My son’s been refusing to go to school for three days. He’s two and a half. I don’t know enough developmental psychology to know whether to worry.
So I ran a knowledge mine — a structured extraction of what Claude knows about toddler school refusal. In two minutes, I had: five distinct mechanisms that look identical from outside but need different responses, a drop-off protocol backed by attachment theory, a table distinguishing normal phases from warning signs, and a developmental timeline for separation comfort by age.
Could the model be hallucinating? Sure. Some of the specific thresholds (“settles within 15 minutes”) are approximate. The age ranges are synthesised across studies, not cited from one definitive source.
But here’s the thing: the structure is reliable. The distinction between autonomy assertion and separation anxiety. The operant conditioning argument against returning after you’ve left. The overjustification effect on bribing. These aren’t invented — they’re well-replicated developmental psychology, compressed into the model’s weights from thousands of papers and textbooks.
The question isn’t whether the model “understands” developmental psychology. It’s whether the output is consistently good enough to act on. And for parenting decisions — not medical prescriptions, not journal submissions, but “should I linger at drop-off tomorrow?” — it is.
Ten blue links would have given me one article’s perspective, probably from a parenting blog with an affiliate link to a children’s therapist. The mine gave me a synthesised decision framework across the literature in one pass.
You couldn’t have done this on GPT-2. Or GPT-3. The shift happened somewhere around the point where models crossed from producing plausible prose to producing structured frameworks — where they could hold five mechanisms in working memory, distinguish them on signal rather than vibes, and flag their own uncertainty at the right points.
A calculator doesn’t understand arithmetic. I still use it over mental math. The interesting question isn’t “does it understand?” It’s “when did the output get good enough that the understanding question stopped mattering?”
For knowledge extraction from well-studied domains, we’re past that threshold. Not for everything — volatile facts, niche subfields, anything post-training-cutoff still needs verification. But for stable cognitive operations across established fields? The parrot is good enough. And getting better.
The trick is how you use it. Most people type a question and get an answer — the search engine mental model. The real value is treating the model as a subject matter expert you’re debriefing: probe, push past the first answer, look for the structure underneath. “What drives school refusal at this age?” gets you a paragraph about separation anxiety. “What are the distinct mechanisms, and how do their signals differ?” gets you a decision framework. The first answer is always too general. The second and third layers are where the structure lives.
This has implications beyond parenting. A manager handling their first performance issue. An engineer evaluating a vendor. Anyone negotiating anything. The common shape: you need a decision framework in a domain you visit infrequently. Books take hours. Blog posts give you one perspective. The debrief gives you the compressed structure of the field — and for most decisions, an imperfect framework beats no framework at all.