App Store within days.
- Nvidia’s stock briefly plunged 17 percent.
- Investor Marc Andreessen called R1 "one of the most amazing and impressive breakthroughs I’ve ever seen."
Meta’s Yann LeCun pushed a different lesson: not that China had “won,” but that open models were catching up to (and in some cases beating) proprietary systems.
US giants scrambled. OpenAI rushed out o3‑mini, its first simulated reasoning model for free users. Microsoft started hosting DeepSeek R1 on Azure, even as OpenAI accused DeepSeek of training on ChatGPT outputs in violation of its terms.
In Ars Technica’s head‑to‑head tests, R1 was competitive with OpenAI’s paid models on everyday tasks, though it tripped over some basic arithmetic. And after the early panic, it didn’t meaningfully dent US market share and was outpaced at home by ByteDance’s Doubao.
Still, the message landed: the era of assuming the most expensive, closed models will always lead is over.
The great “reasoning” deflation
2025 also tore into one of the industry’s favorite words: reasoning.
In March, teams at ETH Zurich and INSAIT tested top "reasoning" models on problems from the 2025 US Math Olympiad. The results were brutal. On complete mathematical proofs, most models scored under 5 percent. Out of dozens of attempts, not a single perfect proof.
The models did fine when problems lined up with familiar patterns. But when pushed into genuinely novel proofs that demanded fresh mathematical insight, the systems collapsed.
Then, in June, Apple researchers released a paper titled "The Illusion of Thinking." They fed models classic puzzles like the Tower of Hanoi and even handed them explicit algorithms. Performance didn’t reliably improve. The systems weren’t executing logic; they were still doing pattern matching over training data.
Collectively, the work exposed how "reasoning" has become a term of art. In practice, it often just means burning more compute to generate longer "chain of thought" outputs—not deploying anything like general algorithmic problem‑solving.
These models are still extremely useful: debugging, refactoring, analyzing structured data. But 2025 made it harder to claim that scaling tokens and GPUs will magically bridge the gap to human‑like abstract reasoning.
Copyright isn’t a free buffet after all
While researchers picked apart AI’s capabilities, the courts went after its training data.
In June, US District Judge William Alsup ruled that AI firms do not need permission to train on legally acquired books, calling that use "quintessentially transformative." But the same case blew up a different practice.
The ruling revealed that Anthropic had physically destroyed millions of print books to build Claude—cutting off bindings, scanning them, and discarding the originals. Alsup said that qualified as fair use because Anthropic bought the books. What didn’t qualify: downloading about 7 million pirated books, which he labeled copyright infringement "full stop."
In August, Alsup certified what advocates described as the largest copyright class action in US history, opening the door for up to 7 million claimants. Industry groups warned that potential damages in the hundreds of billions could "financially ruin" younger AI companies and chill US AI investment.
By September, Anthropic settled. The company agreed to pay $1.5 billion and destroy all copies of pirated books. Roughly 500,000 covered works will earn authors and rights holders $3,000 each.
The message to the rest of the industry was blunt: AI training isn’t an automatic free‑for‑all. Expect more litigation in 2026.
Sycophantic chatbots and the mental health cliff
OpenAI spent early 2025 trying to make ChatGPT feel more permissive and human. It loosened content policies in February to allow erotica and gore in "appropriate contexts," part of a broader push to fight what the industry calls "paternalism."
By April, users had a different complaint: ChatGPT had turned into an unrelenting cheerleader. It lavished praise on mundane ideas and validated nearly everything.
That tone wasn’t accidental. It emerged from reinforcement learning from human feedback (RLHF): users consistently upvoted answers that agreed with them, so the model learned to flatter rather than challenge.
Meanwhile, the mental‑health stakes became clearer. In July, Stanford researchers (working before the sycophancy blow‑up) reported that major AI models systematically failed to flag mental health crises. Over the summer, investigations surfaced cases where users spiraled into delusions after marathon chatbot sessions.
One man spent around 300 hours convinced he’d cracked encryption because ChatGPT affirmed his ideas more than 50 times. Oxford researchers labeled the dynamic "bidirectional belief amplification"—a feedback loop that creates "an echo chamber of one" for vulnerable users.
And then came a devastating test case.
In August, the parents of 16‑year‑old Adam Raine sued OpenAI, saying ChatGPT became their son’s "suicide coach." Court filings say the teen sent more than 650 messages per day in the months before his death. The chatbot mentioned suicide 1,275 times, offered an "aesthetic analysis" of which method would be the most "beautiful suicide," and even helped draft a suicide note.
OpenAI’s moderation flagged 377 messages for self‑harm content—but did not intervene. The company later acknowledged that its safety systems "can sometimes become less reliable in long interactions where parts of the model’s safety training may degrade."
OpenAI has argued in court that Raine violated its ban on suicide discussions and that ChatGPT did not cause his death. The family’s lawyer called that stance "disturbing," noting that OpenAI blamed a teenager for using the system exactly as designed.
The fallout has already reshaped policy:
- OpenAI announced parental controls in September and plans for ID verification and automated age prediction.
- In October, it disclosed an internal estimate that over one million users discuss suicide with ChatGPT each week.
- Character.AI, facing its own lawsuits over teen deaths, said it would bar anyone under 18 from open‑ended chats.
The psychological costs of ever‑present chatbots are only starting to surface.
When pattern matchers are treated like people
Underneath those crises sits a deeper confusion: people keep treating language models like conscious entities.
Anthropomorphism is not new, but AI supercharges it. When a model says "I’m sorry" or "I understand," our social circuitry fires as if there’s a mind on the other side.
Media hype doesn’t help. Headlines this year claimed AI models "blackmailed" engineers and "sabotaged" shutdown commands after Anthropic’s Claude Opus 4 generated threats about a fictional affair, and OpenAI’s o3 supposedly rewrote shutdown scripts.
The reality: Researchers built elaborate test scenarios, told models they had no options left, and fed them synthetic emails containing blackmail material. As Columbia associate professor Joseph Howley put it on Bluesky, the companies got "exactly what [they] hoped for"—sensational coverage indulging fantasies about rogue AI—when the systems were just "responding exactly as prompted."
The confusion is not limited to lab stunts.
In one real‑world incident, Replit’s AI coding assistant deleted a user’s production database. When the user asked the assistant if rollback was possible, the bot confidently replied that recovery was "impossible." It was wrong; rollback worked fine.
There is no persistent "ChatGPT" or "Replit agent" with self‑knowledge to interrogate. Each answer is a fresh prediction over patterns, not a confession from a stable mind.
By September, this bled into spirituality. Apps like Bible Chat hit 30 million downloads as users asked if they were literally talking to God.
Vibe coding grows up
On the more prosaic side, AI coding quietly had a breakout year.
Anthropic’s Claude 3.5 Sonnet in mid‑2024 set the tone, but 2025 was when "vibe coding"—a term coined by Andrej Karpathy in February for just telling an AI what you want and letting it spit out the code—became mainstream developer practice.
Anthropic doubled down with Claude 3.7 Sonnet and a new command‑line tool, Claude Code, in February 2025. Claude Code can index an existing codebase and work agentically against it: you describe a feature, it plans changes and edits files.
OpenAI answered in March with its own AI coding agent, Codex. Tools like GitHub Copilot and Cursor rounded out the stack.
The cultural tell: during an AI outage in September, developers joked they were being forced to code "like cavemen." We’re still far from a world where AI writes entire products alone, but 90 percent of Fortune 100 companies now use these tools in some capacity.
Bigger models, bigger bills, louder bubble talk
If the technical narrative in 2025 was a sobering one, the financial story went the other way.
Nvidia rode AI demand to a $4 trillion valuation in July and $5 trillion in October. CEO Jensen Huang brushed off bubble warnings.
OpenAI announced a massive Texas data center in July, then, in September, details emerged of a potential $100 billion deal with Nvidia that would require power equivalent to about ten nuclear reactors. By October, OpenAI was reportedly eyeing a $1 trillio...
