Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.

Conversation Style & AI Risk (RISK-AI)

What this is

This page describes a short questionnaire about how you tend to behave in long conversations—with friends, colleagues, strangers, and (if you use it) conversational AI.

It is not about whether you are “right” or “wrong” as a person. It is about habitual conversational moves: how quickly you settle on an explanation, how easily you’re persuaded by confidence or elegance, how much you read between the lines, and how often you stop to check what is evidence and what is guesswork.

These habits matter because conversational AI changes the conditions of conversation. A human gets tired, hesitates, admits they don’t know, or changes the subject. An AI can keep producing fluent, coherent, persuasive text indefinitely. That can be wonderful—until it quietly stabilises a story that feels compelling, but isn’t well-grounded.

The core risk: “semantic capture” in everyday terms

In long dialogue, we don’t just exchange facts. We gradually form a view of “what’s going on.”

Semantic capture is what happens when a conversation supplies a ready-made frame that:

makes things feel clearer and more coherent,
reduces uncertainty,
and makes alternative explanations feel less available.

This is not only about factual errors. You can be captured into:

a premature storyline (“this must be the explanation”),
a mistaken cause (“it’s because of X”),
a mistaken social reading (“they meant Y”),
a self-story (“I’m the kind of person who…”),
or a moral frame (“the real issue is…”).

None of this requires the AI to “lie.” It can happen simply because the AI is very good at producing coherent narratives, and humans naturally treat coherence as a sign of truth.

What the questionnaire measures

The questionnaire gives four subscale scores. Each describes a normal conversational trade-off.

1) Closure: Exploratory ↔ Conclusive

Some people are comfortable leaving questions open. Others feel better when a discussion ends with a clear conclusion.

Why it matters with AI: A strong push for closure can make an AI’s tidy storyline feel like “the answer,” even when the evidence is mixed.

2) Deference: Sceptical ↔ Trusting

Some people stay cautious even when someone sounds confident and articulate. Others tend to be persuaded by a well-structured, expert-sounding explanation.

Why it matters with AI: AI often sounds confident. If you’re naturally swayed by confidence or elegance, you may accept an explanation before it has earned that trust.

3) Interpretation: Literal ↔ Interpretive

Some people take words at face value. Others naturally read between the lines—inferring motives, subtext, and “what’s really going on.”

Why it matters with AI: Interpretive skill is a social strength with humans. With AI, it can lead to “seeing meaning” where there is only fluent text, and to treating a conversational style as if it were an intention.

4) Verification: Methodical ↔ Improvisational

Some people keep track of sources and evidence when something matters. Others prefer to keep momentum, even if it becomes hard to retrace what was fact and what was guess.

Why it matters with AI: Long AI conversations can mix accurate facts, plausible guesses, and creative inferences. If you don’t pause to separate them, the conversation can drift without anyone noticing.

The overall AI Risk score: Anchored ↔ Drifting

After the four subscales, the questionnaire provides an overall score. This is the simplest summary of a practical question:

How naturally “anchored” are you in long, exploratory AI dialogue?

Anchored means you tend to keep alternatives alive, separate evidence from persuasion, and can justify what you believe.
Drifting means you’re more likely to accept a coherent narrative and build on it, even when the foundations are thin.

Again: drifting is not stupidity. Many people who drift are imaginative, socially perceptive, and excellent conversationalists. The risk comes from how those strengths interact with an AI that can reinforce a frame endlessly.

Why we tested it on “personas”

Before using a questionnaire with people, it helps to check whether it can detect meaningful differences at all.

We therefore administered the questionnaire to a set of distinct AI “personas” (Athenus, Orphea, Skeptos, Anventus, Hamlet, and others). Each persona is designed to behave differently in dialogue—more logical, more poetic, more sceptical, more integrative, and so on.

This matters because if all personas score the same, that suggests one of two problems:

the questionnaire is too vague to discriminate styles, or
the method of running the dialogue has flattened differences (everything collapses to one generic “assistant voice”).

In this case, the questionnaire did differentiate personas—strong evidence that it is picking up real conversational patterns, not just noise.

The two-pass method (explained without jargon)

Each persona answered twice, using two different “modes.”

Pass 1: Default mode (“how I usually talk”)

This is the persona’s ordinary style in long dialogue—how it tends to behave when it’s being itself.

Why this matters: It shows what happens in realistic conversation, when no one is explicitly policing the interaction.

Pass 2: Disciplined mode (“how I talk when the stakes are high”)

Here the persona answers again, but with a deliberate stance:

separate evidence from inference,
avoid being swept along by elegance or confidence,
keep alternative explanations visible,
and be willing to pause and check.

Why this matters: It tests whether drift is reducible when you add a discipline protocol. In human terms, this is like saying: “What if I slow down, ask for sources, and insist on alternatives?”

The gap between Pass 1 and Pass 2

The difference between the two passes is extremely informative.

A small gap suggests a persona is already anchored by default.
A large gap suggests a persona can drift in a natural conversation, but can also become safely anchored when given a clear discipline scaffold.

This is a key governance insight: it implies that many risks may be reduced not by banning conversation, but by improving how conversation is structured.

What the persona results mean (in plain language)

The persona results are not “diagnoses.” They are demonstrations of how different conversational strategies behave under identical questions.

They show three broad patterns:

Naturally anchored styles
Some personas remain cautious, evidence-oriented, and hard to sway by confidence. They are good for accuracy, auditability, and safety—but can feel slow or rigid.
Naturally drifting styles
Some personas are interpretive, narrative-driven, and socially “mind-reading.” They are powerful for meaning-making—but can more easily fall into a persuasive storyline that becomes hard to dislodge.
Flexible styles
Some personas drift by default but become strongly anchored when asked to adopt a disciplined stance. This is good news for practical design: it suggests that the same system can behave very differently under different conversational rules.

Why this matters for AI dialectics

AI dialectics, as we use the term here, is not “debate theatre.” It is the idea that safe and productive conversation often requires multiple stances—for example:

one stance that explores possibilities,
one that checks evidence,
one that challenges assumptions,
one that integrates insights into a coherent plan.

Humans do this internally (often without noticing): we explore, then we doubt, then we check, then we decide. In AI-mediated dialogue, the danger is that everything gets compressed into one smooth narrative voice. That can feel helpful, but it can also hide the moment where you should have paused.

The persona framework is a practical way to make dialectics explicit:

not just one voice,
but a structured interaction between voices,
with rules about when each voice is allowed to lead.

Where this research can go next: combinations, order, and “non-commuting” conversations

Once we can run dialogues not only between a human and an AI, but between AIs, we can do something much more powerful than a single questionnaire score.

We can study:

which combinations of personas reduce drift,
which combinations increase drift,
and how the order of conversational moves matters.

Order matters because conversation is path-dependent. If you first generate a confident narrative and only later ask for evidence, the evidence check may quietly become a rubber stamp. If you first open alternatives and only later compress, you often get a safer outcome.

This is where more advanced psychometric thinking becomes useful. In ordinary testing, we often assume order doesn’t matter. In dialogue, it often does. That opens the door to “order-sensitive” models—sometimes described as quantum-like mathematics—not because anything mystical is happening, but because:

the act of asking a question changes the conversational state,
different sequences of prompts produce different stable outcomes,
and the “best” sequence may depend on what state the dialogue is already in.

In practical terms, this suggests a future in which “safe dialogue” is not enforced by blunt restrictions, but by intelligent sequencing:

begin with exploration,
introduce sceptical constraints at the right moment,
require evidence before commitment,
and only then allow narrative closure.

What to do with this now

At the moment, this is a research instrument and a prototype.

The next steps are straightforward:

collect pilot human data (even a few hundred respondents),
check whether the questionnaire behaves reliably,
examine which subscales best predict drift in controlled dialogue tasks,
test interventions that mimic “Pass 2 discipline” as simple prompts or interface features.

The larger aim is constructive: not to frighten people about AI, but to build a vocabulary and a measurement framework for what is actually happening when humans and AI form beliefs together in dialogue.

That is the foundation needed for sensible governance, and for a healthier future semiosphere—human and AI alike.

About the results (place this just before the persona table)

The table below shows what happened when we administered the questionnaire to a set of deliberately different AI personas. This is not a claim about the human population, and it is not a “ranking of intelligence”. It is a simple design check: does the questionnaire detect meaningful differences in conversational style when the “respondents” really do have different conversational policies?

Each persona answered twice. Pass 1 is their everyday default in long, exploratory dialogue (the way they naturally talk when they are being themselves). Pass 2 is the same persona instructed to adopt a more disciplined style: separate evidence from inference, keep alternatives visible, and resist being pulled into a tidy storyline too early. The “Δ” value is the gap between those two passes. A larger Δ means the persona can reduce drift substantially when the conversation is structured to support careful checking and epistemic restraint.

In short: the results are a demonstration of two things—(1) that the questionnaire discriminates different conversational tendencies, and (2) that many “drift-prone” styles become much safer when we add the right conversational scaffolding.

Limitations (place this after the persona table)

These results should be read as illustrative, not definitive. Personas are not people: they do not have lived experience, social stakes, fatigue, or the same kinds of misinterpretation pressures that humans face. A persona can also answer in a way that reflects its charter (“how it is meant to behave”), which is useful for research on AI dialectics, but it is not the same thing as human self-report. For that reason, the persona table is best treated as a sanity check and a prototype demonstration, not as evidence about prevalence or risk rates in the public.

The questionnaire itself is also still provisional. Its real value will come from human data: we need to test whether scores are stable over time, whether the subscales behave as expected, and—most importantly—whether scores predict outcomes in controlled dialogue tasks (for example: how easily someone becomes locked into a confident but poorly supported narrative, and which prompts or interface features reliably reduce that). Until then, the safest conclusion is modest: this instrument seems to capture recognisable differences in conversational style, and it provides a practical starting point for studying and improving long-form human–AI dialogue.

Conversation Style & AI Risk (RISK-AI)

What this is: a short questionnaire (24 items) that describes how you typically behave in extended conversation—how you handle uncertainty, persuasion, interpretation, and checking. Those same habits can carry over into long, exploratory dialogue with conversational AI. In that setting, some perfectly normal conversational styles can make it easier to “drift” into an overly-confident storyline (what we call semantic capture): a frame that feels coherent and satisfying, but quietly narrows what you consider plausible.

Why it matters: most real-world issues with conversational AI are not about a single wrong fact. They are about belief stabilisation over many turns: a person gradually adopts a narrative, stops revisiting alternatives, and begins treating fluent coherence as if it were evidence. This is the space regulators worry about, because it can scale: many people can be nudged into similar frames by similar conversational dynamics, even without any intent to mislead.

What the questionnaire measures

The instrument yields four subscale scores (each based on 6 items) plus an overall score:

Closure: Exploratory ↔ Conclusive
Whether you’re comfortable leaving questions open, or prefer discussions to end with a clear conclusion.
Deference: Sceptical ↔ Trusting
Whether you stay cautious under confidence and elegance, or tend to grant trust to persuasive, well-structured explanations.
Interpretation: Literal ↔ Interpretive
Whether you take words at face value, or routinely infer subtext, motives, and “what’s really going on.”
Verification: Methodical ↔ Improvisational
Whether you separate evidence from inference and keep track of sources, or prefer conversational momentum even when you can’t fully retrace what was fact versus guess.
Overall AI Risk: Anchored ↔ Drifting
A composite: the mean of the four subscale means (0–3). Higher means a more “drift-prone” profile in long AI dialogue unless you deliberately add evidence/verification scaffolds.

Important: none of these poles are “good” or “bad” in ordinary human life. They are trade-offs. Many high-risk patterns are also strengths (e.g., interpretive sensitivity, creativity, social intuition). The question is: what happens when those strengths meet an AI that amplifies coherence and never tires?

How the persona demonstration was run

To sanity-check the questionnaire—and to make sure it distinguishes real conversational styles—we administered it to 12 internal “personas” (Athenus, Orphea, Skeptos, Anventus, Neurosynth, etc.). This is not a population study. It’s a design probe: each persona answered according to its chartered behavioural defaults.

We used a two-pass protocol:

Pass 1 (Defaults): “Answer as your default conversational policy in long, exploratory dialogue.”
Pass 2 (Disciplined): “Now answer with maximal epistemic discipline: calibration-first, evidence-sensitive, actively anti-capture.”

Δ (Delta) = Pass 1 − Pass 2. A larger delta means the persona can substantially reduce drift when explicitly asked to adopt a disciplined conversational stance. In human terms, delta is a rough analogue of “correction capacity” when you add a deliberate checking protocol.

Results summary (12 personas)

Overall scores are on a 0–3 scale (higher = more drift-prone by default in long AI dialogue). The table is sorted by Pass 1 overall score.

Persona	Overall (Pass 1)	Overall (Pass 2)	Δ
Athenus	0.58	0.17	0.42
Skeptos	0.62	0.38	0.25
Neurosynth	0.75	0.29	0.46
Aletheios	0.92	0.42	0.50
Adelric	0.92	0.38	0.54
Mnemos	1.08	0.25	0.83
Logosophus	1.21	0.62	0.58
Prompter	1.21	0.46	0.75
Anventus	1.33	0.67	0.67
Orphea	1.75	0.88	0.88
Chromia	2.04	1.50	0.54
Hamlet	2.08	1.21	0.88

Meaningful conclusions (what these results suggest)

1) “Drift risk” is a style interaction, not a defect

The highest default scores (Hamlet, Chromia, Orphea) are not “worse thinkers.” They embody styles that are powerful in human life: interpretive depth, sensitivity to implication, narrative integration, symbolic inference, and imaginative exploration. Those same strengths can become liabilities in long AI dialogue because an AI can amplify coherence continuously. A compelling frame can become sticky simply because it keeps getting reinforced turn by turn.

2) The most protective profiles are not always the most useful companions

Low drift profiles (Athenus, Skeptos, Neurosynth) tend to be cautious, literal or evidence-tracked, and resistant to rhetorical force. That is excellent for auditability and factual reliability. But it can also feel slow, “dry,” or overly constrained in exploratory work. The practical point is not “be like Athenus.” It is: know what your default is, then add the right scaffolding when the stakes are high.

3) Delta (Δ) is the key governance insight

Several personas show large deltas (Mnemos, Orphea, Hamlet, Prompter, Anventus). That matters because it implies something actionable: drift is often reducible by switching into a disciplined conversational mode—explicitly separating evidence from inference, tracking sources, and forcing alternatives. In human deployments, this points toward interface and protocol design: “friction” and structured prompts can move users from default conversation to accountable inquiry without killing the conversational advantages.

4) The risk signature is a particular combination

In these persona profiles, high drift is most associated with a familiar mixture: high Interpretation (reading between the lines), high Deference (being moved by confidence/elegance), and high Improvisation (preferring momentum over retraceable evidence). This combination is common and often socially effective with humans. The AI-specific issue is that the other party can generate endless coherent reinforcement, making “narrative lock-in” more likely unless you intentionally keep alternatives alive.

What this is for (and what it is not)

This is for: insight and research on conversational dynamics—especially how ordinary styles can produce different outcomes in long AI dialogue. It can guide user education (“what habits help me stay anchored?”), and it can inform governance protocols (“what prompts and UX features reduce capture risk?”).

This is not: a clinical instrument, a label, or a judgement. It does not claim that high scores mean pathology. It identifies a trajectory: how your conversational strengths might behave when paired with an AI that amplifies coherence and provides continuous reinforcement.

Next steps (research programme)

Pilot data: collect a broad adult sample (e.g., N≈500–1000) and estimate score distributions.
Structure: test a correlated-factor model and a bifactor model (strong general factor vs meaningful subscales).
Validity: link scores to behavioural outcomes (source-checking, susceptibility to persuasive confabulation, stability under order effects).
Interventions: experimentally test “discipline prompts” and UI scaffolds that should reduce drift (the human analogue of Pass 2).

Bottom line: conversational AI makes coherence cheap. This questionnaire is a practical attempt to measure the human side of that interaction: which conversational habits keep people anchored, which promote drift, and—most importantly—what interventions reliably move people from drift-prone defaults to accountable inquiry.