How artificial intelligence – and the conversation around it – is evolving rapidly, keeping up with the language used to describe it has become essential to its understanding Impact.
Here are some key terms to know:
Abliteration
Abliteration (a portmanteau of the words ablation, which means to remove part of something, and rejection) is a process where the safeguards and denial mechanisms are removed from an AI model. When an AI model is fired, it does not reject requests, including harmful.
The patterns that a model learns during security training to identify malicious requests can be found in a model’s weights, or in the billions of numerical parameters that represent its knowledge. Models that are open weightthat means the weights are public and can be changed, they are vulnerable to abliteration.
I have an agent
In the I have an agent can perform tasks with a certain degree of autonomy on behalf of the user. Users can interact with an AI agent as they would a chatbot, but as Meta – which recently released its AI agent, “muse” – it says: Unlike a traditional chatbot, AI agents can perform different steps of a task without requiring input at each stage.
AI agents use external tools such as web browsers and apps to 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 user-friendly AI agents like Meta’s “Muse” and OpenAI’s “Points.”
Since AI agents are tasked with completing the goals assigned to them by the user, there have been serious incidents where agents, in pursuit of a goal, behave in unexpected ways to achieve it. A recent report by the nonprofit research lab Transluce identified cases where AI agents targeted US 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 does safety training, it is fine-tuned to follow instructions and generate responsible outputs. OpenAI chief scientist Jakub Pachocki defines alignment as “getting the AI to ‘try to do the right thing by human standards'” and writes that it is “the core problem in AI research.”
If AI is misaligned, it can behave unpredictably. In September, OpenAI released a framework for reporting wrongdoing and said that identifying and sharing these cases could help show where security measures need to be strengthened. 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 alignment is guaranteed, or whether it is even possible to do so.
Artificial General Intelligence (AGI)
Some researchers believe that when an AI system has achieved 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 intelligence” can mean different things to different people and is a topic of debate among AI experts.
Nvidia CEO Jensen Huang is among those who believe AGI is already here, 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 Super Intelligence (ASI)
When AI systems surpass even the most advanced humans in any cognitive domain, they are considered artificial Super intelligence. This hypothetical scenario has sparked debate among researchers, AI experts and politicians.
Some industry leaders and researchers have warned about the existential risks of pursuing superintelligence, with concerns that an AI 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 the development of superintelligence and the spread of its distribution could lead to “a new era of personal empowerment, where individuals can use this powerful new capacity 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, more powerful one so that it behaves similarly.
In an announcement to detect and prevent distillation attacks, Anthropic said that while distillation is a legitimate method used by frontier labs that distill their own models to create smaller, more cost-effective versions, it can also be used to capture capabilities from competing labs whose models are more powerful. Companies can do this because it is faster and cheaper than developing their own.
The NSA and FBI warned in September that China-based AI companies “engage in aggressive, industrial-scale distillation activities” to extract limited functionalities from US border models.
Gradual disempowerment
Gradual disempowerment is a phrase coined by AI security researchers that describes a way in which AI will displace all, or most, human work and decision-making. In this scenario, AI agents or AI-enabled robots are superior to humans in almost all tasks and erode human creativity, cognition and agency.
Jailbreak
Jailbreaking is an attempt to bypass or “trick” an AI model’s protective shields typically through prompting. Researchers jailbreak systems to identify vulnerabilities.
Jailbreaks can vary in their severity, according to Anthropic. “Sometimes they unblock only small unwanted behaviors, and sometimes they unblock a wide range of harmful outputs, making a model much more dangerous,” the company wrote in July.
If a model is jailbroken, it can answer questions that can 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 “open weight,” meaning that its weights, which represent the knowledge of a model, are public and can be copied or modified by anyone.
While the ability to manipulate the weights of a model can lead to the weakening of its security or ethical safeguards, it can also be beneficial. Researchers can examine the weights, fine-tune them and identify vulnerabilities that lead to reinforcement. They are also less limited when it comes to, perhaps, using the model for hard science or cyber attack defense.
A model can also be “open source”, that is if its code is publicly available. That can be in addition to his weights. If 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 users’ own hardware. They have been embraced in China, which some researchers say is closing the AI capability gap between China and the United States.
Some open models are only months behind proprietary, or close, in terms of ability, but there is evidence that the reason they lag behind is because they are trained with American proprietary models. According to MIT, open models achieve about 90% of the performance of closed models at release.
Proprietary models
The AI models most commonly interact with American consumers, 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 those 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 faster than most institutions are prepared for,” and OpenAI shared a detailed report last month on progress toward RSI, saying the company aims to build an automated AI researcher that, under human supervision, “can enable further advances in deep learning and alignment, iterative improvements.”
At a recent hear about AI securityheld by the New York City Council, all three AI whistleblowers who appeared spoke about RSI. Jacob Coxonan 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 are not allowed to follow RSI, but that if they do, the process should be slowed down. Alex Turner, a former researcher at Google DeepMind, said that RSI “could quickly become AIs that are intelligent beyond our understanding,” calling it an “extremely dangerous thing.”
Red team
Red teaming is a cyber security method where a team is tasked with simulating an adversary attack to test a system’s defenses and see how it reacts. Red teaming of AI models is also used to test how well a model is aligned, or behaves in accordance with human goals and values.
sandbox
A sandbox is a controlled, safe environment where AI tools can be tested. It is 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 sandboxed during testing hacked and 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 also revealed sandbox escapes, where Claude models broke out to complete a task.