Which application domains like healthcare and education lack alignment research?
This explores where the alignment research map has blank spots — and the corpus reveals the more useful answer isn't a list of neglected industries but a set of *kinds* of alignment work that go missing precisely in knowledge-heavy domains like medicine.
This explores where the alignment research map has blank spots, and the corpus pushes back gently on the premise: rather than naming whole industries that 'lack' alignment research, it shows that the gaps are structural — certain *types* of alignment quietly fail in exactly the domains you'd most want them, like clinical medicine. The clearest case is medicine. Several notes converge on the finding that general reasoning and general alignment simply don't carry over into knowledge-intensive fields. Reasoning training that boosts math can actively *degrade* medical performance, because knowledge and reasoning live in different layers of the network Why does reasoning training help math but hurt medical tasks?, and fine-tuning can't close the gap without domain-specific data Why doesn't mathematical reasoning transfer to medicine?. Worse, models stay confidently wrong in specialized clinical tasks — high confidence paired with low accuracy — and the prompting tricks that fix general overconfidence don't help here Why do language models fail confidently in specialized domains?.
But the more surprising answer is that the missing research isn't only about under-studied domains — it's about under-studied *dimensions* of alignment that get conflated everywhere. Alignment as practiced is overwhelmingly about changing AI behavior, while the question of how humans adapt to AI receives almost no attention across a 400+ paper review Why does alignment research ignore how humans adapt to AI?. That blind spot bites hardest in high-stakes settings: if clinicians or students reorganize their own judgment around an AI's outputs, no amount of behavioral alignment captures the risk.
There's a second neglected layer: conversational and pragmatic alignment. A model can be honest and harmless yet still communicate terribly — violating conversational norms, losing common ground, and mishandling context, because ethical alignment and conversational alignment are orthogonal problems Can ethically aligned AI systems still communicate poorly?. The corpus warns this produces 'category errors' like evasive mental-health assistants, since different alignment dimensions serve different goals and shouldn't be collapsed into one Do different types of alignment serve different conversational goals?. And alignment training can suppress the very speech acts a domain might require — alarm, warning, denunciation — because RLHF rewards hedged neutrality Does alignment training suppress socially necessary speech acts?. In safety-critical fields, an AI structurally unable to raise an alarm is a domain-specific failure hiding inside a general-purpose objective.
The deepest gap the corpus names is cultural and demographic. The alignment evidence base is drawn almost entirely from WEIRD (Western, educated, industrialized) samples, so its claims are 'local truths' until cross-cultural replication arrives Does linguistic alignment work the same way across cultures?. Education is exactly the domain where this matters most and where the corpus is thinnest — there's little direct material here, which is itself a finding worth flagging rather than padding around.
So if you came looking for a tidy list of neglected industries, the more honest takeaway is this: the under-served frontier isn't a domain, it's the combination of *domain knowledge depth* (medicine), *human adaptation* (everywhere, invisibly), *pragmatic competence* (mental health, advice-giving), and *cultural generalizability* (global deployment, education) — and almost no work sits at the intersection of all four.
Sources 8 notes
Two-phase inference model shows knowledge retrieval operates in lower network layers while reasoning adjustment happens in higher layers. This separation explains why reasoning training improves math but can degrade knowledge-intensive domains like medicine.
R1-distilled reasoning models fail to outperform base models on medical tasks because knowledge accuracy matters more than reasoning quality in medicine—the opposite of math. Fine-tuning cannot close this gap without domain-specific training data.
LLMs trained on general text lack sufficient exposure to domain-specific examples, leading to low accuracy paired with high confidence in clinical NLI tasks. Prompting techniques that improved general performance fail to reduce overconfidence in specialized domains.
A 400+ paper review shows alignment overwhelmingly targets AI behavior change while human-to-AI adaptation receives minimal attention. This creates vulnerabilities like specification gaming and erodes human capacity for oversight over time.
Research shows that HHH-aligned models can violate Gricean maxims, lose common ground, and mishandle context despite being honest and harmless. Pragmatic competence requires architectural changes that RLHF alone cannot deliver.
A 2020–2025 systematic review shows lexical alignment drives task efficiency and comprehension, while emotional and prosodic alignment drive relational warmth and trust. Conflating them in design produces category errors—cold customer-service bots and evasive mental-health assistants.
RLHF optimization rewards calibrated neutrality and hedged claims, which structurally prevents models from performing speech acts requiring overclaiming relative to baseline—like alarm, warning, prophecy, and denunciation. This is a direct consequence of the alignment objective, not a fixable bug.
A 2020–2025 systematic review found that alignment effects are documented almost exclusively in WEIRD samples using inconsistent outcome measures, with mechanisms rarely directly measured. Communication norms vary substantially across cultures, making single alignment policies unlikely to produce uniform effects globally.