Should Israel have AI, and is how they are using it, is it right?

Yep, definitely thanks for the debate. Maybe the LGBTQIA+ community and the TPUSA community can finally understand the other side. It's good to see no hate toward each other. And we both have a bunch of letters in the names.

So Israel does indeed do some not super great things with their A.I.. So the bots don't really know what side someone is on. They can end us messing things up. Also, they can't tell who is actually fighting in the war either, meaning civilians can get attacked for no reason.

One major concern that someone could have is that Israel might be using A.I. to help decide who and what should be targeted in Gaza. Systems analyze large amounts of surveillance and intelligence data and then they suggest people, buildings, or locations as possible targets. The main problem with this is that A.I. doesn't actually "know" if a certain person is a fighter or a civilian. It looks for patterns in data but that data could end up being wrong, not complete, or be misunderstood. A person might get labeled as suspicious because of who they called, where they go, or just because they were connected with another person they consider suspicious. If they treat it as reliable without careful human investigation, innocent people can be wrongly targeted.

There's also reports that A.I.s are used to track suspected targets to their homes. That creates a serious risk to families and neighbors, because attacking someone’s home can kill many civilians who have nothing to do with the conflict.

Another concern is the speed of A.I., because it can generate potential targets much faster than humans can investigate them.

The systems are also difficult for the public (and sometimes even for those carrying out the attacks) to understand. If an algorithm makes a mistake, it can be unclear who is responsible: the programmers, the intelligence officers, the commanders, or the person who approved the strike. The lack of transparency makes it harder to investigate deaths and hold anyone accountable for it.

So the central problem is not simply that A.I. is being used. The problem is using imperfect technology in life-and-death decisions, especially when the data could be unreliable, when it isn't reviewed by humans, and the process is kept secret. In those circumstances, A.I. can make warfare more impersonal, faster, and more destructive rather than making it more precise. So maybe A.I. needs to be on pause until we get these problems fixed.

Sources:

1. +972 Magazine and Local Call — “Lavender: The AI machine directing Israel’s bombing spree in Gaza”
2. Investigates claims from Israeli intelligence officers about Lavender, Where’s Daddy?, limited human review, attacks on targets in family homes, and the use of unguided bombs.
3. 972mag.com
4. The Guardian — reporting on the Lavender investigation
5. Summarizes testimony from six Israeli intelligence officers who described Lavender as identifying suspected Hamas and Palestinian Islamic Jihad members and reported that the system’s recommendations were sometimes reviewed very quickly.
6. Wikipedia
7. The Guardian — reporting on “The Gospel” and automated targeting
8. Describes The Gospel as a system used to recommend buildings and other infrastructure as potential targets, while explaining concerns about the speed and scale of AI-assisted target generation.
9. Wikipedia
10. Human-rights and international-law analysis of AI-assisted targeting
11. Discusses the risks of false positives, automation bias, inadequate human verification, civilian harm, and the difficulty of assigning responsibility when algorithms influence lethal decisions.
12. Wikipedia1
13. Georgetown Security Studies Review — “The Dehumanization of ISR: Israel’s Use of Artificial Intelligence in Warfare”
14. Provides an overview of reported systems including The Gospel, Lavender, and Where’s Daddy?, along with concerns about surveillance, facial recognition, errors, proportionality, and excessive reliance on algorithmic recommendations.
15. Georgetown University

So the argument says Lavender has “about a 10%” error rate and that this will fall to 2% within 20 years. That needs something like some specific study, methodology, and definition of “error.” The reported 90% looks like it comes from a small sample described by intelligence sources, not from a publicly available, independently peer reviewed accuracy study. Even a highly accurate classification system wouldn't automatically make a strike lawful or morally acceptable. AI accuracy is only one part of the issue. A system can be 98% accurate at predicting whether someone resembles known militants but still be producing unlawful or unjustified strikes.

The argument treats human involvement as if it solves the problem. It doesn't automatically do so. An investigation based on interviews with Israeli intelligence officers alleged that Lavender outputs were sometimes accepted after extremely brief reviews. One source reportedly described the human role as little more than a “stamp of approval.” The IDF disputes important parts of that account and says analysts must independently verify targets. That disagreement itself means the argument should not state as fact that robust human verification is always occurring.

You says that humans also help identify and conduct strikes, so people should be criticized too. That is true, but it doesn't defend the AI systems either. AI could make the problem get worse by increasing the speed and number of targets. A human error affecting one decision is different from a flawed system generating tens of thousands of recommendations. The humans don't check everything the bots do. And if they did, what would be the point of using the bots in the first place?

The discussion of Rafael’s PUZZLE system appears to describe a separate battlefield-information product. Its advertised function is to combine large amounts of sensor and intelligence data, display relevant information, and let operators trace their recommendations back to the sources. That isn't the same as proving that Lavender is transparent, independently audited, and/or capable of explaining its classifications. In fact, publicly available expert analysis specifically raises concerns about systems such as Lavender and The Gospel. Being able to trace some information to a source is not the same as knowing why the model assigned a particular probability or how it weighed conflicting data.

The fifth section makes several specific claims, like “hundreds of millions of shekels,” decentralized regulation, and dedicated explainability laboratories—but no source is provided for them. The publicly reported national AI plans I found discuss a National AI Institute, regional “acceleration reactors,” AI-security laboratories, and a proposed ethics council. That's not the same as evidence that the government already operates dedicated explainability labs that audit public-sector decisions. Some reporting also describes uncertainty and disputes over the scope, budget, and coordination of Israel’s national AI program.

A country’s general technological sophistication does not prove that one military system is accurate or responsibly used.

The strongest criticism is not simply that AI can make mistakes. It is that AI can make mistakes faster and on a much larger scale than humans ever could.

Sources:

https://www.972mag.com/lavender-ai-israeli-army-gaza/

https://www.courthousenews.com/israel-using-ai-to-fine-tune-air-raid-alert-system/

https://www.ksl.com/article/news/features/science-and-tech/israel-upgrades-its-air-raid-warning-system/46572198

https://www.ynetnews.com/tech-and-digital/article/s1y11bs1czl

https://lieber.westpoint.edu/ai-based-targeting-gaza-surveying-expert-responses-refining-debate/

https://sg.finance.yahoo.com/news/israel-launches-national-ai-action-175500976.html

https://en.globes.co.il/en/article-israels-ai-directorate-launches-modest-national-plan-1001552634

Well, first, the types of errors humans make are different from the ones an AI will make. A human is probably going to remember which side is which, but an AI may forget in the middle of a war which side is which and start attacking its own people. The AI may not even be able to tell the difference between those fighting and those on the sidelines.

The alternative to questionable AI targeting is not “get rid of humans and AI.” It could be slower intelligence analysis, more corroboration, better surveillance, smaller or more discriminating weapons, stricter rules of engagement, or choosing not to strike when uncertainty is too high. Comparing AI-assisted targeting with completely unaided human judgment does not establish that the particular AI system improves the overall decision.

I'd like to see how the humans and AIs work together. If humans need to check so much, why use the AIs at all?

Human fatigue and bias do exist, but the relevant comparison is not human imperfection versus AI perfection. To establish superiority, the argument needs evidence about false positives, false negatives, civilian-harm rates, and performance across relevant populations and not just the general claim that humans make mistakes. Saying “criticize the workers” does not defend the system.

“More accurate than human judgment” is incomplete. A system can be accurate at predicting that someone resembles a known militant while still being poor at answering the legally important questions. A 90% classification rate could still mean thousands of false positives if the system is applied to tens of thousands of people. It also matters whether “accuracy” means overall accuracy, precision, recall, or something else. If only 10% of a large list are false positives, that can still produce substantial harm.

So about your binoculars analogy: Binoculars display information that a person can directly inspect. A system may select the data, identify patterns, rank people, assign scores, and determine who receives human attention. That makes it closer to a filtering and recommendation system than to passive observation. The key issue is not whether the missile is physically fired by a human; it is how much the algorithm shapes the set of available choices.

And also you keep making assertions like that it's just the old versions that are less accurate. Where is all the data to prove these claims?

Machine learning does not automatically “get exponentially better” as it processes more data. More biased, incomplete, or circular data can make a system confidently repeat and amplify existing errors.

AI decision-support systems may be permissible in military targeting only when they demonstrably improve the full targeting process, are independently validated, use reliable and representative data, provide explainable and contestable recommendations, preserve meaningful human review, and are subject to strict legal and civilian-harm safeguards.

Good debate, done

...the end...