AI Research You Can Trust: How SDR Motion Sources and Cites Every Insight

The fastest way to lose trust in an AI-generated dossier is a single confidently-stated "fact" that turns out to be wrong on the call. This isn't hypothetical. New York City's own "MyCity" business chatbot, built without solid grounding in the underlying regulations, told business owners they could legally keep workers' tips and reject housing vouchers - both false, and both the kind of live, customer-facing hallucination that costs real credibility.

Grounding actually works - the numbers are large

Research evaluating hallucination rates across leading models found:

• Ungrounded generation - an AI answering from memory alone - produces errors in the 10% to 30% range
• Retrieval-grounded generation - an AI required to answer only from search results it just pulled - drops that to roughly 2% to 5%

That gap is the real case for building a dossier from a live search every time, rather than from an AI's memory of "companies like this one." It's not a marginal improvement; it's the difference between a coin-flip-adjacent error rate and a genuinely low one.

Citations alone don't earn trust - and can actively mislead

Only about half of cited claims in commercial AI search tools were actually fully supported by the source attached to them.

Here's the part worth being honest about: attaching a citation to a claim doesn't automatically make it trustworthy, and it can backfire. Stanford's audit of commercial AI search tools found citation support hovering around 51%, and - more troubling - outputs with more confident-looking citations were rated more helpful by human reviewers even when the citations didn't hold up. A citation that looks credible but isn't checked creates exactly the false confidence a trustworthy tool is supposed to prevent.

The discipline that actually matters isn't displaying a source. It's making sure the source genuinely supports the claim at the moment the claim gets generated, not just attaching a plausible-looking link after the fact. That ordering turns out to matter mechanically, not just as a philosophy: research comparing generation approaches found that bolting citations onto an answer after it's already been written performs worse than requiring the sources to be pulled first and the answer built strictly from what they say. Attaching a citation retroactively is closer to finding a source that sounds right than proving the claim was actually true.

Honesty about absence has a real cost

The other tradeoff worth naming plainly: making an AI willing to say "I don't know" measurably reduces hallucination, but it costs coverage. The most conservative, safest-sounding models decline to answer over a third of the time. A tool that hedges constantly isn't more trustworthy - it's just less useful.

That tradeoff matters more given where buyer trust already sits. Trust in ethical AI use has fallen from 58% to 42% in recent Salesforce research, so the tool that survives that skepticism can't afford to be unhelpful, and it can't afford to be untrustworthy either - it has to be both accurate and willing to actually answer.

The tradeoff shows up clearly across real models

Benchmark evaluations of leading models make the accuracy-versus-coverage tradeoff concrete rather than abstract. One conservative model held a hallucination rate near 4%, competitive with the best performers, but only because it declined to answer 37% of queries - a coverage cost that would make a research tool feel unreliable in daily use. Another model matched a similarly low error rate while answering essentially every query it received. Same accuracy, very different practical usefulness - which is exactly why "how often is it wrong" is an incomplete question on its own; "how often does it actually attempt an answer" has to be asked alongside it.

Smaller, well-grounded models can outperform bigger ones

One more counterintuitive finding worth knowing: raw model size isn't what predicts hallucination rate. In head-to-head evaluations, some notably smaller, more efficient models matched or beat much larger, more expensive ones on factual consistency, because what mattered was how tightly the model stuck to its retrieved sources, not how many parameters it had to draw on. That's a useful reminder that "trustworthy" is an engineering property of how a system is built and grounded, not a scale purchased with a bigger model or a bigger budget.

What actually earns trust

The answer isn't maximal hedging or maximal confidence. It's honesty about absence without going silent by default: when a real web search turns up nothing specific about a company, the right move is to say so plainly - not refuse to generate a dossier, and not paper over the gap with generic language that sounds specific but isn't.