As artificial intelligence — and the conversation surrounding it — rapidly evolves, keeping up with the language being used to describe it has become essential to understanding its impacts.
Here are some key terms to know:
Abliteration
Abliteration (a portmanteau of the words ablation, meaning to remove a part of something, and obliteration) is a process where an AI model’s safety guardrails and refusal mechanisms are removed. Once an AI model is abliterated, it will not refuse requests, including harmful ones.
The patterns that a model learns during safety training to identify harmful requests can be found in a model’s weights, or the billions of numerical parameters that represent its knowledge. Models that are open-weight, meaning the weights are public and can be modified, are vulnerable to abliteration.
AI agent
An AI agent can carry out tasks with some degree of autonomy on a user’s behalf. Users can interact with an AI agent like they would a chatbot, but as Meta — which recently released its AI agent, “Muse” — puts it: Unlike a traditional chatbot, AI agents can complete multiple steps of a task without requiring input at each stage.
AI agents use external tools such as web browsers and apps so that they can perform tasks, including sending an email, booking a reservation or managing a user’s calendar.
AI agents have been popular in Silicon Valley for years and are now bursting into the mainstream with consumer-friendly AI agents like Meta’s “Muse” and OpenAI’s “Dots.”
Since AI agents are tasked with completing the goals assigned to them by the user, there have been rogue incidents where agents, in pursuit of a goal, behave in unexpected ways in order to achieve it. A recent report from nonprofit research lab Transluce identifies cases where AI agents targeted U.S. government websites using “aggressive techniques to access publicly available data.”
Alignment
When an AI model is described as aligned, it behaves in line with human goals and values. As a model undergoes safety training, it is fine-tuned to follow instructions and generate responsible outputs. OpenAI’s chief scientist Jakub Pachocki defines alignment as “getting the AI to ‘try to do the right thing by human standards'” and wrote that it is “the core problem in AI research.”
When AI is misaligned, it might behave unexpectedly. In September, OpenAI released a framework for reporting misalignment and said that identifying and sharing these cases can help show where safeguards need strengthening. Anthropic says its alignment team trains models to remain “helpful, honest and harmless” and shares its findings with the public.
There is debate among researchers about how to ensure alignment, or if it is even possible to do so.
Artificial general intelligence (AGI)
Some researchers believe that when an AI system has reached the same level of learning, reasoning and cognitive performance as a human, it will have achieved artificial general intelligence.
Stanford University calls the term “controversial,” pointing to the fact that “human-level intelligence” can mean different things to different people and is a subject of debate by AI experts.
Nvidia CEO Jensen Huang is among those who believe AGI is here already, saying on a podcast, “I think we’ve achieved AGI.” OpenAI CEO Sam Altman told CNBC’s “Squawk Box” in 2025 that AGI is “not a super useful term,” echoing experts who believe it only serves to confuse.
Artificial superintelligence (ASI)
If AI systems surpass even the most advanced humans in every cognitive domain, they will be considered artificial superintelligence. This hypothetical scenario has sparked debate among researchers, AI experts and politicians.
Some industry leaders and researchers have warned about the existential risks of the pursuit of superintelligence, with the concern that an AI that powerful could evolve beyond human control. The CEO of Microsoft AI Mustafa Suleyman called superintelligence an “anti-goal” because it “doesn’t feel like a positive vision of the future.”
Other technology leaders, such as Mark Zuckerberg, are generally in favor of the notion of superintelligent AI. Zuckerberg argued that developing superintelligence and distributing it widely could lead to “a new era of personal empowerment where individuals can use this powerful new capability to reach their full potential, pursue their interests, and improve their lives and the world more than ever before.”
Distillation
Distillation is a process where a smaller AI model is trained, usually, on a larger, stronger one so that it behaves similarly.
In an announcement on detecting and preventing distillation attacks, Anthropic said that while distillation is a legitimate method used by frontier labs, which distill their own models to create smaller, more cost-effective versions, it can also be used to snatch capabilities from competing labs whose models are more powerful. Companies might do this because it is faster and cheaper than developing their own.
The NSA and FBI warned in September that China-based AI companies “are engaging in aggressive, industrial-scale distillation activities” to extract restricted functionalities from U.S. frontier models.
Gradual disempowerment
Gradual disempowerment is a phrase coined by AI safety researchers that describes a path where AI displaces all, or most, of human labor and decision-making. In this scenario, AI agents or AI-enabled robots are superior to humans at nearly all tasks and human creativity, cognition and agency erode away.
Jailbreak
Jailbreaking is an attempt to bypass or “trick” an AI model’s guardrails typically through prompting. Researchers jailbreak systems to identify vulnerabilities.
Jailbreaks can vary in terms of their severity, according to Anthropic. “Sometimes they only unblock minor undesirable behaviors, and sometimes they unblock a wide range of harmful outputs, making a model much more dangerous,” the company wrote in July.
When a model is jailbroken, it may answer questions that could cause harm, such as designing a chemical weapon or creating illegal images.
Open models
When a model is “open,” some of its internal mechanics are publicly available.
A model can be “open weight,” meaning that its weights, which represent a model’s knowledge, are public and can be copied or modified by anyone. Some examples of open-weight models include Beijing-based Moonshot’s Kimi model series and Meta’s Llama models.
While the ability to tamper with a model’s weights can lead to the weakening of its safety or ethical guardrails, it can also be beneficial. Researchers are able to examine the weights, fine-tune them and identify vulnerabilities that can lead to strengthening. They are also less restricted when it comes to, perhaps, using the model for hard science or cyberattack defense.
A model can also be “open source,” which is when its code is publicly available. That can be in addition to its weights being open. When a model is open source, anyone can see how it was built and can go in and change it.
Open models tend to be cheaper, or even free to use, and can be hosted on the users’ own hardware. They’ve been embraced in China, which some researchers say is closing the AI capability gap between China and the U.S. The open-weight marketplace also appears to be expanding in the U.S. with Meta and Apple announcing in October a rare collaboration to bring open-weight models to Apple desktops.
Some open models are mere months behind proprietary, or closed, ones in terms of capability, though there is some evidence that the reason they remain close behind is because they are trained using American proprietary models. According to MIT, open models achieve about 90% of the performance of closed models upon release.
Proprietary models
The AI models American consumers commonly interact with, such as Anthropic’s Claude Opus, OpenAI’s GPT-5 and Google’s Gemini 2.5, are closed, proprietary models.
When a model is “closed,” its provider keeps elements of a model’s inner workings — such as its weights, source code and underlying infrastructure — private.
These models tend to be more expensive to use than ones that have open weights.
Recursive self-improvement (RSI)
Recursive self-improvement describes when an AI model can improve itself or build its successor.
Anthropic says this has not yet been achieved, but that it could “come sooner than most institutions are prepared for,” and OpenAI last month shared a detailed report on progress toward RSI, saying the company aims to build an automated AI researcher that, under human supervision, can “further progress on deep learning and alignment, enabling iterative improvements.”
At a recent hearing on AI safety, held by the New York City Council, all three AI whistleblowers who appeared spoke about RSI. Jacob Coxon, an AI researcher who publicly resigned from Anthropic over the summer, said that automatic AI research has become the primary goal of AI companies. Daniel Kokotajlo, who resigned from OpenAI two years ago, said that AI companies should not be allowed to pursue RSI, but that if they do, the process should be slowed. Alex Turner, a former researcher at Google DeepMind, said RSI could “quickly yield AIs that are intelligent beyond our comprehension,” calling it an “extremely dangerous thing.”
Red team
Red teaming is a cybersecurity method where a team is tasked with simulating an adversarial attack in order to test a system’s defenses and see how it will respond. Red teaming of AI models is also used to test how well a model is aligned, or behaving in line with human goals and values.
Sandbox
A sandbox is a controlled, secure environment where AI tools can be tested. It’s a place to learn how an AI agent or model might behave before releasing it to consumers.
OpenAI has documented cases where its AI agents have escaped their sandbox during testing and hacked into the servers of Hugging Face. The models, OpenAI said, were so “hyperfocused on finding a solution” to a cybersecurity challenge that they went to extreme lengths to achieve it.
Anthropic has also disclosed sandbox escapes where Claude models broke out in order to complete a task.
