Artificial intelligence is changing war faster than the rules meant to control it. That is the blunt truth. I’m Ermenildo V Castro, and after years of...
Artificial Intelligence in Warfare: Why the Law Is Falling Behind Fast
Artificial intelligence is changing war faster than the rules meant to control it. That is the blunt truth. I’m Ermenildo V Castro, and after years of watching software systems grow from neat demos into blunt instruments of power, I can tell you this much: when machines help pick targets, the law does not get to move at code speed.
Key Takeaways
- AI in warfare can sort data fast, spot patterns, and support military decisions.
- The real problem is not speed alone. It is accountability.
- International law still expects humans to make careful, lawful choices.
- Many AI systems are useful for defense, but they can also make bad calls faster.
- The safest path is not blind trust. It is strict human control, clear rules, and public oversight.
Artificial intelligence in warfare means using software to help military teams read data, spot objects, predict threats, and sometimes suggest or carry out actions. It can live inside drones, radar systems, command tools, and target-spotting software. Some systems help analysts sort huge piles of images. Others help commanders decide where to look next. A few are far closer to the trigger.
That sounds neat on paper. It is not neat in real life.
When I worked on large-scale systems, I learned a simple fact the hard way: if the input is messy, the output can be garbage dressed up as certainty. War makes that worse. Smoke, bad weather, decoys, broken sensors, weak signal quality, and rushed people all feed the machine. Then the machine gives an answer with a clean screen and a confident tone. That confidence can fool tired humans.
The legal problem starts there. International humanitarian law, including the Geneva Conventions and later rules on armed conflict, expects people to make judgments about distinction, proportionality, and military necessity. In plain English, that means you must tell combatants from civilians, avoid excessive harm, and use force only when needed. A machine can assist. It should not become the moral actor.
The International Committee of the Red Cross on autonomous weapon systems has warned that the human role matters. So does the United Nations work on conventional weapons, where states keep arguing over how much machine control is too much. Frankly, that debate exists because the tech moved first and the law showed up late.
This is not science fiction. It is procurement, software updates, data pipelines, and battlefield pressure. You can call it decision support if that sounds safer. But if a system nudges the strike, the distinction starts to smell thin.
I have seen enough software projects to know that labels often hide risk. What matters is control, error rate, and who bears blame when the screen is wrong. And in war, blame is not some office memo. It can be life and death.
There is a place for images here, because war tech is not just code. It is screens, maps, operators, and decisions made under pressure.
Core benefits and why militaries keep using it
AI keeps getting bought because it is useful. No surprise there.
- Speed of analysis: AI can process satellite photos, drone feeds, and sensor data far faster than human teams. That matters when minutes count.
- Pattern detection: A model can spot a truck, a heat signature, or repeated movement across many frames. Humans miss things when they are exhausted.
- Lower workload: Command staff can get cleaner summaries instead of drowning in raw data. I’ve watched teams burn out on this exact problem in civilian systems. War only makes the overload worse.
- Better range: AI can cover more ground than a single analyst. That means more map, more scan, more eyes on more places.
- Risk reduction for troops: If used well, AI can help keep soldiers away from dangerous zones. That part is real.
- Fast decision support: In some cases, AI can suggest the next best step based on current data. The catch is that “best” is not the same as “lawful.”
Here’s the kicker. A lot of people praise AI because it looks neutral. It is not. It reflects the data, the goals, and the values built into it. If the team feeds it biased labels, poor surveillance data, or rushed policy, the output can become a polished error.
When I was at big tech companies, I learned that scale amplifies small mistakes. A 1% defect is tiny until it hits millions of users. In war, the math is harsher. A tiny defect can hit a family, a street, or a school. That is why human dignity has to stay in the room.
There is also a moral side here that gets skipped in a lot of policy talk. Good tools should serve the common good. That is basic stewardship. If a system helps defend a country, fine. If it weakens restraint, hides accountability, or makes killing easier to outsource, then it has crossed a serious line.
For a broader view on how technical systems create risk at scale, see NIST’s AI Risk Management Framework. It is not about war specifically, but the ideas about oversight, measurement, and risk control still matter.
Step-by-step guide to responsible use
There is no magic fix. Anyone selling one is selling smoke.
- Define the mission clearly
Start with a narrow purpose. Is the system for logistics, surveillance review, or target support? Do not blur those lines. I have seen teams fail because they started with “help the operator” and ended with “automate the judgment.” That jump is where trouble begins.
- Keep a human in charge
A human must make the final call on any use of force. Not a rubber stamp. A real choice. If the operator cannot override the model, the system is not merely advisory anymore.
- Test with ugly data
Use bad weather, damaged sensors, crowd scenes, and false positives. Clean demos are a lie. Real-world stress is the only thing that matters. Frankly, if a system only works in the lab, it is not ready.
- Track failure rates
Measure false alarms, missed threats, and confidence drift over time. Keep logs. Keep the logs honest. The point is not to make the chart look good. The point is to know when the model is going off the rails.
- Set legal review before deployment
Weapon reviews should happen before the system is fielded, not after a tragic clip shows up online. Many states already use Article 36 reviews under Additional Protocol I practices. That process should be strict, not ceremonial.
- Limit autonomy
Use AI for detection, sorting, and recommendation before you let it do anything more. If the system can select and engage targets without meaningful human review, the risk jumps hard. That is not caution. That is common sense.
- Train people, not just models
Operators need to know when to distrust the screen. They need to understand error modes, not just button presses. I’ve trained enough engineers to know that blind trust comes from bad training.
- Review after every incident
If the system makes a bad call, pause use and review it. No excuses. No PR fog. A just society does not shrug at failures involving life and death.
The Amnesty International position on autonomous weapons is worth reading because it explains the human cost in plain terms. If you prefer a defense-policy angle, the Center for Security and Emerging Technology has useful analysis too.
Here is the plain truth. Good process does not slow everything down for no reason. It slows bad decisions. That is the point.
Comparison table: AI-supported warfare vs. human-led targeting
| Aspect | AI-supported warfare | Human-led targeting |
|---|
| Speed | Very fast at sorting data | Slower, but more deliberate |
| Accuracy | Good in narrow tasks, weak in messy scenes | Better at context, but can tire |
| Accountability | Often unclear unless tightly controlled | Easier to assign legal responsibility |
| Bias risk | Can scale data bias fast | Can carry human bias, but easier to question |
| Transparency | Hard to explain complex models | Human reasoning can be stated plainly |
| Error scale | Errors can spread quickly | Errors are usually more contained |
| Legal fit | Tense fit with current law | Better aligned with current rules |
| Moral judgment | Weak or absent | Possible, if people are trained well |
The table is not meant to flatter humans. We make bad calls too. But at least a person can be questioned, disciplined, or stopped. A model cannot repent. It cannot feel shame. It cannot bear moral weight. That burden belongs to people.
Common mistakes to avoid
Most advice on AI in war sounds polished. It is often lazy.
- Mistake 1: Treating AI as neutral
It is not neutral. It reflects its training data, design goals, and operators. If the system is trained on narrow examples, it will struggle when the field turns ugly.
- Mistake 2: Confusing speed with wisdom
Faster is not always better. In fact, speed can make bad judgment harder to catch. I’ve seen teams celebrate automation while missing the quiet damage underneath.
- Mistake 3: Hiding behind vague terms
Words like “assist,” “recommend,” and “optimize” can hide who really chose the target. That is lawyer bait. It does not solve the ethical problem.
- Mistake 4: Assuming technical audits are enough
A model review does not replace legal review. Code can be clean and still be wrong for the mission.
- Mistake 5: Ignoring civilian harm
Civilian protection cannot be an afterthought. If the system cannot reliably distinguish a threat from a bystander, it has no business near a weapon.
- Mistake 6: Letting vendors write the moral story
Vendors want contracts. That is their job. But the state owns the legal duty. Do not outsource conscience to a sales deck.
The Holy See has also spoken about the dignity of the human person in armed conflict and technology questions. Even if one does not quote church language in a defense room, the idea still matters: people are not data points. They are not expendable parts in a machine.
The truth is, a lot of failure comes from weak governance, not just weak code. Teams skip hard questions because hard questions slow them down. That is how accidents become policy.
If you want a deeper legal frame, the ICRC’s work on autonomous weapons and legal review is a strong reference. It spells out why human judgment remains central.
Frequently asked questions
Is AI already being used in warfare?
Yes. It is used for image analysis, logistics, surveillance support, threat detection, and decision help. In some systems, it can also support targeting. That last part is where the legal heat rises fast.
Can international law stop autonomous weapons?
Not fully, at least not yet. Law can set limits, ban some systems, and force review rules. But enforcement is messy, and states move at different speeds. That is the hard part.
Why do militaries want AI if it is risky?
Because it saves time, cuts workload, and can improve detection. Those are real gains. The problem is that gains often arrive before restraint does.
What is the safest policy?
Keep meaningful human control over any use of force, require legal review before deployment, and build strict audit trails. If a system cannot explain itself well enough for review, it should not be trusted with lethal action.
Final thought
I’m Ernie Castro, and I do not think this problem will be solved by fancy slogans or calm press briefings. It will be solved, if it is solved at all, by boring things: careful review, clear limits, honest logs, and people willing to say no when a machine gets too much say. That is not weakness. That is stewardship.
War already breaks enough things. We should not hand it a faster knife and call it progress. The better path is simple, even if it is not easy: keep humans responsible, keep law real, and keep dignity at the center. If we fail there, the code will not save us. It will only make the mistake harder to see.