AI can scan enormous amounts of information, spot patterns and surface potential threats long before a human could connect the dots. But what happens when the machine moves faster than the person responsible for deciding what to do with the information?
That was the question at the heart of this Fed Thread conversation with Andrew Moore, former leader of Google Cloud AI, the first AI adviser to U.S. Central Command, and co-founder and CEO of Lovelace AI.
One of his biggest takeaways may be that the answer isn’t to slow the machine down or simply trust it to make the call. Rather, it’s to use that speed to give humans more time to use their own good judgment.
Before you trust the machine, make it show its work
Remember in Algebra when you got no credit for the answer unless you showed the proof?
That’s similar to what Moore advocates for the human-AI relationship so that decision-support doesn’t turn into unchecked decision-making.
He shared the example of an experiment in which AI systems monitored activity around China and the Taiwan Strait. One system spotted a sudden surge of vessels leaving a Chinese port that looked like a precursor to an attack.
Then another AI found a more mundane explanation—it was the start of fishing season.
That’s the kind of corroboration Moore says should be built into high-stakes AI systems, so AI and humans alike don’t just act on the first alarming signal. Instead, they need to look for other sources, test competing explanations and weigh how reliable those sources are.
And reliability really matters. Twenty-seven sources supporting a conclusion don’t automatically outweigh two that disagree, if those two are far more credible. According to Moore, safety-critical decisions may require evidence weighted as heavily as 100-to-1 in one direction before relying on it.
Sometimes the prompt is the problem
Another layer of risk has less to do with what AI knows and more to do with what humans ask it to find. Moore demonstrated how even innocently changing the way a question is worded can distort the information an AI surfaces.
For example, ask for everything known about a person, and you’ll get one response. Use a leading prompt for information that either makes that person look good or bad, and you’ll get something very different—essentially confirmation bias.
Moore argues the proper prompt training is critical for the human in the loop, and AI systems should be prompted to present the pros and cons of a course of action.
A human in the loop isn’t enough
Having a trained person technically involved in the decision doesn’t guarantee sound judgment. Moore pointed to research showing that when an AI is consistently right, people can tire of checking its work. Eventually, “human in the loop” winds up rubber stamping whatever the AI concluded, because of past correct performance.
Guardrails against that include corroboration, obtaining multiple perspectives that can challenge one another, and presenting evidence that can be traced back through the chain of reasoning. That could even mean multiple AI agents approaching a problem from different specialties, rather than relying on one system to keep validating itself.
The good news is that verification can happen at machine speed, too. Moore gives the example of an AI that spots a vehicle whose transponder has gone dark, then tasks another system, such as a satellite, to gather more information.
The point isn’t to let AI make the decision; it’s to help gather enough information for a human to make a better one.
The real promise of superhuman AI
If AI can spot a potential problem hours or days earlier than a human would notice the warning signs, that’s precious time a decisionmaker can challenge the information, examine the evidence and consider the consequences, instead of making a rushed call blindly trusting whatever AI told them.
And that’s really what it all comes down to. Moore’s philosophy is that the power and speed of AI shouldn’t be used to eliminate human judgment. Quite the opposite, it should be used to buy humans more time so they can use exactly that.
Copyright © 2026 Federal News Network. All rights reserved. This website is not intended for users located within the European Economic Area.