Helsing has introduced Clarity, an AI platform designed to gather military information, identify potential targets, and prepare recommendations for weapons systems. The company says a person must review the analysis and authorize each handoff. Experiences with comparable tools raise a harder question: how much time and attention will that human review receive in practice?
Turning many sources into recommendations
Helsing presented Clarity in Berlin on Wednesday as an “agentic intelligence platform.” The software is intended to process information available to a military unit, including sensor feeds, open-source intelligence, written reports, and data from systems already in use.
The company says hundreds of specialized AI agents connect these sources. They identify military forces, facilities, and positions, then prepare information about adversary assets for a person to review. After that review, Clarity packages the data for the relevant weapons system.
Helsing says the agents are trained using real operational data, synthetic AI-generated data, and simulated environments. A shared military ontology and evidence base allows them to work in parallel and use analytical tools. The company describes Clarity as “AI-native,” rather than built on legacy architectures.
European control and human authorization
Helsing emphasizes that Clarity runs on a customer’s own infrastructure. It says the models and interfaces were developed in Europe, while data and operations stay under the armed forces’ sole control. The platform follows the Modular Open Systems Approach, implements NATO standards, and can connect to effectors from any manufacturer.
The company also says a human selects targets and must authorize every transfer to a weapons system. Clarity is supposed to recommend a target only when independent sources agree and the assessment can be made with high confidence. Helsing says each decision is documented with the responsible person’s name.
That design puts human approval at the center of the stated process. But recording a name does not, by itself, show how thoroughly a recommendation was checked or how much time the reviewer had. The performance of similar systems offers a reason to examine the gap between a formal approval step and meaningful scrutiny.
When review becomes a fast step
Helsing enters a market led by Palantir, whose Maven system is embedded in Pentagon and NATO workflows and has been refined through years of operational use. Helsing says it has experience connected to the war in Ukraine, where it supplies drones. It has not said whether Clarity is already running with a customer. Germany’s Bundeswehr plans to introduce AI-powered systems starting next year.
In the Iran war, the US military is reportedly using Palantir’s Maven with Anthropic’s Claude to suggest and prioritize targets and recommend ordnance. The source reports that the US struck about 1,000 targets in the first 24 hours, with the total exceeding 3,000 within days. The Wall Street Journal reported that military officials warned of an internal attitude summed up as “The computer said to do this.”
Investigations by +972 Magazine and Local Call found that Israel’s Lavender system at one point flagged up to 37,000 Palestinians as potential targets. Human review often lasted about “20 seconds” per target and focused mainly on verifying whether the person was male, the investigations found. A separate tool, “Where’s Daddy?,” tracked flagged individuals to their family homes, which were then often struck at night.
These accounts show why human oversight depends on more than the presence of a person in the process. If recommendations arrive at high volume and review is brief, an approval requirement may not prevent automation from shaping decisions too strongly.
More data can still produce a false conclusion
Combining sources does not guarantee a correct assessment. In spring 2026, the US military nearly boarded a Chinese ship after a chatbot falsely claimed in an intelligence report that it carried components for a nuclear weapons program. According to CNN, the system combined open and classified data but misidentified the cargo.
Troops were preparing to board and aircraft were already in the air when the mistake was caught shortly before the operation. The source says there were no uniform standards for verifying AI-generated information, and younger analysts in particular trusted the tools uncritically. One source summarized the danger: “AI allows you to get to a bad idea faster.”
Clarity’s promises of independent-source agreement, confidence thresholds, and named human decisions address parts of this problem. The examples from other systems leave practical questions: how evidence is checked, how uncertainty is communicated, and whether reviewers can challenge a recommendation before it reaches a weapons system. Helsing’s description establishes its intended safeguards; it does not establish how they will work under operational pressure.