AI research techniques
Not the research process — that's covered in AI Research Workflows. This is the technical layer underneath: what a model actually knows versus what it just looked up.
A distinction most people never think to make
An AI response can come from two genuinely different places, and the interface rarely makes it obvious which one you're getting. Sometimes the model is recalling a pattern from its training data — like a very well-read person answering from memory. Other times, it's actually run a live search and is summarizing what it found. These have different failure modes, different freshness, and different levels of trust they deserve, and treating them as the same thing is where a lot of research mistakes start.
What "memorized" actually means
A model's training data has a cutoff date — it doesn't know about anything that happened after that point unless it searches. Ask it something well-established from before that cutoff, and it can answer accurately and instantly, drawing on patterns absorbed during training. Ask it about something recent, or something obscure enough that it wasn't well-represented in that training data, and memory alone becomes a much shakier foundation — this is exactly where hallucination risk climbs.
Telling the two modes apart
Verifying a specific citation, step by step
Don't stop at "the source exists" — that's necessary but not sufficient. Locate the actual source independently, confirm the title, author, and date match exactly what was cited, and then check that the source actually supports the specific claim attached to it, not just a related one. A real source cited to back up the wrong point is a subtler, more common failure than a fabricated source outright.
Cross-referencing multiple AI research tools
Different AI tools have different search access, different training data, and different default behaviors around citing sources — which means the same question can get meaningfully different answers from two different tools, and that gap is itself useful information. When two independent tools converge on the same answer through different search paths, that's a real signal of reliability. When they diverge, that's exactly the point worth digging into further, not averaging away.
Getting more out of a research session
Ask explicitly for sources rather than assuming none exist just because they weren't offered upfront. Follow an answer with a specific verification question — "what's your source for that number specifically" — rather than a general one like "are you sure." The general question invites a reflexive "yes, I'm confident"; the specific one forces the system to actually point at something checkable.
Mistakes specific to this layer
Trusting stale memory
- Treating a training-data answer as current when the topic clearly needed a live search
Citation, wrong claim
- A real source that doesn't actually say what it's cited for
False agreement
- Two tools "agreeing" because they both memorized the same outdated claim, not because they verified it independently
The short version
Before trusting any AI-provided fact, ask which mode produced it: recalled from training, or actually searched for. The two deserve different levels of scrutiny, and conflating them is where a surprising amount of AI-assisted research quietly goes wrong. Verify citations against their actual claim, not just their existence, and treat agreement between tools as informative only when you know they reached it independently. For the broader process this technical layer sits inside, see AI Research Workflows.