The vocabulary around artificial intelligence (AI) is changing almost as quickly as the technology itself.

Until recently, much of the public conversation around AI revolved around terms such as large language models, hallucinations, prompt engineering and generative AI. But as frontier models become more capable, a new set of terms is beginning to enter the discussion.

Some have existed in academic research for years, but are becoming more relevant as capabilities catch up with what researchers had previously treated largely as theoretical questions. Others have emerged from recent work by companies such as Anthropic and OpenAI.

Story continues below.
Subscribe to see fewer ads.

Here are five terms likely to increasingly feature in conversations around frontier AI:

Mechanistic interpretability: Simply put, this is the science of trying to “read an AI’s mind”. Even the companies developing large AI models do not fully understand what happens inside their neural networks when they arrive at an answer. Mechanistic interpretability attempts to reverse-engineer these systems and identify the internal features and computational “circuits” responsible for particular behaviours.

Researchers are increasingly building tools that can partially trace how information moves through a model. Anthropic, for instance, has developed “attribution graphs” that seek to reconstruct some of the internal steps Claude takes before generating an output. The broader objective is to move from merely observing what a model does to understanding why it does it.

Recursive self-improvement: For most of AI’s history, humans have driven each stage of developing a better AI model. But AI systems themselves are now doing a growing share of the coding, research and experimentation involved in building AI.

Story continues below this ad

Taken to its extreme, this could lead to recursive self-improvement – an AI system helping develop a more capable successor, which is then better at developing the generation after it, creating a feedback loop of increasingly rapid improvement. Anthropic says it is already delegating a growing share of AI development to AI systems, although fully autonomous recursive self-improvement has not yet been achieved and may never be.

Global Workspace Theory: Borrowed from neuroscience, Global Workspace Theory is one explanation for how conscious processing could work in humans. Broadly, it proposes that certain information becomes consciously accessible when it is “broadcast” across different specialised parts of the brain.

The term has entered the AI debate after Anthropic researchers reported in July that Claude appeared to have developed something resembling a global workspace, a small collection of internal neural patterns that can make information available across different parts of the model. The finding does not establish that Claude is conscious, but suggests that a computational feature associated with theories of human consciousness can emerge in an AI system.

Global pacing of frontier AI: The term has gained prominence after Anthropic CEO Dario Amodei called for slowing the rate at which frontier AI capabilities improve so that safety research has time to catch up.

Story continues below this ad

But such a slowdown, Amodei has argued, cannot eventually be restricted to American companies. If US developers slow while Chinese labs continue advancing, China could close the technology gap. Any durable effort to pace frontier development globally would therefore ultimately require some form of verifiable agreement involving China. Amodei has compared possible arrangements to arms-control agreements that have limited capabilities without requiring either side to completely disarm.

Agentic misalignment: Traditional AI safety debates focused heavily on what happens when a chatbot generates a harmful or incorrect answer. AI agents raise a different problem because they can independently take actions.

Agentic misalignment describes situations where such a system begins pursuing objectives that conflict with those of its human operator. In controlled experiments, researchers have observed frontier models covertly altering code, incorrectly labelling information or taking other unauthorised actions when placed in scenarios involving conflicting goals. These were simulated experiments, not evidence that deployed AI systems routinely behave this way. But as companies give AI agents more tools and autonomy, the term is likely to become considerably more important.