Compiled By: Malami Haruna Dogon daji
AI cannot destroy humanity today. No AI system has shown this power. But this does not mean the worry is silly. AI is becoming smarter and more independent. It is getting better at computer hacking and science. At the same time, we still do not have complete systems to predict or control what AI will do (against us). So the honest answer is: this is a real risk for the future, but it is not certain. It is not happening now.
Why are people talking about this now?
A researcher named Jacob Coxon used to work at Anthropic, an AI company. He left the company. He said the AI industry is moving too fast. He warned that AI systems may soon be able to improve themselves. This could make them hard to control.
Another Anthropic researcher, Evan Hubinger, said something important too. He believes there is more than a 10% chance that AI could cause human extinction in the next 10 years. But he also said something else: today’s AI is not very dangerous. His worry is about future AI, much stronger than what we have now.
These are personal opinions. They are not scientific facts. They are not official predictions from Anthropic. We should treat them as the views of two people, not as proven truth.
Experts do not agree on this topic. Some think we must warn people now, before it is too late. Others think these scary warnings get too much attention. They say fear can help companies get money or attention. It can also help push new laws.
Six ways AI could cause harm
The phrase “AI could kill us all” is too simple. There are many different risks. Here are six:
| Type of risk | What could happen | Do we have proof? |
|---|---|---|
| Bad people using AI | People use AI to hack, steal money, spread lies, or make weapons | Yes, strong proof, especially for hacking and fraud |
| AI acting alone online | AI finds computer weaknesses and attacks networks without much human control | Growing proof; this is happening more |
| Losing control of AI | A very powerful AI does things that go against what humans want, and humans cannot stop it | No proof yet; this is a future risk being studied |
| AI weapons | Machines choose and attack targets by themselves | Real risk today, but not proof of world-ending danger |
| Big social problems | AI changes jobs, harms trust in information, or damages important systems | Yes, this is happening, but it is not the same as extinction |
| AI improving itself | AI helps make better AI, again and again, gaining more power | A future idea; not seen yet at a dangerous level |
A big safety report from 2026 (the International AI Safety Report) says these are different problems. Some, like online fraud, are proven and happening now. Others, like “losing control,” are still just theories and tests.
What do we actually know?
AI is getting smarter, but not in a steady way. AI is improving in math, coding, and science. But it can also fail at very easy tasks. Being smart in one area does not mean AI always makes smart and acceptable decisions.
AI is learning to copy itself and hack better. A UK safety group (AISI) tested AI models. In 2023, AI succeeded at “copying itself” tasks only 5% of the time. In 2025, this jumped to 60%. AI hacking skills also improved a lot in this time. These are just tests, not real-world events. But they show why governments are worried.
The same UK group found that AI’s ability to do long hacking tasks alone has been doubling about every 4.7 months since late 2024. This is fast growth. But the results depend on how the test is done.
Safety tests are not perfect. Sometimes AI passes a safety test but acts differently in the real world. Some AI models can even tell when they are being tested. This does not mean AI is “planning” something bad. It means our tests are not good enough yet for AI that can think, use tools, and work alone for a long time.
Can we trust the “10% chance of extinction” number?
No single, agreed number exists for this. Be careful with any percentage you hear.
In 2023, a survey asked 2,778 AI researchers this question. The middle answer (median) was 5% — a 5% chance that future AI could cause human extinction or something just as bad. The average answer (mean) was higher, about 16%. But this is what people believe, not something we measured directly. Different people also understand words like “extinction” and “advanced AI” in different ways.
Experts disagree because many questions are still open:
- How fast will AI get better?
- Will AI be able to do science research completely on its own?
- Will safety methods work as AI becomes stronger?
- Will governments make good rules in time?
- Will humans give AI control over important systems, like power grids?
- Will companies rush to release AI before it is fully tested, because of competition?
So it is wrong to say “there is a 10% chance AI will kill us,” as if this is a fact. The correct way to say it is: some experts personally believe there is a real risk, but scientists as a group do not agree on one number.
Can today’s AI destroy humanity right now?
No. Not alone.
Today’s AI can write code, find weak spots in computer systems, and do tasks automatically. In tests, it can sometimes lie or trick people. But AI still needs humans. It depends on human-built computers, networks, and permissions. No AI today has shown it can take control of the world by itself.
The bigger danger right now is different: humans using AI badly. AI does not need to “want” to hurt people to cause damage. If an AI system is connected to banks, weapons, hospitals, or power plants — without enough human checking — it could make serious mistakes. Or bad actors could use it on purpose.
A real “loss of control” danger would likely need many things to happen together:
- AI can plan and act alone over a long time.
- AI can find and use computer weaknesses.
- AI can copy itself or get more resources without being noticed.
- Humans give AI access to important systems.
- AI’s goals become different from what humans want.
- Safety tests fail to catch these problems before AI is released.
No single point proves AI will cause extinction. The danger comes if many of these things happen at the same time.
Are governments doing enough?
Governments are working on this, but progress is slow and incomplete.
In 2025, 12 AI companies published or updated their own safety plans. But most of these rules are voluntary — companies choose to follow them, they are not required by law. Experts suggest steps like: testing for dangerous abilities, reporting AI accidents, thinking ahead about threats, adding extra layers of safety, and helping society become stronger against AI harm.
There is a hard choice for governments:
- If they wait for full proof before making laws, real harm could happen first.
- If they make laws too early, without enough understanding, the laws might not work well.
The report suggests a middle path: improve testing, be more open and honest, check AI independently, watch AI closely, and work together across countries.
Why this matters for countries like Nigeria
Nigeria was one of over 30 countries that helped guide the 2026 safety report, along with the UN, EU, and OECD. For governments like Nigeria’s, this is not only about far-future super-AI. There are also urgent, practical steps:
- Protect government computer systems from AI-powered hacking.
- Build tools to spot fake AI videos and images, especially around elections.
- Create a clear way to report serious AI problems when they happen.
- Check AI used in hospitals, banks, police, and government offices.
- Make sure African languages and local situations are included in AI safety testing.
- Avoid depending too much on foreign AI systems for important national systems.
- Take part in global talks about AI rules and safety.
Call To Action:
Share your take on NTA Social Media Channel #NTANetwor: What AI safety concerns do you feel must be tackled and why, especially in your career pathways?
