Why this page exists
This manual is organized as chapters, because most of these questions don’t have a one-sentence answer that stays honest. This page is the exception: short, direct answers to the specific questions people actually type into a search box, each linking to the full page where the sourcing, hedging, and nuance actually live. Treat the answer here as the on-ramp, not the destination.
What is AGI (artificial general intelligence)?
AGI is a hypothetical AI system that can match human performance across the full range of cognitive tasks a person can do, rather than excelling narrowly at one domain the way current systems do. No system meeting that bar is known to exist yet; experts disagree sharply about how close current systems are to it. See AGI, narrow AI and superintelligence: the capability ladder for the precise definitional boundaries, and the glossary for how the term is used across this manual.
How is AGI different from the AI I already use (ChatGPT, Gemini, and similar)?
Today’s widely used systems are narrow in an important sense: they’re extremely capable at language, code, and many reasoning tasks, but their competence is uneven — genuinely superhuman at some tasks, unreliable at others a person finds trivial. AGI, as the term is generally used, implies that unevenness has closed. See emergent capabilities: what appears, what only looks abrupt and the current AI capability dashboard for where that gap actually stands.
Is AGI real, or is it hype?
Both camps in this debate include credentialed researchers, which is itself informative — it means no one has a validated model of AGI timing, only extrapolations that could bend, break, or continue. This manual takes neither side as settled. See how to weigh claims about AGI for a framework for evaluating any specific claim, and AGI timeline forecasts for what labs and surveys actually say.
When will AGI happen?
Estimates span from “within a few years” to “not this century, if ever,” from people with comparable expertise — including lab leaders themselves. Google DeepMind’s CEO, for instance, put human-level AI five to ten years out in a March 2025 interview, with meaningful evidence already in play by his own account. The spread across the field as a whole is itself evidence: no one has a validated method for this prediction. See AGI timeline forecasts: what the labs and surveys say for the actual range of expert and forecasting-market estimates, sourced and dated.
Who is building AGI?
A small number of well-funded labs — OpenAI, Google DeepMind, Anthropic, Meta AI, and xAI are the names most often discussed as pursuing frontier-level general capability, alongside major efforts in China and elsewhere. See how the frontier AI labs differ for what actually distinguishes their approaches, and power concentration: why a few labs matter for why the list is so short.
Is a humanoid robot the same thing as AGI?
No — a robot is a body; AGI (or its absence) is a claim about the intelligence controlling it. Today’s humanoid robots are overwhelmingly narrow-purpose and often more teleoperated than autonomous. See the state of humanoid robotics and how AGI-era foundation models are changing what a robot body can do for the actual state of that intersection.
Will AGI (or AI generally) take my job?
The honest answer is: it depends heavily on the job, and current evidence shows transformation more often than elimination so far. See labor displacement: what the data actually shows for the real numbers, and which trades actually resist automation, and why for what holds up best and why.
How dangerous is AGI, really — and what’s the actual probability of extinction?
There is no single agreed-upon number, and this manual treats any specific probability as a scenario weight, not a fact. Surveyed AI researchers’ median estimates for extremely bad outcomes have varied by question wording alone; see AI risk estimates and expert disagreement and why expert extinction estimates diverge so widely for the actual surveys, their numbers, and why they spread so widely.
Can AGI be controlled, or “aligned” with human interests?
This is an open, actively researched problem, not a solved one. See corrigibility: systems that accept correction, scalable oversight: supervising more capable systems, and interpretability: reading what a model is doing internally for the actual state of the research, including its real limits.
What do AI companies’ own safety policies actually commit them to?
Each major lab publishes its own framework defining capability thresholds and what happens if a model crosses one — and they differ from each other in real, consequential ways, including how well each lab’s practice matches its own stated policy. See frontier AI labs’ safety frameworks, compared for the specifics, sourced to each lab’s current published document.
Has AI actually caused any real, documented harm so far?
Yes — a real, growing, and verifiable record exists, separate from hypothetical future risk. See the AI incidents timeline for a chronological, sourced record of documented incidents from 2016 through today, including a section on stories that circulated widely but turned out not to be true.
Do AI experts even agree on any of this?
Less than headlines suggest. Disagreement is substantial and genuine among credentialed people looking at the same evidence — on timelines, on risk levels, and on what current systems can actually do. That disagreement is itself a finding this manual takes seriously rather than resolving artificially; see how to weigh claims about AGI for how to read conflicting expert claims responsibly.
What should I actually do to prepare?
Start with the vocabulary in this chapter, then follow whichever reading path matches your situation — household, policy, or technical. The practical answer differs enormously by audience, which is why this manual has three of them rather than one, and it is deliberately built to stay useful whether AGI-level capability arrives in a few years or several decades from now.
Where do I go next?
If you read nothing else on this site, read what AGI is and what AGI changes first — they set the vocabulary and the consequence model everything else depends on. From there, the chapter list shows the full manual, and every page links forward and backward to the ones that inform it, so following a link is rarely a wrong turn.