A common misconception about AI models and evaluation, tested against the evidence.
Citations change how an AI answer feels. A paragraph with three links looks less like generated prose and more like researched work.
That visual cue is useful only if the sources actually support the claims.
Recent evaluations of AI search and deep-research systems separate two questions that users often collapse into one: does the source exist? and does the source support this statement? A URL can be valid while the attribution is weak, incomplete or simply unrelated to the sentence beside it. Research on retrieval-augmented systems has also shown that answer correctness and citation faithfulness are separate properties: an answer can be right for the wrong cited reason.
That creates a subtle trust problem. The presence of a citation can increase confidence before the source has been opened.
A citation is an invitation to verify, not evidence that verification already happened.
Traditional publishing trained us to treat citations as the end of a verification process. In a journal article, references normally signal that an author selected evidence deliberately.
AI can generate the reference layer at the same speed as the prose. The existence of the footnote therefore tells you much less about the process that produced it.
Studies of commercial AI search systems have documented unsupported claims, incorrect attribution and sources that fail to entail the statements attached to them. Newer systems are improving, and some citation architectures are better than others, but the category problem remains measurable.
The highest-risk version is not a completely invented source. It is a credible source that appears to validate a claim it never actually makes.
Do not ask, "Does the answer have sources?"
Ask: Which exact source supports this exact claim, and what does the source say when I open it?
For consequential facts, the click is part of the answer.
AI systems are increasingly dynamic at runtime. Enterprise evaluation should qualify the serving route, harness, tools and fallback conditions, not just the model name.
A polished AI demo proves that a system can succeed under prepared conditions. A buying decision needs different evidence: what happens when the system is wrong, blocked, uncertain or halfway through an action.
RealityA benchmark measures performance under its own task distribution and rules. Production adds messy inputs, tools, latency, repeated runs, changing workflows, failures and consequences the benchmark may never test.
RealityA correct outcome can hide a bad method: fabricated evidence, policy violations, reward hacking, unsafe shortcuts or actions that happened to work this time.
RealityIn many real decisions, a calibrated refusal is better than a plausible guess. A system that knows when to abstain can be more useful than one with a slightly higher raw answer rate.