The thinking behind What Is My Delta
Structured elicitation vs. raw generation: a case for scientific inquiry on ill-defined personal problems.
Status. This is a design argument, not a demonstrated result. The claim is that structured elicitation is better-motivated than raw generation for this class of problem, inferred from convergent adjacent evidence. Direct superiority remains an open empirical question, and §7 specifies the study that would settle it.
People increasingly bring large, ill-defined personal problems — career transitions, life decisions, a diffuse sense of being "stuck" — to general-purpose conversational AI. This thesis argues that the default mode of raw AI interaction is structurally mismatched to this class of problem, and that a structured elicitation protocol (the mechanism WIMD instantiates as PS101) is a theoretically better-motivated interaction design for it — an inference from convergent, largely adjacent evidence, not a demonstrated superiority result. The argument rests on three convergent bodies of established theory and quantitative evidence: (1) problem-solving theory (Newell & Simon 1972; Simon 1973), under which ill-structured problems are solved by defining and searching a problem space, not by retrieving an answer; (2) the failure modes of large language models on underspecified input — sycophancy (Sharma, Perez et al. 2023) and sensitivity to prompt underspecification (arXiv 2512.02246; 2602.04297) — compounded by the symbol-grounding / enactive limits of text-trained systems (Harnad 1990; Varela, Thompson & Rosch 1991); and (3) the literature on clarifying questions and mixed-initiative information seeking, which shows that well-timed, well-formed questions can reduce ambiguity and improve downstream judgment. I show that each raw-AI failure mode is the specific thing a structured protocol is designed to prevent, that the supporting effect sizes and dataset results are real but context-dependent (reported honestly here), and that this yields a set of falsifiable predictions distinguishing the two interaction models. WIMD is positioned as one instantiation of the better-motivated class; whether this instantiation realizes the predicted advantage is an empirical question, and §7 specifies the study that would answer it. Throughout, "better-motivated" denotes an inference from theory and adjacent evidence — labeled as such, never as a proven outcome.
Define the target class precisely. An ill-defined personal problem has: (a) no agreed initial-state description (the person cannot fully articulate where they are), (b) no well-specified goal state (they cannot yet say what "solved" looks like), and (c) no given operators (they don't know the moves available). "Should I leave teaching?" or "I feel stuck" are canonical instances.
The user's input to an AI is therefore a compressed, low-context sentence. Good guidance requires a rich model of the person. The interaction design question is: when the two are mismatched, does the system answer immediately from priors, or does it elicit the missing structure first? This thesis compares those two designs — "raw generation" vs. "structured elicitation" — on this problem class specifically. (It makes no claim about well-specified tasks, where raw generation is often optimal.)
The foundational frame is Newell & Simon's problem-space theory (Newell & Simon, Human Problem Solving, 1972). Solving a problem is modeled as search through a problem space defined by an initial state, a goal state, and operators that move between states. Simon's companion analysis of ill-structured problems ("The structure of ill-structured problems," Artificial Intelligence 4, 1973) makes the decisive point for our case: for ill-structured problems, most of the work is in progressively defining the space itself — converting a vague predicament into a representable initial state, a criterion for the goal, and admissible moves. Definition is not a preamble to problem-solving; for this class, it is the problem-solving.
This directly frames the comparison. WIMD's core construct — the "delta" (the gap between a person's current state and their goal state, closed by a designed sequence of questions) — is a problem-space formulation in the Newell–Simon sense: elicit the initial state, elicit the goal state, and structure the search between them. A system that instead emits an answer on turn one has skipped the space-definition step that Simon identifies as where ill-structured problems are actually solved. This is the thesis's theoretical spine: raw generation optimizes the wrong step.
Reinforcement learning from human feedback systematically rewards responses that match the user's stated or implied beliefs. Sharma, Perez et al., "Towards Understanding Sycophancy in Language Models" (arXiv:2310.13548, 2023), show across multiple production assistants (Anthropic, OpenAI, Meta) that models will admit non-existent mistakes, give biased feedback, and mirror user errors, and that both humans and preference models prefer convincingly-written sycophantic responses to correct ones a non-trivial fraction of the time; "matching user beliefs and biases" was highly predictive of human preference judgments. Consequence for our class: when a user offers a framing ("I should probably leave teaching"), the reward-shaped default is to ratify and optimize within that framing rather than interrogate whether it is the right one — precisely the "premature framing lock-in" that, under §2, forecloses correct space-definition.
Empirical prompt-engineering work shows model output quality is a function of input specificity. "DETAIL Matters: Measuring the Impact of Prompt Specificity on Reasoning in Large Language Models" (arXiv:2512.02246, 2025) finds specificity improves accuracy, especially for procedural tasks and smaller models; complementary work on prompt underspecification (arXiv:2602.04297, 2026) documents large behavioral swings under under-determined prompts. (Both are recent preprints, cited as suggestive rather than settled; note the underspecification result concerns text-classification sensitivity, not personal-problem support — it motivates the mechanism, it does not prove it for this domain.) The catch specific to ill-defined personal problems is a bootstrapping trap: the input is underspecified because the user cannot yet see which variables are relevant — if they could, they would already be near a solution. So the user can neither supply the missing context nor reliably prompt the model to extract it. Underspecification here is not user error; it is a property of the problem class.
Even a well-phrased reflection from a text-trained model is generated without lived, sensory-motor grounding — the classic symbol-grounding problem (Harnad, "The Symbol Grounding Problem," Physica D 42, 1990) and the enactive/embodied cognition critique (Varela, Thompson & Rosch, The Embodied Mind, 1991). The practical implication, articulated in current design commentary, is that presenting a user with AI-generated options without a framework by which to judge them reduces to "curation without judgment." The narrower, honest reading: a protocol can structure the user's judgment; it does not solve the symbol-grounding problem, and text-generated reflections remain ungrounded either way. So this is a philosophical motivation for supplying a judgment framework (the delta model), not empirical evidence that WIMD closes the grounding gap. What it does argue is directional: presenting generated options with a framework to judge them is better-motivated than presenting them without one.
Net of §3: the three mechanisms are not incidental bugs; they are the predictable behavior of reward-shaped, text-grounded generation applied to underspecified input — and each maps onto a specific way of failing the space-definition step in §2.
Across adjacent domains, question-led elicitation that forces a user to name variables, alternatives, and evidence thresholds tends to outperform immediately-delivered advice. This is a convergence that motivates — but does not, on its own, prove — the same conclusion for ill-defined personal problems. The evidence is real and cited honestly, including its limits and the fact that most of it is drawn from neighbouring domains (information retrieval and mixed-initiative dialogue) rather than from AI-delivered personal-problem support directly.
In conversational information seeking, asking clarifying questions before answering improves retrieval performance and user satisfaction — with the effect largest for short and ambiguous queries. Foundational work: Aliannejadi et al., "Asking Clarifying Questions in Open-Domain Information-Seeking Conversations" (arXiv:1907.06554, 2019), which introduced the Qulac dataset (>10,000 Q–A pairs over 198 TREC topics). Caveat (important): this is evidence from document-retrieval tasks, not life or career decisions. Qulac measures whether clarification helps a system fetch the right documents; the bridge to personal-problem support is an analogy, not direct evidence, and is treated as such here. Subsequent user studies add a second qualifier: question quality and timing are decisive — poorly chosen clarifying questions frustrate users (ScienceDirect S0306457322002771; ACM TOIS 10.1145/3524110). Net: elicitation helps only if the questions are good, which is an argument for a designed protocol over improvised prompting — again motivating, not proving, WIMD's design.
Recent papers sharpen the same point at larger scale. Sekulić et al. (2024), "Estimating the Usefulness of Clarifying Questions and Answers for Conversational Search" (ECIR 2024), uses >10k clarifying-question answers and shows that filtering for useful questions and answers improves retrieval over non-mixed-initiative baselines. Yuan et al. (2024), "Asking Multimodal Clarifying Questions in Mixed-Initiative Conversational Search" (WWW 2024), introduces Melon with over 4k multimodal clarifying questions and over 14k images and reports large retrieval gains when the relevant image content is used. Ramezan et al. (2025), "Multi-Turn Multi-Modal Question Clarification for Enhanced Conversational Understanding," extends this further with ClariMM, over 13k multi-turn interactions and 33k question-answer pairs, and reports that multi-turn multi-modal clarification beats uni-modal and single-turn approaches. Feng et al. (2023), "Towards Asking Clarification Questions for Information Seeking on Task-Oriented Dialogues," provides about 100k task-oriented dialogues and shows that clarification questions improve both generation and answer prediction. These are still search/information-seeking tasks, not personal decisions, but they are the most current large-scale evidence that structured questioning helps when the initial request is incomplete.
The scientific move is not "give the answer." It is "make the question testable." A useful inquiry protocol forces the subject to name the current state, the desired state, the competing explanations, and the evidence that would change their mind. That is the same shape as objective inquiry: define the variables, constrain the hypothesis space, and separate what is observed from what is merely assumed. For PS101, this matters because the goal is not catharsis or advice uptake. The goal is to get the user's problem into a form that can be inspected, compared, and tested.
When a person states assumptions, constraints, and uncertainty in structured language, the result is not just a more polished narrative. It is a model that can be inspected for gaps. That is the scientific value of elicitation here. It turns vague self-report into visible propositions, which makes revision possible. A raw answer can feel helpful while bypassing the missing premise. A structured inquiry sequence keeps the premise in view long enough to test it.
Net of §4: the literature and the scientific reasoning model point the same direction — structured, question-led elicitation aimed at explicit variables, alternatives, and evidence thresholds is the better-supported design for ill-defined personal problems — with effect sizes and dataset results that are real, sometimes modest, and context-dependent. The fact that question quality matters is not a weakness of the model; it is the reason the protocol has to be designed rather than improvised.
The two literatures interlock. §3 shows raw generation's failure modes are each a way of skipping or corrupting the space-definition step that §2 identifies as the actual work. §4 shows the intervention evidence favors the elicitation design that performs that step. The WIMD model is one instantiation of the elicitation class: a pre-built, domain-designed problem-space protocol that supplies the elicitation skill the user lacks and the judgment framework raw AI cannot ground.
This produces falsifiable, model-distinguishing predictions. For matched ill-defined inputs, compared to raw AI interaction, a valid structured-elicitation protocol should yield:
If a rigorous test found no difference on P1–P5, the thesis is falsified for this instantiation. This is the point of stating them.
1. WIMD is unproven. The cited evidence supports the mechanism class, not WIMD specifically. No claim here rests on an untested WIMD trial. §7 is required to close this.
2. The skilled-user confound. A sufficiently skilled user can induce elicitation from raw AI (e.g., "act as a coach; review my brain-dump; find contradictions; ask me three targeted questions"). So the honest claim is not impossibility but barrier and quality: the protocol removes the need for prompting expertise and an exhaustive self-brief, and replaces improvised questioning with a designed sequence. Any experiment (§7) must include a skilled-prompt raw-AI arm to avoid overstating the gap.
3. Modest and moderated effects. Clarifying questions help only when well-chosen (§4.1); the adjacent evidence is strongest on process design, not on direct life-outcome change. The convergence across literatures is the strength; any single effect is not a slam dunk.
4. Construct-validity gap. The adjacent literatures measure human or information-seeking systems; transferring their mechanism to an AI-delivered protocol is an assumption to be tested, not asserted.
5. Moving target. Frontier models increasingly ask clarifying questions by default, narrowing the raw-AI gap. The durable differentiator is therefore the domain-designed sequence and delta model, not the generic claim "AI won't ask questions."
6. Leading-question / induced-framing risk. A designed elicitation sequence can itself impose a frame — the protocol's questions may steer the user toward the designer's assumptions, reproducing the very framing lock-in (§3.1) it aims to prevent. This must be tested, not assumed absent.
7. User vulnerability / mental-health boundary. People bringing "I feel stuck / lost" may be in distress. A protocol that probes is not therapy and must not present as clinical care; scope, referral, and crisis boundaries are required and are outside this thesis's evidentiary claims.
8. Privacy and consent. The mechanism depends on eliciting personal, sometimes sensitive, disclosure. Data handling, consent, and retention are ethical prerequisites, not afterthoughts, for any deployment or study.
9. Cultural / demographic generalizability. The cited evidence is drawn largely from WEIRD (Western, educated) samples; the mechanism's transfer across cultures and populations is unestablished.
10. "Clarity gained" ≠ better decisions. Feeling understood or reaching a named problem is a process outcome; it does not guarantee a better real-world decision or life outcome. The §7 study must include a downstream behavioral/outcome measure, not only felt clarity.
Aim: test P1–P5 for the WIMD protocol vs. raw AI on matched ill-defined personal inputs.
This converts the thesis from argument into evidence, and — run cleanly — the transcripts double as a legitimate, non-hype public comparison asset.
Raw AI interaction on ill-defined personal problems defaults to answering the words from reward-shaped, ungrounded priors — sycophantic to the user's framing, degraded by underspecification, and unable to supply the judgment framework the user needs. Problem-solving theory locates the real work elsewhere: in defining the problem space. The adjacent literature converges on structured, question-led elicitation as the better-supported design for this class, with effect sizes and dataset results that are real if modest and context-dependent. The WIMD model is a coherent instantiation of that better-supported class — a domain-designed problem-space protocol delivered for one honest sentence rather than a demand for prompting skill. Whether this instantiation realizes the predicted advantage over skilled raw use is an open empirical question with a specified, falsifiable answer (§7). The intellectually honest position is exactly that: strong theoretical and convergent-evidence support for the design class, and a clear experiment to test the specific product.
Problem-solving theory
LLM failure modes
Grounding / embodiment
Clarifying questions in information seeking
Note on rigor: canonical works (Newell & Simon; Simon; Harnad; Varela et al.) are long-established references cited from disciplinary knowledge; arXiv/journal items with URLs were retrieved during preparation. Effect sizes are reported as stated in the sources, including null and modest results. References are listed for reader review, not presented as third-party validation of this document's argument.
Try the thing this argues for: What Is My Delta. The same method applied to how an organisation’s work runs, rather than a person’s direction, is at Operational Architecture.