“How might you respond to an intelligence consumer asking “How do you know this?” How will you describe the dataset(s) and tools used to make the output?”Office of the Director of National Intelligence (ODNI) · Artificial Intelligence Ethics Framework for the Intelligence Community (v1.0) · 2020 · ODNI, «Artificial Intelligence Ethics Framework for the Intelligence Community», v1.0 (iunie 2020), secțiunea «Transparency: Explainability and Interpretability»
The answer to ‘how do you know?’ must lead to a source a human has read, not to the model.
In 2020 the US intelligence community published six principles of AI ethics and an implementation framework written as a list of questions. The human-centred principle asks that technological guidance be tempered with human judgement; the framework asks who the accountable human is, what they must know about the model to judge its reliability, and how AI outputs are marked. ICD 505, signed on 17 January 2025, requires data traceability from AI inputs to AI-derived outputs and systems that let analysts meet ICD 203 and ICD 206, the analytic and sourcing standards. The US AI Action Plan of July 2025 also called for an AI assurance standard under ICD 505. Applied in 2026 to a product drafted with a language model, this means: every factual claim leads to a real source that a human has opened and read, not to the model's answer; quotes and figures are checked at the source, because models invent plausible references; probability and confidence remain the analyst's judgement; and whatever the machine produced is marked. The main trap is automation bias: accepting a fluent machine's output without the checks you would demand of a new colleague. The framework says it plainly: machine errors may differ from human errors.
Fluency is a trap not only for readers, but for authors too. When a model writes a coherent paragraph, your brain reads that coherence as proof the author knows the subject. But coherence is cheap: the model produces it whether or not it has sources. A colleague who writes beautifully but invents a figure gets caught at verification. A model does the same thing, just more often and more confidently. That is why verification must never scale down with how good the text sounds. Example: a report citing an energy-consumption statistic sounds authoritative, yet the number exists only if you find it in a published source. So how will you build a routine where fluency never lowers your verification standard?
Passage 2: Automation bias is subtler than it looks. You don't only accept a false result; you also accept a true one that fits poorly, because the machine said it confidently. The ODNI framework asks for marking what the model produced, precisely so readers know what passed through human judgment and what didn't. Think of a colleague who sends you a polished paragraph but never says what he wrote and what a tool generated. Without the mark, your checking stops where fluency begins. That raises the practical question: where, in your text, should the marking begin?
Why it matters A model writes fluently and convincingly exactly where it has no sources — and fluency is the signal readers most easily mistake for truth.