What happens when it answers, but ignores crucial data?
In Washington, NASA's first dedicated AI director David Salvagnini has turned the conversation about artificial intelligence toward a quieter, more insidious risk than the one most people fear. While the world worries about machines that lie visibly, Salvagnini warns that the deeper danger is the machine that tells a partial truth so convincingly that no one thinks to ask what was left unsaid. His appointment, born of a 2023 federal mandate to bring AI governance into every agency, reflects a broader reckoning with how humanity chooses to share — and surrender — its judgment to tools it does not yet fully understand.
- The threat Salvagnini identifies is not the AI that fabricates boldly, but the one that omits silently — leaving critical data out of its reasoning without any signal that something is missing.
- As AI systems embed themselves deeper into high-stakes decisions at agencies like NASA, the risk of acting on incomplete outputs without knowing they are incomplete grows more acute.
- Salvagnini is pushing a disciplined three-question framework: know where the data came from, verify its accuracy, and reduce confidence sharply when either answer is unclear.
- He is also challenging the language itself — arguing that calling these systems 'artificial intelligence' obscures their nature as tools, and that renaming them 'digital assistants' would restore the human responsibility that the term AI quietly erodes.
- NASA's creation of a dedicated AI director role signals that federal governance is beginning to catch up with the technology, though the harder work of cultural discipline around AI outputs is still ahead.
When David Salvagnini addressed space agency officials in Washington, he redirected a familiar anxiety about AI toward something less visible and more dangerous. The hallucination — the AI-generated falsehood that can be caught and corrected — is not what keeps him up at night. What concerns him is the omission error: the moment an AI delivers a coherent, confident answer while silently ignoring vast amounts of relevant data. The system offers no warning. The user has no idea what is missing. And decisions get made in that gap.
Salvagnini took over as NASA's first-ever AI director in mid-May, a role created in response to a Biden administration executive order requiring federal agencies to establish dedicated AI oversight positions. He came to the job with two decades of military experience and more than twenty years working with intelligence agencies on technology — credentials that shaped his pragmatic, accountability-first approach to the technology.
At the forum, he offered a simple but demanding framework: before trusting any AI output, ask where the data came from and whether it can be verified. If either question goes unanswered, confidence in the result should be discarded or sharply reduced. It is unglamorous work — tracing sources, interrogating assumptions — but Salvagnini sees it as non-negotiable.
Underpinning all of it is a conviction about responsibility. AI systems are tools; the humans who act on their outputs bear the consequences. He would rather the technology be called a digital assistant than artificial intelligence, believing the grander term quietly encourages people to treat machine outputs as objective truth. In a world where these systems increasingly shape consequential decisions, Salvagnini's core warning is that the silent omission — invisible, unchallenged — is more dangerous than any lie you can see.
David Salvagnini stood before a room of space agency officials in Washington and posed a question that cut to the heart of a problem most people don't think about when they worry about artificial intelligence. Everyone talks about the moment an AI system generates something obviously false—a hallucination, a made-up fact, something you can catch and correct. But what troubles Salvagnini more is the opposite: when an AI gives you an answer that sounds reasonable, that fits together logically, and you have no idea what information it left out of the calculation.
This is the danger of omission errors, and Salvagnini, who took over as NASA's first-ever director of artificial intelligence in mid-May, has made it his mission to make people understand the distinction. An AI system might produce a response while ignoring vast amounts of relevant data sitting right in front of it. The system doesn't tell you it ignored anything. You don't know what you're missing. And that's where the real risk lives.
Salvagnini's appointment came in response to an executive order issued by the Biden administration in October 2023, which directed all federal agencies to create dedicated positions overseeing artificial intelligence development and deployment. NASA, like other government bodies, needed someone to think strategically about how the technology would be used across the organization. Salvagnini, who arrived at NASA in 2023 as director of data, brought substantial credentials to the role: two decades in the U.S. military and more than twenty years working with intelligence agencies on technology implementation.
At the Washington forum, he laid out what he sees as the essential framework for trusting AI outputs. Before you accept what an AI system tells you, he argued, you need to ask three hard questions: Do we understand where this data came from? Can we verify that it's accurate? If the answer to either question is no, then your confidence in the result should be discarded or significantly reduced. It's a simple framework, but it demands discipline. It requires people to do the unglamorous work of tracing back through sources and assumptions.
What struck Salvagnini most forcefully was the need to reframe how people think about responsibility. Artificial intelligence doesn't bear responsibility for its outputs—people do. The system is a tool, and the person wielding it, the person deciding to act on its answer, carries the weight of that decision. He went further, suggesting that the term "artificial intelligence" itself might be misleading. He would prefer to call it what it actually is: a digital assistant, a tool for supporting human decision-making, not replacing it.
This distinction matters because it shifts the burden back where it belongs. In a world where AI systems are increasingly embedded in consequential decisions—from resource allocation to safety protocols—the temptation is to treat the machine's output as objective truth. Salvagnini's warning is that this temptation is precisely where danger lies. The omission error, the silent gap in the data considered, can be more consequential than any obvious falsehood because it operates invisibly. You can't correct what you don't know is missing.
Notable Quotes
Humans are responsible for AI outcomes, not the systems themselves; people must understand the origins and accuracy of data before trusting results— David Salvagnini, NASA AI Director
AI should be understood as a digital assistant for decision-making support, not as an autonomous intelligent agent— David Salvagnini
The Hearth Conversation Another angle on the story
When you say omission errors are more dangerous than false responses, what makes them harder to catch?
A false answer is like a broken clock—you can verify it against reality and see it's wrong. An omission error is silent. The AI gives you a coherent answer, but it's built on incomplete information. You don't know what data it never looked at.
So the system isn't lying, it's just... incomplete?
Exactly. And that's insidious because it feels authoritative. The answer is clean, logical, well-formed. There's no red flag telling you something's missing.
How do you even detect that? How do you know what the AI didn't consider?
That's the hard part. You have to trace back. You have to ask: what sources did this draw from? What relevant datasets exist that it didn't touch? It requires human scrutiny at every step.
And Salvagnini is saying that responsibility for catching this falls on the person using the system, not the system itself?
Yes. The AI isn't accountable—it's a tool. The person who acts on its output is. That's why he wants people to stop thinking of it as intelligent and start thinking of it as an assistant. It changes how you approach the work.
Does that shift the burden too much onto humans?
Maybe. But the alternative is pretending the machine can think for itself, and that's when real mistakes happen.