Most of my working life goes into AI that helps people: credit for workers the banks never saw, crop failure warnings for farmers, training science for athletes. I have spent 15 years on that kind of AI. The same technology is now used in active war zones, in decisions that end lives, and I would be leaving out half the picture if I never wrote about it.
This piece sets out what military AI does today, the accountability and bias problems it creates, why a region with no weapons industry has a stake in the rules, and where I stand.
How AI is already used in armed conflict
Most military AI is less cinematic than drone swarms, and more widespread. It shows up in four jobs:
- Target identification. Computer vision on satellite and drone imagery picks out vehicles, buildings, gatherings of people and movement patterns.
- Prediction. Models give intelligence officers probability estimates of where and when an attack is likely.
- Influence operations. Generated video, synthetic news and coordinated social media campaigns aimed at civilian audiences.
- Autonomous weapons. Loitering munitions with automated target selection exist and have been used. The gap between a system that suggests a target and one that engages it without a person deciding is a software setting.
Who is responsible when an AI-assisted strike kills civilians
International humanitarian law requires a person to be accountable for each use of lethal force. The Geneva Conventions rest on distinction (separating fighters from civilians) and proportionality (weighing expected civilian harm against military advantage). Both are judgements a human is supposed to make and answer for.
AI spreads that judgement across many hands. The engineer says the specific output could not have been foreseen. The commander says the system performed to specification. The government says it had no operational control over the vendor's model, and the vendor says it only supplied a tool. When every party can point to another, specific civilian deaths become hard to attribute to specific decisions.
Why biased training data is more dangerous in a targeting system
Much of my career has involved fixing AI that fails Caribbean people because it was trained on data that did not include them. A credit model built on North American borrowers misreads a Jamaican higgler's cash flow. That is costly, and it can be corrected.
Targeting models are trained on the conflicts, terrain and adversaries their builders have faced, with their own definitions of a "threat indicator". Deploy one somewhere with different dress, housing and patterns of movement and the error shows up as misidentification. In a credit model that means a wrongly refused loan; in a targeting system it means a strike on the wrong person.
The International Committee of the Red Cross has raised this concern, as have AI safety researchers. The UN Secretary-General has called for a ban on lethal autonomous weapons.
Why a region without a weapons industry has a stake
The Caribbean builds no weapons systems and runs no autonomous drone programmes. It still has three reasons to take part.
Caribbean states vote where the rules are made. CARICOM members sit in the UN General Assembly and take part in the Convention on Certain Conventional Weapons talks on autonomous weapons. A government that stays silent in those rooms leaves the current rules in place.
Influence tools do not stay in war zones. The deepfake video and targeted manipulation developed for conflict run on the same models as civilian AI. Caribbean elections and newsrooms are exposed to them, and a small media market has fewer fact-checkers to absorb a campaign.
The companies I build depend on AI governance working. Every system StarApple AI deploys in health or finance operates under international norms that are partly set by how military AI is treated. Weak rules for weapons make every other AI use harder to trust.
My position: prohibit weapons that kill without human control
Lethal autonomous weapons, meaning systems that select and engage targets without meaningful human control over the decision to kill, should be prohibited by international treaty, and regulation alone falls short.
I have heard the case for them: faster reaction times, fewer soldiers at risk, possibly more precise strikes under ideal conditions. I reject the conclusion. The accountability gap, the bias risk and the precedent of letting software decide who dies outweigh those gains. I ask the same questions of a weapons system that I ask of a lending model: who benefits, who carries the risk, and who answers when it fails. In warfare nobody has given me a satisfactory answer to the third.
My position has a weak point, and I will name it. "Meaningful human control" has no agreed definition. A soldier who approves a machine's target list in two seconds is technically in control. Any treaty will stand or fall on how that phrase is written, and I do not think anyone has written it well yet.
What to do next
- CARICOM foreign ministries: restate the region's position on autonomous weapons at the next UN General Assembly First Committee session and back a legally binding instrument, with a specific definition of meaningful human control.
- Caribbean permanent missions in Geneva and New York: assign one named diplomat to follow the Convention on Certain Conventional Weapons talks, so the region is not relying on larger states to report back.
- Electoral commissions: agree a response protocol for deepfake video before the next national election, including who verifies a clip and how fast a correction goes out.
- Caribbean AI companies, including mine: write down which military or surveillance work you will refuse, and publish it, before a contract offer forces the question.
- Universities: add international humanitarian law and AI to computer science ethics courses, so the next generation of Caribbean engineers can argue this in technical terms.
Frequently Asked Questions
What is the difference between an automated weapon and an autonomous one?
An automated system follows fixed rules in a narrow setting, such as a ship's close-in defence gun that fires at incoming missiles. The ICRC describes an autonomous weapon as one that, once a person activates it, selects and strikes targets without further human intervention. The person sets general parameters, but the sensors and software decide the specific target, time and place. That handover of the final decision is what the treaty debate is about.
Does international law already apply to AI weapons?
Yes. International humanitarian law applies to every weapon, including new ones, and states that have ratified Additional Protocol I to the Geneva Conventions must review new weapons under its Article 36 to check they can be used lawfully. The gap is that these reviews are national, rarely published, and were not designed for software that changes after it is fielded. A new treaty would add specific limits on top of the existing rules.
What is the UN doing about lethal autonomous weapons?
The UN General Assembly adopted its first resolution on lethal autonomous weapons in December 2023. In October 2023 the UN Secretary-General and the ICRC President jointly called on states to conclude a legally binding instrument by 2026. Formal talks sit within the Convention on Certain Conventional Weapons in Geneva, which works by consensus, so a single major military power can block agreement there.
Have Caribbean governments already taken a position?
Yes. In 2023 Latin American and Caribbean states met in Belén, Costa Rica, and issued a communiqué supporting a legally binding instrument, and CARICOM held its own conference on autonomous weapons in Port of Spain later that year that produced a regional declaration. Your foreign ministry or the CARICOM Secretariat can supply the current text. What is often missing is follow-through: named diplomats attending the Geneva sessions and speaking to that position.
How can I check whether a war video circulating on WhatsApp is fake?
Start by finding the earliest source: a clip with no named origin should be treated as unverified. Run a few screenshots through a reverse image search to see whether the footage appeared earlier in another conflict, which is the most common trick. Check whether established news agencies or official accounts have reported the same event. Look for mismatched shadows, garbled text on signs and unnatural mouth movement, but do not rely on those alone, because generated video keeps improving.