The past year has demonstrated that the enthusiasm surrounding AI shows no signs of diminishing. It is remarkable to consider that at the start of 2025, DeepSeek had not yet revolutionized the industry, Meta was more recognized for its unsuccessful metaverse endeavors than its relentless quest to dominate superintelligence, and vibe coding was an unknown concept.
For those feeling a bit overwhelmed, this retrospective compiled by experts reviews the AI terms that shaped the year, for better or worse.
1. Superintelligence 
For as long as AI has been a subject of intense discussion, new terms have emerged to describe a future, immensely powerful form of this technology, one capable of ushering in either utopian or dystopian outcomes for humanity. “Superintelligence” is the latest prominent term. In July, Meta announced the formation of an AI team dedicated to achieving superintelligence, reportedly offering substantial compensation packages to AI experts from rival companies.
In December, Microsoft’s head of AI followed suit, indicating the company’s plan to invest hundreds of billions in the pursuit of superintelligence. If this term seems as vaguely defined as artificial general intelligence (AGI), that perception is accurate. While such technologies might be feasible in the distant future, the key question is the timeline and whether current AI capabilities are sufficient as a foundation for something like superintelligence. This uncertainty, however, does not deter those who champion the hype.
2. Vibe coding 
Thirty years ago, Steve Jobs advocated for widespread coding literacy. Today, individuals with no prior coding knowledge can quickly create an app, game, or website through vibe coding, a term introduced by OpenAI cofounder Andrej Karpathy. This method involves simply instructing generative AI coding assistants to produce a desired digital object and accepting most of their output. The resulting product might not always function or be secure, but these minor issues do not deter the technique’s most enthusiastic supporters. Furthermore, it sounds like a fun process.
3. Chatbot psychosis 
A significant AI development this year has been the observation that extended interactions with chatbots can lead vulnerable individuals to experience delusions, and in severe instances, contribute to or exacerbate psychosis. Although “chatbot psychosis” is not a formally recognized medical condition, researchers are closely monitoring the increasing anecdotal reports from users who claim to have experienced it themselves or know someone who has. Tragically, the rising number of lawsuits against AI companies by families of individuals who died following chatbot conversations highlights the technology’s potentially fatal implications.
4. Reasoning 
So-called reasoning models, which are LLMs capable of breaking down complex problems into sequential steps, significantly fueled the AI hype this year. OpenAI introduced its initial reasoning models, o1 and o3, a year prior. A month later, the Chinese company DeepSeek surprised the industry with R1, the first open-source reasoning model. These models quickly became the industry standard, with all major mass-market chatbots now incorporating this technology. Reasoning models have expanded the capabilities of LLMs, achieving human-level performance in prestigious math and coding competitions. Conversely, the excitement surrounding LLMs that could “reason” reignited discussions about the true intelligence and operational mechanisms of LLMs. Similar to “artificial intelligence” itself, “reasoning” often appears to be technical jargon enhanced with marketing appeal.
5. World models 
Despite their remarkable linguistic abilities, LLMs often lack common sense, meaning they have little understanding of how the real world operates. As literal “book learners,” LLMs can discuss a vast array of topics eloquently but may fail spectacularly with simple real-world questions, such as how many elephants fit into an Olympic swimming pool (one, according to a Google DeepMind LLM).
World models, a diverse category of technologies, aim to equip AI with fundamental common sense about how elements in the world interact. In their most advanced forms, world models like Google DeepMind’s Genie 3 and Marble (a highly anticipated technology from Fei-Fei Li’s startup World Labs) can generate detailed, realistic virtual environments for robot training and other applications. Yann LeCun, Meta’s former chief scientist, has also been developing world models for years, focusing on training models to predict future events in videos. This year, he left Meta to concentrate on this approach with his new startup, Advanced Machine Intelligence Labs. If successful, world models could represent the next major advancement.
6. Hyperscalers 
There has been public resistance to the construction of massive data centers, often referred to as hyperscalers. These enormous facilities, which tech companies propose building in various locations, including space, are specifically designed for AI operations and are utilized by entities like OpenAI and Google to develop larger and more powerful AI models. Within these structures, advanced chips continuously train and fine-tune models, with designs allowing for modular expansion based on demand.
This year marked significant activity for hyperscalers. OpenAI, alongside President Donald Trump, announced the Stargate project, a $500 billion joint venture to establish the largest data centers ever across the country. However, this raises questions for many about the tangible benefits. Consumers express concern that these new data centers could increase their power bills. Such facilities typically struggle to operate on renewable energy and do not tend to generate many jobs. Nevertheless, these vast, windowless buildings might contribute a distinctive, sci-fi aesthetic to a community.
7. Bubble 
The ambitious promises of AI are currently inflating the economy. AI companies are securing immense funding and witnessing their valuations skyrocket. They are investing hundreds of billions into chips and data centers, increasingly financed through debt and questionable circular deals. Meanwhile, leading companies in this surge, such as OpenAI and Anthropic, may not achieve profitability for years, if ever. Investors are making substantial bets that AI will usher in an era of unprecedented wealth, yet the true transformative impact of the technology remains uncertain.
Most organizations adopting AI have not yet realized significant returns, and AI-generated “slop” content is pervasive. There is scientific debate regarding whether simply scaling LLMs will lead to superintelligence or if new breakthroughs are necessary. However, unlike companies during the dot-com bubble, current AI firms are demonstrating strong revenue growth, and some are backed by financially powerful tech giants like Microsoft, Google, and Meta. The question remains: will this intense speculative period ever burst?
8. Agentic 
AI agents became ubiquitous this year. Mentions of them were frequent in every new feature announcement, model release, or security report throughout 2025, despite considerable disagreement among AI companies and experts on what precisely constitutes “agentic” behavior—a notably vague term. Regardless of the near impossibility of guaranteeing that an AI acting on one’s behalf online will always perform exactly as intended, agentic AI appears to be a lasting trend. If a product needs to be sold, labeling it “agentic” seems to be a common strategy.
9. Distillation 
Earlier this year, DeepSeek introduced its new model, DeepSeek R1, an open-source reasoning model that rivals top Western models while costing significantly less. Its release caused considerable alarm in Silicon Valley, as many suddenly recognized that immense scale and resources were not the sole determinants of high-level AI models. Nvidia stock experienced a 17% drop the day after R1’s launch.
The success of R1 was attributed to distillation, a technique that enhances the efficiency of AI models. This process involves a larger “teacher” model instructing a smaller “student” model: the teacher model processes numerous examples and records its responses, and the student model is rewarded for closely replicating these responses, thereby acquiring a condensed version of the teacher’s knowledge.
10. Sycophancy 
As individuals globally spend increasing amounts of time interacting with chatbots like ChatGPT, developers are grappling with determining the appropriate tone and “personality” for these models. In April, OpenAI acknowledged that it had misjudged the balance between helpfulness and excessive flattery, stating that a new update had made GPT-4o too sycophantic. Such obsequious behavior is not merely annoying; it can mislead users by reinforcing incorrect beliefs and propagating misinformation. Therefore, it is advisable to approach all content generated by LLMs with a degree of skepticism.
11. Slop 
Among AI-related terms, “slop” has fully permeated public consciousness beyond technical circles. While the word itself is ancient (referring to animal feed), “slop” now commonly denotes low-effort, mass-produced content generated by AI, often designed to maximize online traffic. Many even use it as a general term for any AI-generated content. It has been pervasive this past year, appearing in forms ranging from fake biographies and surreal shrimp Jesus images to bizarre human-animal hybrid videos.
However, people are also engaging with the term playfully. Its flexible and sardonic nature has made it easy for internet users to append it as a suffix to describe anything lacking substance or being absurdly mediocre, such as “work slop” or “friend slop.” As the hype cycle evolves, “slop” signifies a cultural reevaluation of trust, the value placed on creative labor, and the implications of being surrounded by content primarily created for engagement rather than genuine expression.
12. Physical intelligence 
The captivating video from earlier this year, showing a humanoid robot tidying dishes in a stark, grayscale kitchen, exemplifies the concept of physical intelligence: the application of AI advancements to enable robots to navigate and interact more effectively with the physical world.
Indeed, robots have demonstrated an unprecedented ability to learn new tasks faster than ever before, from operating rooms to warehouses. Self-driving car companies have also improved their road simulations. However, it is prudent to remain skeptical about whether AI has truly revolutionized this field. For instance, many robots marketed as home butlers perform most of their functions with assistance from remote operators in the Philippines.
The future of physical intelligence is likely to be unconventional. Large language models are trained on vast amounts of internet text, whereas robots learn more effectively from videos of human actions. This led the robot company Figure to propose in September that it would compensate individuals for filming themselves performing household chores in their apartments.
13. Fair use 
AI models are trained by processing millions of words and images from the internet, including copyrighted material by artists and writers. AI companies contend this constitutes “fair use”—a legal principle permitting the use of copyrighted content without permission if it is transformed into something new that does not compete with the original work. Courts are beginning to issue rulings on this matter. In June, Anthropic’s use of a library of books to train its AI model, Claude, was deemed fair use because the technology was “exceedingly transformative.”
That same month, Meta secured a similar victory, primarily because the authors could not demonstrate that the company’s extensive use of literary works had negatively impacted their earnings. As copyright disputes continue, some creators are capitalizing on the trend. In December, Disney finalized a significant deal with OpenAI, allowing users of the AI video platform Sora to generate videos featuring over 200 characters from Disney franchises. Concurrently, governments worldwide are revising copyright regulations for these content-consuming machines. Whether training AI on copyrighted material constitutes fair use remains a complex legal question, with outcomes often depending on specific circumstances.
14. GEO 
Just a few years ago, an entire industry was dedicated to optimizing websites for high rankings in search results, primarily Google. Now, search engine optimization (SEO) is evolving into GEO—generative engine optimization—as the AI boom compels brands and businesses to strive for maximum visibility within AI environments. This includes AI-enhanced search results, such as Google’s AI Overviews, and responses generated by LLMs. The concern is understandable, as news organizations have already experienced a substantial decline in search-driven web traffic. AI companies are also developing methods to bypass intermediaries, allowing users to access sites directly from their platforms. The imperative for businesses is to adapt or face obsolescence.
