Data is Everything. : 'Poisoned' Data Could Compromise Military AIs

The U.S. Army's software acquisition chief has raised concerns about the vulnerability of artificial intelligence systems to 'poisoned' data in wartime scenarios. Adversaries could subtly undermine algorithms the US plans to use in future conflicts by "poisoning" the data AI systems rely on.

The primary issue is that every machine-learning algorithm requires extensive data. The Pentagon is diligently working to collect, organise, prepare, and refine its data to ensure that analytic algorithms and emerging AIs interpret it accurately. To avoid inaccurate learning, the preparation team must eliminate any erroneous data points.

Instead of training a Large Language Model (LLM) on the open internet as others have done, the military would train it on a trusted, verified military dataset within a secure setting.

Read more about the risks and the Army's response here.