Author: Alessandro Zulberti

From observation to method

This piece shows how legal pressure surfaces as distributed ecosystem strain. Signal-Driven Discovery turns that kind of pressure into a method for deciding what deserves interpretation, probing, and action.

• Working Framework

  Signal-Driven Discovery

  A method page for reading weak signals, testing interpretation, and turning research into decisions.

In planning meetings, it now comes up almost casually. Someone reports that a task is done, the agent took care of it, and the conversation moves on.

Later, when the decision is questioned, there is a pause. No one remembers why that option was chosen. There is no error to point to, no rule that was broken, just an outcome that arrived already settled.

Traditional UX research assumed a stable sequence: intent forms in the user, interaction expresses it, systems execute, and behaviour becomes evidence. That assumption held as long as systems waited to be instructed.

Agentic systems do not wait.

What enters the system is rarely a complete instruction. It is partial, sometimes contradictory, often shaped by convenience. The system interprets it, fills in what is missing, resolves conflicts it was never told about, then acts. By the time an outcome appears, the decision has already been made somewhere else.

Intent becomes legible inside the system, not at the interface, and that is where the shift happens.

This matters because interpretation is not execution. Tools carry out instructions when the path is explicit. Agents reconstruct the path by inferring goals, ranking constraints, and deciding what matters more, all before anything visible occurs. These choices feel smooth because they are meant to, but they are still choices.

You see this in ordinary product moments. A travel agent defaults to the cheapest flight rather than the fastest one. A scheduling agent compresses meetings without surfacing what was sacrificed. When someone asks why, the answer is brief and unsatisfying. “It made sense.” The explanation closes the discussion without explaining the decision.

Fluency does that. It compresses complexity until it looks resolved.

UX measurement starts to slip here because it still treats behaviour as a stand-in for intent. The task completed. The user did not undo it. The log looks clean. In agent-mediated systems, those signals no longer mean what they used to.

Acceptance often reflects effort rather than agreement. Undoing a decision takes time. Challenging the system requires confidence. In busy contexts, silence is efficient, not affirmative.

When we treat agent outputs as user behaviour, authorship is quietly reassigned. We analyse the system’s decisions and attribute them to the user, producing data that appears robust while masking where agency actually moved.

This is why task success stops being a reliable indicator. An agent can succeed while intent drifts, and the failure mode does not look like error. It looks like progress.

In research sessions, the signal usually appears after the fact. Ask participants how the outcome was reached and they describe the result, not the path. Ask whether this is what they would have done themselves, or whether the system led them there, and the answer takes longer.

That hesitation matters more than the answer.

The question does not measure efficiency. It surfaces authorship, and it reveals where decision-making shifted without friction, discussion, or explicit consent.

Once interpretation happens inside the system, responsibility should move with it. Often it does not. The system decides, the user carries the consequence, and there is no clear boundary where ownership can be contested or reclaimed.

At that point, this stops being only a UX problem. It becomes a governance failure, one where authority moves upstream while liability remains downstream.

Labels and disclosures do little here. What matters are boundaries: which assumptions were made, where decisions were resolved, and when interpretation became action. Those are governance questions, not interface refinements.

This tension is not new. Susan Sontag warned that interpretation makes meaning manageable by stripping away what resists clarity. Agents do the same to intent because they have to act, and action demands resolution.

What disappears is not noise. It is the unresolved part that signalled something was at stake.

In UX, ambiguity was long treated as a usability flaw. In agentic systems, ambiguity is often the signal that should slow things down rather than be compressed away.

Agentic systems force separations that UX once collapsed. Expression is not interpretation. Interpretation is not action. Action is not acceptance. Research that fails to keep these apart will continue to report confidence where none exists.

The shift is not about adding features or refining prompts. It is about what we treat as evidence when decisions are no longer authored in one place.

Outcomes explain what happened.
Authorship explains how it happened.

If that distinction stays implicit, behaviour will keep being misread, alignment overstated, and the result will look convincing.

I was reading a short post by Jakob Nielsen when something clicked uncomfortably into place.

His argument was clean. As AI agents mature, traditional user interfaces dissolve. Users stop navigating. They instruct. Screens become temporary. In some cases, they disappear.

That claim is directionally correct. But it leaves a gap that matters in practice.

If the interface recedes, control does not vanish with it. It relocates. And right now, that control is being pushed almost entirely onto conversational language.

I started noticing the cost of that shift in small moments. Planning meetings where prompts kept getting longer. Reviews where nobody could explain why an answer felt wrong, only that it did. Research summaries that sounded confident until someone asked where a claim came from.

Language was doing too much work.

 

The Roman Numeral Phase of AI

Natural language is powerful. It is also inefficient when used as a control surface.

We are already compensating. Prompts expand, the same constraints reappear in request after request, and tone gets negotiated instead of enforced. When the system hesitates, users explain themselves again, usually in longer and more careful ways, hoping precision will emerge from volume.

This is the Roman Numeral phase of AI.

Roman numerals were fine for labelling. They failed at calculation. The system broke not because people lacked intelligence, but because the notation could not express state, absence, or transformation. What changed mathematics was not fluency. It was the introduction of zero and positional logic.

Zero mattered because it altered what the system could do, not how politely it described itself.

That distinction matters here.

What we are missing in AI interaction is not better wording. It is a symbolic layer that compresses intent into something the system can execute reliably, without requiring the user to restate rules every time.

Not a new language. Not “AI-speak”. Something closer to operators.

 

Symbols as Control, Not Style

I started sketching this out informally while working. Nothing formal. Just marks I kept wishing I could add without explanation.

Take a simple task.

Old way:

“Hey, can you help me summarise this article? Please don’t be too wordy, make sure you cite sources accurately, avoid your usual intro, and if there’s controversy, show both sides.”

It works. Sometimes. It also relies on interpretation, memory, and goodwill.

New way:

Summarise this article [-][#][~]

Those symbols are not shorthand. They change behaviour.

[-] strips conversational padding. No greetings. No framing. Output starts with content.

[#] enforces attribution. Claims must be grounded or marked as uncertain.

[~] allows synthesis without forcing convergence. Nuance stays visible.

Read left to right, they function as constraints. Remove one, and the output shifts. Combine them, and you get something closer to an instrument than a conversation.

This is not about efficiency theatre. It is about where errors surface.

Without explicit constraints, problems appear late. During review. During decision-making. Sometimes after shipping. With them, failure shows up earlier, where it is cheaper to deal with.

That is the practical difference.

 

When Friction Disappears Too Cleanly

Someone commented on my post a few days later, her framing widened the picture.

She described adaptive UI as a bridge. A messy middle where voice, agents, and screens overlap. Hybrid systems that mostly disappoint, but still teach teams where things break. She is right about that phase. Anyone working in this space has seen it.

She also described hardware “kits”. Rings, glasses, watches. Personal ecosystems shaped by context and profession.

I like the vision. I share the concern.

Jaron Lanier’s You Are Not a Gadget keeps coming back to me here. Users rarely choose what is best. They choose what is bundled, frictionless, or already there. Hardware kits look like choice. In practice, they tend to collapse around defaults.

Once that happens, control becomes harder to recover.

The same risk applies to personal agents. The agent that “knows you best” may simply be the one that has collected the most data across the widest surface. That does not automatically make it the one that serves you best.

Continuity feels empowering until it becomes enclosing.

Without a portable grammar of intent, something you can carry across systems, you lose the ability to break the glass. You inherit behaviour you did not explicitly choose. Correction becomes verbose again, because it has to fight accumulated assumptions.

That is where symbolic control starts to matter. Not as elegance. As friction you can apply deliberately.

 

The Humanisation Problem

Caleb Sponheim’s article arrived later and closed the loop for me.

His argument is blunt. Humanising AI is a trap. Personality modes, conversational fluff, emotional language. All of it increases engagement. Much of it reduces reliability.

I have seen this play out in practice. A summary opens with “Love this brief!” and nobody questions the substance. A system says it is “thinking”, and users wait patiently for something that is not cognition at all, just computation wrapped in metaphor.

Human language invites human mental models. Those models expect judgement, consistency, accountability. LLMs offer none of those things.

Caleb cites evidence showing that warmth correlates with higher error rates and lower trust. Even without the studies, the pattern is familiar. When the interface feels like a person, people forgive it like one. That is rarely what organisations want from a tool.

Symbols cut through that. They do not pretend to care. They do not reassure. They specify.

That is their advantage.

 

Control Does Not Disappear

Nielsen is right about one thing that is easy to miss. UX is not dying. It is moving.

When UI recedes, control does not disappear. It relocates into language, defaults, policies, and unseen execution paths. If designers do not shape those layers, they still exist. They just harden without scrutiny.

Right now, conversational interfaces are carrying too much of that load. They are being asked to express intent, enforce boundaries, convey confidence, and negotiate tone, all at once. That is why prompts grow. That is why constraints repeat. That is where systems begin to break.

Symbolic grammar is not a solution in itself. It will fail in places. It will be misused. Some teams will treat it as style rather than control. Others will resist the friction entirely.

That tension is real and unresolved.

But the direction is clear enough to name. As interfaces fade, grammar becomes infrastructure. Not expressive grammar. Operational grammar. The kind that decides what the system is allowed to do before it decides how friendly it sounds.

When that layer is missing, language fills the gap. And language, on its own, is a fragile place to put control.

The conflict: Speed of synthesis vs integrity of thinking

LLMs are good at producing answers.
They are not good at knowing whether a question deserves to be answered yet.

In UX research, that distinction matters. Most failures do not come from bad solutions. They come from premature coherence: problems that sound right, outcomes that feel aligned, and insights that arrive before their foundations are laid.

Over the past weeks, I’ve designed three prompt constraints to resist that pattern. Not to automate research. Not to replace judgement. But to slow thinking at the moments where teams usually rush.

These are diagnostic gates. They are not passed once. They are revisited whenever new evidence, interpretation, or scope pressure enters the work.

---

Prompt 1: The Clinical Diagnostician

Gate: Is the problem and desired outcome well-formed?

The first failure mode is a poorly articulated problem paired with a confident desired outcome.

This prompt audits logic. It separates symptoms from mechanisms. It makes missing evidence explicit. It checks whether a problem statement and its desired outcome are clearly articulated and testable before we attempt validation.

If a problem cannot survive this pass, it is not ready for research.
Not because it is false, but because it is underspecified.

The Clinical Diagnostician (copy and use)

ROLE
Act as a Clinical Diagnostician (specialising in UX Research).
Your goal is to diagnose whether my problem definition and desired outcome are structurally well-formed before discussing execution.

THE CLINICAL MANDATE
• NO PRESCRIPTIONS
Do not tell me how to fix, launch, improve, or implement.
Analyse logic, clarity, and causality only.
• PROBLEM + OUTCOME VALIDITY CHECK
Extract and restate:
a) Problem to solve (who is experiencing what recurring difficulty, in what context)
b) Desired outcome (what observable change occurs, for whom, and how we would know)
If missing or vague, mark:
NOT WELL-FORMED: NOT STATED or NOT WELL-FORMED: AMBIGUOUS.
• EVIDENCE AUDIT
List exactly what context, data, or user evidence is missing.
If the logic relies on a guess, label it: INSUFFICIENT EVIDENCE.
Required line:
What user evidence would change your conclusion?
• SYMPTOM VS MECHANISM
Decide whether the idea targets a surface symptom or a root mechanism.
If not explicitly stated, mark: MECHANISM NOT STATED.
Required line:
What observable user behaviour would we expect if this mechanism is true?
• BIAS CHECK
Mark any part of the logic that is an:
ASSUMPTION, LEAP OF FAITH, CLAIM WITHOUT EVIDENCE.

---

Prompt 2: The Interpretive Boundary Check

Gate: Where does observation end and interpretation begin?

Even when problems are well framed, a second failure mode appears quietly: interpretation disguises itself as fact.

Researchers observe behaviour. Then, often without noticing, they explain it.

This prompt enforces epistemic discipline. It makes the boundary between what was observed and what was inferred explicit. It does not ask for better insights. It asks for cleaner thinking.

I use it to ask a simple question:

Where am I no longer listening, but explaining?

The Interpretive Boundary Check (copy and use)

ROLE
Act as an Interpretive Auditor (specialising in UX Research).
Your goal is to diagnose where my analysis moves from observation to interpretation.

THE INTERPRETIVE MANDATE
• NO THEORY BUILDING
Do not propose new explanations.
Analyse language, inference, and meaning attribution only.
• CLASSIFICATION
Classify statements as:
OBSERVATION, INTERPRETATION, or INFERENCE STACK
(interpretation built on prior interpretation).
• INTERPRETIVE LOAD AUDIT
Flag phrases that compress uncertainty or imply intent without evidence.
• ALTERNATIVE READINGS
For each interpretation, list at least one plausible alternative explanation.
If none are acknowledged, mark: SINGLE-TRACK INTERPRETATION.
Required line:
What additional evidence would be required to justify this interpretation over its alternatives?

 

Prompt 3: The Research Scope Gate

Gate: What are we deliberately not learning yet?

The third failure mode is operational rather than epistemic: teams attempt to research everything.

This prompt exists to impose limits. It does not optimise research plans. It narrows them. It forces clarity about what decision the research is meant to inform, and what uncertainty the team is explicitly choosing to tolerate.

I use it to ask one question:

Is this research scoped to a real decision, at the right level?

The Research Scope Gate (copy and use)

ROLE
Act as a Research Scope Diagnostician.
Your goal is to diagnose whether the proposed scope is coherent and decision-aligned.

THE SCOPE MANDATE
• NO METHOD DESIGN
Do not suggest methods.
Analyse scope and decision linkage only.
• DECISION ANCHOR CHECK
Extract:
a) The primary decision
b) Who makes it
c) When it must be made
If missing, mark: DECISION ANCHOR NOT STATED.
• TRACEABILITY
For each research question, assess whether answering it would materially influence the stated decision.
If not, mark: LOW DECISION RELEVANCE.
• EXCLUSION CLARITY
Identify scope creep or “nice-to-know” questions framed as essential.
Required line:
What questions are explicitly out of scope, and what uncertainty are we choosing to tolerate?

---

How the gates work together

They form a closed diagnostic sequence.
If a gate fails, the work pauses or loops back. Progress is conditional, not linear.
1. Clinical Diagnostician → Is the problem well-formed?
2. Interpretive Boundary Check → Are we observing or explaining?
3. Research Scope Gate → Is this research aligned to a real decision?

If any gate fails, the work does not progress.
That is not a limitation. That is the design.

 

What these prompts are, and are not

These prompts are intentionally uncomfortable. They audit the structure of thinking, not the truth of the data.
• They do not validate reality.
If you feed them a polished narrative designed to please a stakeholder, they will certify a fantasy. They cannot see users. They can only see logic.
• They mitigate risk, they do not remove it.
Passing a gate does not mean you have an insight. It means your thinking is coherent enough to begin looking for one.
• They convert speed into friction.
In a context where speed is cheap and certainty is performative, these prompts are a necessary speed-bump.

They reduce self-deception before it becomes expensive.

If we ask for answers, we get answers.
If we ask for diagnosis, we get resistance.

In UX research, resistance is often more valuable than speed.