Long before ChatGPT, Claude, or Gemini became public-facing products, governments were using artificial intelligence to search data, classify images, screen cargo, and compare records. Police forces compared fingerprints across jurisdictions and checked license plates and facial images against expanding national and transnational databases.

None of this use of AI arrived as science fiction. Adoption was uneven, frequently overpromised, and driven by the belief that software could process material faster than existing personnel and systems. Generative artificial intelligence extends that progression into language itself.

An NSA officer can quickly request an English-language summary of hundreds of intercepted communications initially compiled by France’s DGSE. A military lawyer can request an explanation of competing legal authorities. A diplomat can ask for a comparison of the tone of public statements issued over several months. A physician in a military hospital can organize thousands of clinical notes into a coherent chronology before examining a wounded patient. Thousands of such routine requests are changing how bureaucracies process volumes of information that once required large teams and substantial time. Taken together, they mark an extraordinary shift in how bureaucracies process monumental amounts of information.

The United States has integrated artificial intelligence across numerous civilian and defense functions. Israel’s technology sector has become one of the world’s leading centers for cybersecurity, intelligence technologies, and machine learning research. China continues to invest enormous state resources in artificial intelligence spanning industrial production, surveillance, military modernization, and scientific research. Russia has pursued AI applications in defense, electronic warfare, intelligence analysis, and language processing. NATO members increasingly explore common standards for interoperability, logistics, intelligence support, and defense procurement, while the European Union simultaneously attempts to construct one of the world’s most ambitious regulatory frameworks for artificial intelligence. 

These governments disagree about almost everything that matters geopolitically. Their strategic interests collide. Their legal traditions and alliances differ and their political systems differ even more. Yet each is confronting the same practical question: how should conversational artificial intelligence participate in decisions whose consequences extend far beyond the screen on which an answer appears?

Public debate still concentrates heavily on whether governments should use generative AI at all. As with military aviation and government computing, procurement and operational use are overtaking the abstract debate over adoption. Procurement, institutional need, and strategic competition eventually settled those arguments. Governments adopted the technologies because they solved problems that existing tools could not solve efficiently. The same pattern is unfolding today. The relevant policy question is no longer whether states will employ conversational AI. They already do, and they almost certainly will do so more extensively over the coming decade.

That reality should redirect attention toward a more specific problem. Contemporary evaluation of generative AI focuses overwhelmingly on the answer produced at the end of the interaction. Engineers test hallucination rates, benchmark accuracy, security, bias, information protection, and legal compliance. Those efforts are not just important but critical for efficacy and safety, and many represent genuine advances over earlier generations of artificial intelligence. They nevertheless share a common assumption. They treat the visible response as the primary object worthy of inspection.

Every conversational AI system receives an expression written by a human being. That expression may contain shorthand, slang, errors, omitted context, conflicting evidence, uncertain chronology, or distinctions that depend upon specialist knowledge. Regardless of its quality, the system must first determine what it believes the user is asking before it can produce any answer at all. It cannot skip that step. The reconstruction may occur in fractions of a second, but it remains indispensable. The machine must construct an internal representation of the user’s reasoning, intentions, chronology, relationships, constraints, and evidentiary structure before it can continue.

That intermediate act receives remarkably little public attention despite governing everything that follows. Existing review processes concentrate on outputs, legal consequences, procurement standards, and acceptable-use policies. Yet the computational reconstruction between the prompt and the response remains largely invisible, even though every subsequent answer depends upon it. If that reconstruction subtly changes chronology, certainty, attribution, scope, or relationships among pieces of evidence, an answer may appear perfectly fluent while resting upon reasoning that the human user never intended to provide.

In government, law, diplomacy,........

© The Times of Israel (Blogs)