Why AI Safety Requires Competence, Not Technological Illiteracy
Imagine two people.
One believes artificial intelligence can be dangerous and therefore decides to understand it deeply.
The other believes artificial intelligence can be dangerous and therefore decides never to learn how to use it.
At first glance, the second person may appear more cautious.
But now introduce a third person.
This person wants to misuse AI. They learn it enthusiastically. They study its capabilities. They understand automation. They understand synthetic media. They learn how AI systems can be manipulated. They become extremely proficient.
Now reconsider the two responsible people.
Which one is better equipped to recognise what the malicious person is doing? Which can investigate it? Which can design safeguards? Which can explain it to policymakers? Which can distinguish a genuine threat from technological nonsense? Which can build a countermeasure?
Suddenly technological abstinence looks considerably less like protection.
Refusal Does Not Remove the Capability
There is a simple problem with saying:
“AI is dangerous, therefore good people shouldn't learn AI.”
Your refusal to learn AI does not necessarily remove AI from somebody else's computer.
If a malicious actor retains access, the result may simply be:
Their capability remains.
Yours decreases.
Cybersecurity and Defensive Knowledge
Cybersecurity learned this lesson long ago.
Computer hacking can cause enormous harm. So should ethical computer scientists refuse to study how systems are hacked?
If they did, who would perform penetration testing? Who would discover vulnerabilities? Who would analyse malware? Who would build intrusion-detection systems? Who would investigate attacks? Who would secure critical infrastructure?
Cybersecurity works precisely because responsible specialists understand many of the same systems that attackers attempt to exploit.
Knowledge of an attack is not the same thing as committing an attack.
Competence can be defensive.
Aviation and the Demand for Expertise
Aviation provides another lesson.
Aircraft can crash.
Humanity did not respond by ensuring aviation regulators knew as little about aircraft as possible. Quite the opposite.
Safe aviation required deeper expertise. Pilots had to understand aircraft. Engineers had to understand failure. Investigators had to reconstruct accidents. Regulators had to understand what they were regulating. Meteorologists had to understand conditions affecting flight. Air-traffic controllers had to understand the system surrounding the aircraft.
The danger created a demand for competence.
Regulation Requires Understanding
Now imagine regulating advanced AI while being proudly unfamiliar with how AI works.
How does such a regulator distinguish a serious model vulnerability from marketing hype?
How do they evaluate an audit?
How do they recognise when a company is misleading them?
How do they design rules that malicious actors cannot trivially circumvent?
How do journalists investigate sophisticated deepfakes without understanding generative media?
How do teachers detect inappropriate AI use while knowing nothing about AI?
How do police investigate AI-enabled fraud?
How do doctors evaluate AI clinical tools?
How does a society defend against autonomous systems while refusing to develop expertise in autonomous systems?
At some point, refusing technological proficiency in the name of technological safety becomes contradictory.
Yesterday's Capabilities May Be Inadequate
It is rather like recognising that the other side may arrive technologically equipped for a gunfight and deciding that the responsible thing to do is become exceptionally proficient with a cutlass.
The metaphor is intentionally uncomfortable.
The point is not that society needs an AI arms race.
The point is that yesterday's capabilities may be inadequate for tomorrow's problems.
The stronger somebody believes AI will become, the stronger the argument becomes for responsible people understanding it.
Literacy, Competence and Access
That does not mean giving everybody unrestricted access to every dangerous capability.
We already distinguish between literacy, professional competence and unrestricted access elsewhere.
A doctor can understand controlled drugs without handing them freely to everyone. A cybersecurity researcher can understand malware without deploying it criminally. A nuclear engineer can understand nuclear physics without possessing a nuclear weapon.
AI governance can make similar distinctions.
What We Should Teach
And perhaps this changes what we should be teaching.
Instead of:
“Don't use AI.”
Teach:
“Don't blindly trust AI.”
Learn what it does well. Learn where it fails. Learn verification. Learn privacy. Learn security. Learn bias. Learn appropriate human oversight. Learn how malicious actors use it. Learn when not to use it. And learn how to extract legitimate value from it.
Because technological ignorance is not the same thing as technological safety.
Sometimes ignorance merely ensures that when something goes wrong, somebody else understands the technology better than you do.
Who Needs AI Literacy?
If AI really is going to become as consequential as its strongest critics believe, then the people who need AI literacy are not only AI enthusiasts.
- Teachers
- Doctors
- Lawyers
- Journalists
- Police
- Engineers
- Parents
- Regulators
- Workers
- Politicians
- Ordinary citizens
We should certainly ask who is building powerful technology.
But perhaps we should ask another question too:
Who are we leaving technologically incapable of holding them accountable?
Read the series introduction: The Asymmetric Technology Ledger →