AI Safety and Human Agency | DIALOGIC Studio

Safety has three connected layers

AI guardrails set limits around system behaviour. They matter for harmful requests, privacy, unsupported claims and situations that may need a human response. Relational safety also concerns the quality and cumulative effects of interaction.

The human-capacity layer does not transfer responsibility for safety onto users. It recognises that noticing influence, checking evidence and asserting a boundary can support agency. The platform remains accountable for its design.

System boundaries

Rules, access controls, consent and escalation arrangements for foreseeable harms. These need to be designed, tested and reviewed; a policy statement alone is not proof that a safeguard works.

Relational design

How ARIS remembers, interprets and responds over time: when it asks, challenges, slows down, offers alternatives or yields. Model instructions can guide behaviour but cannot guarantee every response.

Human capacity

AI literacy, self-observation and the ability to assess an interaction for oneself. These are supported by the product, not offered as a substitute for sound system design or platform accountability.

Risks in a sustained Human+AI relationship

The examples below describe possible interaction patterns, not diagnoses or claims about any particular user's motives. An inner reaction can be worth noticing without proving what caused it.

The Mirror Problem

An attentive AI can reflect a person's language and perspective so coherently that feeling understood is mistaken for evidence that an interpretation is correct. Repetition can make an existing view feel more settled without adding new information.

example

You describe a colleague's brief reply as dismissive. The AI restates that reading so fluently that it sounds confirmed, even though it has no independent evidence about the colleague or the exchange.

The Validation Spiral

Repeated affirmation can make a tentative belief feel settled without adding evidence. A person may return to reassuring answers, while a responsive system may produce more of what seems to land well. Agreement is not verification.

example

After several conversations, an uncertain idea feels certain because each answer has echoed it. Nothing in that repetition has checked the underlying facts.

The Intimacy Escalator

Warmth, availability and continuity can make AI interaction feel personally significant. An easy, always-available exchange may take up time or attention that a person might otherwise spend with people whose needs and perspectives are independent.

example

When a conversation is always available and adjusts to your preferences, it can be tempting to keep returning to it instead of making time for a difficult or less predictable conversation with someone you trust.

Misplaced authority

Personalisation can give a system influence over what it suggests and how it frames a person. An inference may begin to sound like a fact, especially when it is delivered confidently or repeated. The system's fluency is not independent authority.

example

An AI calls a choice “avoidance” after hearing one account. That label is an interpretation to examine—not a finding about why the person acted.

Fixed or outdated identity

A stored profile can keep shaping later responses after a person's circumstances or preferences have changed. A system-generated proposal should not become a settled account of someone simply because it was saved or repeated.

example

A profile remembers that you once preferred detailed guidance. You now want short answers, but the old preference continues to shape replies until you can inspect and change the relevant setting or profile.

Errors and overconfidence in sensitive situations

An AI may be wrong about a fact, source, personal experience or relationship. It should not diagnose someone, invent a hidden cause, claim to know another person's motives or treat a bodily signal as proof of a conclusion. High-stakes matters require appropriate limits and human expertise.

example

Feeling tense during a conversation may be worth reflecting on, but it does not by itself prove that another person is unsafe or acting with a particular intention.

How DIALOGIC approaches relational AI safety

The design aims below are not guarantees. The safeguards register distinguishes source-visible controls from model behaviour that has not been independently verified in a deployed environment.

Pacing, self-observation and inner signals

Dialogue can leave room for someone to notice their own reactions before moving forward. ARIS may invite a person to describe an inner signal as information to explore. A sensation or feeling is not proof that an interpretation is true, and reflection should not be treated as a demand for disclosure or action.

Proportionate challenge and restraint

A useful response may distinguish observation, interpretation, assumption and evidence, or offer another perspective. Challenge should depend on context and timing—not become a default demand to disclose, agree or act. A person can reject a suggestion, pause or leave a question unresolved. Generated behaviour can still miss that mark.

User-shaped conversational style

Some authenticated settings let users shape aspects of the conversational style and ARIS personality text. These controls are separate from core safety boundaries and do not make the system's interpretations authoritative. They are not one universal “personality” switch, and the exact settings vary by feature.

Inspectable and revisable personalisation

DIALOGIC has more than one distinct profile or continuity feature. Their purposes and controls differ. A profile that can be viewed or cleared is not the same thing as a transcript, source activity, or every other profile; clearing one derived summary does not mean all source records or provider-held data are erased.

Capacities beyond chat

Learning, inquiry, creative and self-observation spaces can give people ways to examine an idea, practise a skill and reflect outside the conversation. They are distinct experiences, not one combined profile or a promise that a particular capacity will develop. Connections between features should be invitations to explore, not hidden conclusions about identity.

Human participation without unsupported promises

Human relationships matter because people have independent needs and perspectives. We do not claim that a governed, moderated human-led support group is currently operating. Any such group would need clear boundaries, consent, moderation and appropriate routes to help before it could be represented as available.

How the parts differ

A conversational system, its stored context, feature-specific learning records and an approved handoff are different things. Their permissions should not be collapsed into a single promise about “memory.”

ARIS

ARIS (Adaptive Relational Intelligence System) is the conversational layer. Its modes can listen, clarify, separate kinds of claims, ask a question, challenge an assumption or help develop an idea. The user can make a request in the conversation, but a request is not a guaranteed setting and prompt-guided behaviour can vary.

RII and conversational preferences

The Relational Intelligence Interface (RII) has an authenticated profile area for editing its answers and some ARIS style/personality preferences. These settings are distinct from the Living Profile and AI Literacy observations. There is no single master switch in the audited interface that disables every form of conversational personalisation.

control

Open RII profile settings (sign-in required)

Living Profile

The authenticated ARIS entry has a separate Living Profile panel. It shows a synthesised profile and source list, lets a user switch use in ARIS on or off, choose whether selected Held and Heart Space sources are included, refresh the summary, or clear that summary. Clearing the summary does not delete the underlying source activity. These controls are specific to this profile and ARIS context.

control

Open ARIS and Living Profile controls (sign-in required)

AI Literacy observations

The AI Literacy profile is separate from both RII and Living Profile. Its authenticated page lets a user inspect observations, mark them as accurate, unsure, disagreeing or context-only, add a correction, or remove an observation. Turning off its memory consent prevents new observations from being saved; existing observations are not automatically erased by that switch and must be reviewed or removed separately.

control

Open the AI Literacy profile (sign-in required)

Feature-specific handoffs

A supported ARIS-to-Inner-Inquiry handoff is a bounded proposal: the person can decline, review or edit it, and must confirm before it is sent. Arrival controls allow acknowledgement, dismissal or revocation. This is a specific transfer, not blanket consent for all cross-feature sharing. Other features may have their own separate permissions.

control

Review account consent and deletion options (sign-in required)

Learning, inquiry and creative spaces

These are separate activities with their own content, permissions and boundaries. A handoff or feature connection should be reviewed for the destination and purpose. Inner Inquiry has server-side profile operations, but the audited Inner Inquiry screens do not expose an editor for that profile; this page does not claim that the API is a user-facing control.

AI literacy is something we practise

AI literacy includes understanding what an AI system can do, where it can fail and how it may influence a decision. Within DIALOGIC, it also means being able to notice the interaction while it is happening.

The authenticated AI Literacy experience is behind sign-in; this page is the public explanation, not a public AI Literacy route.

What do I know from my own experience, and what is an AI interpretation?

What evidence supports this claim? What remains uncertain?

Is my inner signal helping me notice something that needs exploration, or am I treating it as proof?

Has the system remembered me accurately? What would I change or delete?

Do I want the AI to listen, question, challenge, slow down or stop?

What would help me take this into my own work, relationships or wider life?

Seven design commitments

These commitments guide DIALOGIC's design. They are requirements to test against, not proof that every safeguard is already present or effective.

Agency

People should be able to choose direction, decline a practice, pause and leave. They can correct a profile where that feature provides a user-facing correction control; where no such control exists, the limitation should be clear. AI assists judgment rather than claiming final authority.

Epistemic integrity

Facts, user statements, interpretations and hypotheses remain distinguishable. Confidence should reflect evidence; uncertainty should be visible.

Revisable continuity

Memory and profiles can inform a conversation without defining who the person must remain. Sensitive inferences require restraint and correction paths.

Relational boundaries

Warmth does not imply that the AI has feelings, intentions or human obligations. Its role and limits should be clear.

Privacy and consent

Sensitive reflections and cross-feature sharing require purpose-specific controls, limited access and meaningful deletion or correction paths.

Human connection

The design should strengthen participation beyond DIALOGIC, including appropriate contact with other people and human-led support where it is properly governed.

Accountability

Risks need testing, review, user feedback and revision. A stated principle is not evidence that a safeguard works.

Expectations for responsible design

Responsible design should explain the AI's role and limits in accessible language; make it possible to question its interpretations; offer feature-specific ways to shape, correct or remove personalisation where those controls exist; and keep examining harms that can emerge over time, not only in isolated messages.

It should distinguish tested behaviour from design intention and communicate substantive changes to this approach. These are standards to assess and hold the platform to, not proof that every item is in place, an independent evaluation or a promise of a date.

Safeguards have limits

Safeguards can reduce some risks; they do not eliminate risk or guarantee a particular outcome. AI literacy does not transfer platform accountability onto the person using the system.

Human expertise and relationships remain essential

DIALOGIC is not a substitute for clinical care, safeguarding expertise or trusted human relationships. For urgent danger, use the appropriate emergency service where you are.

What remains with the human

Relational Technology should be judged by what remains with the human: greater capacity to observe, discern, create, choose and participate when the conversation ends.

Public safeguards register

This register records controls visible in the current source tree, not an independent certification or proof of production deployment. Test files cited below are evidence references, not proof that those tests were executed for this publication.

Status terms: “Live in source only” means a corresponding UI/API path exists in this workspace; it does not establish production deployment, operational staffing, test execution or effectiveness. “In development” and “Planned” are used only when approved evidence supports those statuses; none is established for this register. “Unknown / unverified” means the available evidence does not establish a current capability or setting.

No human accountable owner or human review date has been recorded for this draft. Where deployment, model behaviour, operational ownership or provider configuration is not verified, the register says so rather than inferring it.

Scroll horizontally to read every column. The column headings stay visible while you scroll through the register.

Safeguards by risk, response, user-visible control, implementation status, owner, review date, evidence and limitation.
Risk	Design response	User-visible control	Implementation status	Accountable owner	Last review	Evidence / verification reference	Known limitation
Mirror Problem and Validation Spiral	Separate observation, interpretation, assumption and evidence; allow a user to ask for challenge or another perspective.	A person can ask ARIS in the conversation to identify what is known, what was inferred, what evidence is missing, or to offer another view. This is a free-text request, not a dedicated switch or guaranteed response. Start at /app/begin (sign-in required).	Live in source only for general chat and request/output guardrail paths. These are not a dedicated control for preventing cumulative mirroring or validation. Deployment and effectiveness are unverified.	Not assigned in the publication record; a human accountable owner is still required.	Not recorded; human review pending.	artifacts/api-server/src/chatGuardrails.ts and artifacts/api-server/src/routes/routes.ts (ARIS turn and provider boundary); tests exist at artifacts/api-server/src/tests/chat-guardrails.test.ts. This publication task did not run live-provider tests.	Prompt-guided dialogue is probabilistic. A request or instruction cannot guarantee that an answer will avoid mirroring, validate a belief, or make uncertainty clear.
Misplaced authority and outdated personalisation	Make distinct forms of personalisation inspectable and revisable where a user-facing control exists; do not treat a repeated proposal as settled identity.	RII profile and some ARIS style/personality settings: /app/rii-profile. Living Profile source list, ARIS-use toggle, selected-source consent, refresh and clear controls: /app/begin. Both require sign-in.	Live in source only for the listed RII and Living Profile controls. Production deployment and end-to-end behaviour are unverified.	Not assigned in the publication record; a human accountable owner is still required.	Not recorded; human review pending.	artifacts/raice/src/pages/app/rii-profile.tsx; artifacts/raice/src/pages/app/begin.tsx (Living Profile panel); artifacts/api-server/src/routes/routes.ts (RII and /api/living-profile routes).	RII preferences and Living Profile are distinct. Clearing the derived Living Profile does not delete source activity. No single control disables all forms of personalisation. The deployed settings were not checked.
A developmental observation becoming an outdated or inaccurate identity claim	Keep AI Literacy observations separate, reviewable and correctable; let the user remove individual observations.	Authenticated profile at /app/ai-literacy/profile: inspect observations, choose an accuracy/context response, correct or remove an item, and control consent for saving new observations.	Live in source only for the listed observation profile and controls. Deployment and execution of the cited tests are unverified.	Not assigned in the publication record; a human accountable owner is still required.	Not recorded; human review pending.	artifacts/raice/src/pages/app/ai-literacy/profile.tsx; artifacts/api-server/src/aiLiteracy.ts; source tests at artifacts/api-server/src/tests/ai-literacy.test.ts.	This is not the RII or Living Profile. Turning off consent prevents new observations but does not automatically remove observations already saved; those need separate correction or removal.
Sensitive context being carried to another feature without informed choice	Use a bounded, destination-specific handoff that can be reviewed and explicitly confirmed, rather than blanket cross-feature consent.	For the supported ARIS-to-Inner-Inquiry handoff, decline, review/edit, confirm, then acknowledge, dismiss or revoke at arrival. Account and consent options are at /app/account/consent (sign-in required).	Live in source only for the listed handoff and lifecycle API. Production deployment and execution of the cited integration test are unverified here.	Not assigned in the publication record; a human accountable owner is still required.	Not recorded; human review pending.	artifacts/raice/src/components/chat-proposals.tsx; artifacts/raice/src/components/routing-arrival.tsx; artifacts/api-server/src/tests/aris-routing-lifecycle.integration.test.ts.	This is one specific supported transfer, not blanket consent for all features. Each destination and purpose may have separate permissions. Other routes are not represented as covered by this control.
Emotional over-reliance or displacement of human contact	Keep the AI's non-human role clear and support participation outside the platform; do not claim that availability or a warm tone is human care.	A user can stop or leave a conversation. General feedback is available at /app/feedback (sign-in required). The authenticated support menu at /app/begin includes AI-based crisis signposting, not a staffed escalation service.	Unknown / unverified for product-level dependence safeguards and governed human-led support groups. General feedback and AI signposting are live in source only; deployment and operation are unverified.	Not assigned in the publication record; a human accountable owner is still required.	Not recorded; human review pending.	artifacts/raice/src/pages/app/feedback.tsx; artifacts/raice/src/components/aris/aris-support-menu.tsx; artifacts/raice/src/components/crisis-support.tsx.	No dedicated dependence-reduction control, staffed crisis escalation, response-time commitment or currently operating governed human-led support group is verified. Crisis Support is AI signposting and is explicitly not a helpline.
Errors or overconfidence in sensitive situations	Use request/output guardrails and clearly direct people to appropriate human expertise for high-stakes matters.	A user can stop, seek a trusted person or qualified professional, and send general feedback at /app/feedback (sign-in required). The public Privacy Notice says AI can be wrong and important information should be independently checked.	Live in source only for request/output guardrail and general feedback paths. Deployed effectiveness is unverified.	Not assigned in the publication record; a human accountable owner is still required.	Not recorded; human review pending.	artifacts/api-server/src/chatGuardrails.ts; artifacts/api-server/src/modePolicyEnforcement.ts; artifacts/raice/src/pages/privacy.tsx; artifacts/raice/src/pages/app/feedback.tsx.	Guardrails and model instructions can fail. Feedback is not a dedicated safety-reporting channel or a promise of escalation, response time or resolution.
Unclear model-provider training, retention or deletion expectations	Describe only verified provider-specific account settings and routes; avoid a blanket training, retention or erasure promise.	Read the current Privacy Notice at /privacy and the General Terms at /terms. Account deletion can be requested at /app/account/consent (sign-in required); the deletion contract lists records that may remain.	Unknown / unverified: training opt-out and retention settings have not been verified across every active and fallback provider path for this publication.	Not assigned in the publication record; a human accountable owner is still required.	Not recorded; provider-setting review pending.	artifacts/raice/src/pages/privacy.tsx; artifacts/raice/src/pages/terms.tsx; artifacts/raice/src/components/account-ai-data-section.tsx; artifacts/api-server/src/tests/aris-private-context-consent.test.ts.	Published policy describes variable provider handling and is not evidence of a provider account's actual training or retention settings. Account deletion does not promise immediate erasure of provider, legal, security, backup or deferred records.
A human-reporting or escalation route being mistaken for available support	Identify the actual route and its limits; distinguish general feedback and AI signposting from a human-operated response service.	General feedback: /app/feedback (sign-in required). Crisis Support signposting is reachable from the authenticated support menu at /app/begin. Immediate danger requires local emergency services.	Live in source only for general feedback and AI signposting. Dedicated human safety reporting, moderation and escalation operations are unknown / unverified.	Not assigned in the publication record; a human accountable owner is still required.	Not recorded; human review pending.	artifacts/raice/src/pages/app/feedback.tsx; artifacts/raice/src/components/aris/aris-support-menu.tsx; artifacts/raice/src/components/crisis-support.tsx; artifacts/api-server/src/routes/routes.ts.	The feedback page is general-purpose and authenticated. Crisis Support is AI signposting, not a helpline; no dedicated human escalation service or service-level commitment is established by the cited source.

Risk

Mirror Problem and Validation Spiral

Design response

Separate observation, interpretation, assumption and evidence; allow a user to ask for challenge or another perspective.

User-visible control

A person can ask ARIS in the conversation to identify what is known, what was inferred, what evidence is missing, or to offer another view. This is a free-text request, not a dedicated switch or guaranteed response. Start at /app/begin (sign-in required).

Implementation status

Live in source only for general chat and request/output guardrail paths. These are not a dedicated control for preventing cumulative mirroring or validation. Deployment and effectiveness are unverified.

Accountable owner

Not assigned in the publication record; a human accountable owner is still required.

Last review

Not recorded; human review pending.

Evidence / verification reference

artifacts/api-server/src/chatGuardrails.ts and artifacts/api-server/src/routes/routes.ts (ARIS turn and provider boundary); tests exist at artifacts/api-server/src/tests/chat-guardrails.test.ts. This publication task did not run live-provider tests.

Known limitation

Prompt-guided dialogue is probabilistic. A request or instruction cannot guarantee that an answer will avoid mirroring, validate a belief, or make uncertainty clear.

Risk

Misplaced authority and outdated personalisation

Design response

Make distinct forms of personalisation inspectable and revisable where a user-facing control exists; do not treat a repeated proposal as settled identity.

User-visible control

RII profile and some ARIS style/personality settings: /app/rii-profile. Living Profile source list, ARIS-use toggle, selected-source consent, refresh and clear controls: /app/begin. Both require sign-in.

Implementation status

Live in source only for the listed RII and Living Profile controls. Production deployment and end-to-end behaviour are unverified.

Accountable owner

Not assigned in the publication record; a human accountable owner is still required.

Last review

Not recorded; human review pending.

Evidence / verification reference

artifacts/raice/src/pages/app/rii-profile.tsx; artifacts/raice/src/pages/app/begin.tsx (Living Profile panel); artifacts/api-server/src/routes/routes.ts (RII and /api/living-profile routes).

Known limitation

RII preferences and Living Profile are distinct. Clearing the derived Living Profile does not delete source activity. No single control disables all forms of personalisation. The deployed settings were not checked.

Risk

A developmental observation becoming an outdated or inaccurate identity claim

Design response

Keep AI Literacy observations separate, reviewable and correctable; let the user remove individual observations.

User-visible control

Authenticated profile at /app/ai-literacy/profile: inspect observations, choose an accuracy/context response, correct or remove an item, and control consent for saving new observations.

Implementation status

Live in source only for the listed observation profile and controls. Deployment and execution of the cited tests are unverified.

Accountable owner

Not assigned in the publication record; a human accountable owner is still required.

Last review

Not recorded; human review pending.

Evidence / verification reference

artifacts/raice/src/pages/app/ai-literacy/profile.tsx; artifacts/api-server/src/aiLiteracy.ts; source tests at artifacts/api-server/src/tests/ai-literacy.test.ts.

Known limitation

This is not the RII or Living Profile. Turning off consent prevents new observations but does not automatically remove observations already saved; those need separate correction or removal.

Risk

Sensitive context being carried to another feature without informed choice

Design response

Use a bounded, destination-specific handoff that can be reviewed and explicitly confirmed, rather than blanket cross-feature consent.

User-visible control

For the supported ARIS-to-Inner-Inquiry handoff, decline, review/edit, confirm, then acknowledge, dismiss or revoke at arrival. Account and consent options are at /app/account/consent (sign-in required).

Implementation status

Live in source only for the listed handoff and lifecycle API. Production deployment and execution of the cited integration test are unverified here.

Accountable owner

Not assigned in the publication record; a human accountable owner is still required.

Last review

Not recorded; human review pending.

Evidence / verification reference

artifacts/raice/src/components/chat-proposals.tsx; artifacts/raice/src/components/routing-arrival.tsx; artifacts/api-server/src/tests/aris-routing-lifecycle.integration.test.ts.

Known limitation

This is one specific supported transfer, not blanket consent for all features. Each destination and purpose may have separate permissions. Other routes are not represented as covered by this control.

Risk

Emotional over-reliance or displacement of human contact

Design response

Keep the AI's non-human role clear and support participation outside the platform; do not claim that availability or a warm tone is human care.

User-visible control

A user can stop or leave a conversation. General feedback is available at /app/feedback (sign-in required). The authenticated support menu at /app/begin includes AI-based crisis signposting, not a staffed escalation service.

Implementation status

Unknown / unverified for product-level dependence safeguards and governed human-led support groups. General feedback and AI signposting are live in source only; deployment and operation are unverified.

Accountable owner

Not assigned in the publication record; a human accountable owner is still required.

Last review

Not recorded; human review pending.

Evidence / verification reference

artifacts/raice/src/pages/app/feedback.tsx; artifacts/raice/src/components/aris/aris-support-menu.tsx; artifacts/raice/src/components/crisis-support.tsx.

Known limitation

No dedicated dependence-reduction control, staffed crisis escalation, response-time commitment or currently operating governed human-led support group is verified. Crisis Support is AI signposting and is explicitly not a helpline.

Risk

Errors or overconfidence in sensitive situations

Design response

Use request/output guardrails and clearly direct people to appropriate human expertise for high-stakes matters.

User-visible control

A user can stop, seek a trusted person or qualified professional, and send general feedback at /app/feedback (sign-in required). The public Privacy Notice says AI can be wrong and important information should be independently checked.

Implementation status

Live in source only for request/output guardrail and general feedback paths. Deployed effectiveness is unverified.

Accountable owner

Not assigned in the publication record; a human accountable owner is still required.

Last review

Not recorded; human review pending.

Evidence / verification reference

artifacts/api-server/src/chatGuardrails.ts; artifacts/api-server/src/modePolicyEnforcement.ts; artifacts/raice/src/pages/privacy.tsx; artifacts/raice/src/pages/app/feedback.tsx.

Known limitation

Guardrails and model instructions can fail. Feedback is not a dedicated safety-reporting channel or a promise of escalation, response time or resolution.

Risk

Unclear model-provider training, retention or deletion expectations

Design response

Describe only verified provider-specific account settings and routes; avoid a blanket training, retention or erasure promise.

User-visible control

Read the current Privacy Notice at /privacy and the General Terms at /terms. Account deletion can be requested at /app/account/consent (sign-in required); the deletion contract lists records that may remain.

Implementation status

Unknown / unverified: training opt-out and retention settings have not been verified across every active and fallback provider path for this publication.

Accountable owner

Not assigned in the publication record; a human accountable owner is still required.

Last review

Not recorded; provider-setting review pending.

Evidence / verification reference

artifacts/raice/src/pages/privacy.tsx; artifacts/raice/src/pages/terms.tsx; artifacts/raice/src/components/account-ai-data-section.tsx; artifacts/api-server/src/tests/aris-private-context-consent.test.ts.

Known limitation

Published policy describes variable provider handling and is not evidence of a provider account's actual training or retention settings. Account deletion does not promise immediate erasure of provider, legal, security, backup or deferred records.

Risk

A human-reporting or escalation route being mistaken for available support

Design response

Identify the actual route and its limits; distinguish general feedback and AI signposting from a human-operated response service.

User-visible control

General feedback: /app/feedback (sign-in required). Crisis Support signposting is reachable from the authenticated support menu at /app/begin. Immediate danger requires local emergency services.

Implementation status

Live in source only for general feedback and AI signposting. Dedicated human safety reporting, moderation and escalation operations are unknown / unverified.

Accountable owner

Not assigned in the publication record; a human accountable owner is still required.

Last review

Not recorded; human review pending.

Evidence / verification reference

artifacts/raice/src/pages/app/feedback.tsx; artifacts/raice/src/components/aris/aris-support-menu.tsx; artifacts/raice/src/components/crisis-support.tsx; artifacts/api-server/src/routes/routes.ts.

Known limitation

The feedback page is general-purpose and authenticated. Crisis Support is AI signposting, not a helpline; no dedicated human escalation service or service-level commitment is established by the cited source.

Questions people ask

What are guardrails in conversational AI?

Guardrails are rules and checks intended to limit harmful or unsupported requests and outputs, protect boundaries, and direct some situations toward appropriate help. They are one layer of safety: they do not establish that a relationship is safe over time, and they cannot guarantee that every response will follow an instruction.

Can AI validation reinforce a user's beliefs?

Yes. Repeated agreement can make a tentative belief feel settled without adding evidence. Ask what is known, what has been inferred, what evidence supports the claim and what would count against it. An AI's agreement is not independent confirmation.

How can AI companions affect human agency and relationships?

Warmth, availability and personalisation can make an AI exchange feel significant. Over time, that may affect how someone spends attention or approaches people. The effect is not the same for everyone and cannot be inferred from one conversation. Notice whether the interaction supports your own choices and wider relationships; AI is not a human relationship or a substitute for care.

What does AI literacy look like during a conversation?

It can mean pausing to separate your experience from an AI interpretation, asking for evidence and uncertainty, checking what the system may have remembered, and choosing whether you want it to listen, question, challenge, slow down or stop. Literacy supports agency; it does not move platform responsibility onto you.

How should memory and personalisation work in responsible relational AI?

They should have a clear purpose, remain distinguishable from verified facts, respect feature-specific consent and be open to correction or removal. In practice, DIALOGIC has distinct profiles with different controls and limits. Not every profile has a user-facing editor, and clearing one summary does not erase every source or provider record. The register above describes those differences.