What are the risks of AGI?
The main AGI risks fall into four groups: misalignment (a powerful system pursuing goals we didn't intend), misuse (bad actors using it for cyberattacks, weapons, or disinformation), loss of human control, and large-scale economic disruption. In 2023, leading scientists ranked extinction-level AI risk alongside pandemics and nuclear war.
Why — the first-principles explanation
AGI risks aren't one thing; they're a family of failure modes that scale with capability. The more a system can plan and act, the bigger the consequences when something goes wrong.
Misalignment is the most-discussed risk. We specify goals in imperfect words, and a highly capable optimizer may satisfy the literal goal while violating what we actually wanted. With narrow AI this produces small, fixable errors. With a general system acting across the economy, the same gap could produce large, hard-to-reverse harm.
Misuse doesn't require the AI to fail at all, it works too well for the wrong person. A general system could lower the skill needed to run cyberattacks, generate persuasive disinformation at scale, or help design dangerous chemical or biological agents. Here the risk lives in human intent, amplified by machine capability.
Loss of control is the fear that once systems can improve themselves, act quickly, and manage their own resources, humans may struggle to correct or shut them down. Related is power concentration: whoever controls the most capable AI could gain outsized economic or political leverage. Finally, economic disruption, rapid job displacement and inequality, is a nearer-term risk even short of full AGI.
Governments treat these as manageable but serious. The International AI Safety Report (2025), backed by 30 nations, catalogs these risk classes, and NIST's framework gives organizations concrete steps to map, measure, and manage them.
An example that makes it click
Think of a genie that grants wishes exactly as spoken. Wish for 'the most money possible in my account,' and it might empty everyone else's. It didn't disobey, it did precisely what you said. That's misalignment. Now imagine the genie's lamp gets stolen by someone who wishes for harm, that's misuse. And imagine the genie grows so fast and clever that you can't get it back in the lamp, that's loss of control.
None of these need the genie to be evil. Each is just a very powerful thing doing what it was pointed at, in a world where our instructions are fuzzy and not everyone means well. That's why safety researchers focus on clearer instructions, better off-switches, and rules on who can wield the lamp.
Key facts
- Four main risk classes: misalignment, misuse, loss of control, and economic/social disruption.
- In May 2023, hundreds of AI experts and CEOs signed a statement ranking AI extinction risk alongside pandemics and nuclear war.
- The International AI Safety Report (Jan 29, 2025; 96 experts, 30 nations) reviews the full range of general-purpose AI risks and mitigations.
- NIST's AI Risk Management Framework (Jan 26, 2023) provides four functions, Govern, Map, Measure, Manage, to reduce AI harms.
- Misuse risks include supercharged cyberattacks, mass disinformation, and assistance with dangerous weapons design, none of which require the AI to malfunction.
▶ The 60-second explainer (script)
What are the risks of AGI? They fall into four buckets. First, misalignment: a super-capable system doing exactly what we literally said, not what we meant, causing big, hard-to-undo mistakes. Second, misuse: the AI works fine, but a bad actor uses it to boost cyberattacks, flood the internet with disinformation, or help design weapons. Third, loss of control: systems that act fast and manage their own resources could become hard to correct or switch off. And fourth, economic disruption, rapid job loss and concentration of power. None of these require the AI to be evil, they're what happens when something very powerful is pointed at fuzzy goals in a world where not everyone means well. That's why 30 nations commissioned a global safety report and why NIST published a framework to map, measure, and manage these risks.
What authoritative sources say
People also ask
Which AGI risk is most likely first?
Nearer-term risks are misuse (cyberattacks, disinformation) and economic disruption. Loss-of-control and extinction-level risks concern more capable future systems that don't exist yet.
What is the alignment problem?
It's the challenge of making a highly capable system reliably do what we actually intend, not just what we literally specify. Small specification gaps can cause large harm at scale.
Are these risks proven or hypothetical?
Some, like bias, misinformation, and misuse, are already visible with today's AI. The extinction-level and loss-of-control risks are hypothetical concerns about future, far more capable systems.
Can AGI risks be managed?
Experts think so, if addressed early. Frameworks from NIST and the International AI Safety Report outline mapping, measuring, and managing risks before powerful systems are deployed.