Help · AI transparency
Which parts of ConKarma are AI, and which deliberately aren't.
This page answers a different question from our AI privacy page. That one is about data — what the AI can see, who processes it, whether it trains on you. This one is about disclosure: which surfaces are AI at all, which ones we have deliberately kept human, and how you tell generated content from the rest.
ConKarma discloses which of its surfaces are AI and which are deliberately not. AI produces content only in features you turn on: voice-memo and voice-journal summaries, the generated narratives (weekly recap, relationship reflections, family-story drafts), the Family and Relationship Coach, the parent-side Fairness dashboard, tap-to-use helpers, Cosmic readings, optional speech-to-text, image generation where offered, and child-safety moderation of submitted text and uploaded images. There is no general-purpose chat surface inside the cell and no background AI reading along as you use the app.
Three exclusions are deliberate: conflict resolution is never AI-mediated (ADR-029), safety-critical Workshop advice is never AI-generated (ADR-061), and Family Shield is rule-based rather than an AI system. Adult-zone (Ember) content is not general AI input — automated processing occurs only under the explicit adult-feature consent you give (ADR-120, GDPR Art. 9(2)(a)). Disclosure that you are talking to an AI is built into both coaches, which are in preparation until a conversational provider is live; marking the generated output itself is rolling out as of 20 September 2026, with verification across surfaces in progress. Every AI call runs through a vendor-neutral gateway contracted with Anthropic, OpenAI, and Google.
The features that use AI
AI produces content in a bounded set of places, all of them features you turn on: summaries of voice memos and voice journals; the generated narratives — the weekly cell recap, relationship reflections and reports, family-story drafts; the Family Coach and the Relationship Coach, once they ship — both are in preparation today; the parent-side Fairness dashboard; small helpers you tap — suggestions on the dashboard, natural-language task entry, the generational translator, care triage; Cosmic readings (the daily horoscope and the birth chart); speech-to-text when you turn it on for a voice note; image generation where it is offered; and child-safety moderation, which screens free text you submit and classifies uploaded images before they are shown to others. There is no general-purpose chat surface inside the cell, and no AI reading along as you use the app. The complete per-feature list below is read live from the same registry the gateway runs on; the sub-processor list records what each vendor receives.
Every AI feature, read from the registry
The 46 features below are every feature that sends something to an AI provider through our gateway, read live from the gateway's own configuration — the same list the app runs on, not a list we maintain by hand. Grouped by how each one runs.
Runs when you tap it (27)
Each of these runs only because you invoked it.
- Coach Thank YouAI-generated outputchild accounts: text-helper consent
What leaves the system: the short thank-you the family wrote to a coach
- Cosmic Birth ChartAI-generated output
What leaves the system: first name and date of birth
- Cosmic Couple ChartAI-generated output
What leaves the system: both partners' first names and dates of birth
- Echo ExplainerAI-generated output
What leaves the system: the family Echo's text and attribution, to rewrite per age tier
- Echo Personalized ReflectionAI-generated output
What leaves the system: the Echo's text and author name
- Family Story KickstartAI-generated output
What leaves the system: the story seed you typed
- Feature RecommenderAI-generated output
What leaves the system: your user id and the week — no content
- Gaming Concept ExplainerAI-generated output
What leaves the system: the gaming term you asked about and the reader's persona (teen/adult/grandparent)
- Generational TranslatorAI-generated outputchild accounts: text-helper consent
What leaves the system: the phrase you asked to translate and the two generations
- Heritage Narrative WeaveAI-generated outputchild accounts: text-helper consent
What leaves the system: the interview template's questions and your (already transcribed) answers
- Joysticks Year in ReviewAI-generated output
What leaves the system: counts of completed gaming sessions, tournaments and skills for the year — no session details
- Memory Wall LayoutAI-generated output
What leaves the system: the chosen layout template key — no photos
- Mentorship Topic SuggestAI-generated outputchild accounts: text-helper consent
What leaves the system: the mentorship pair's stated interests
- Natural Language Task ParseAI-generated outputchild accounts: text-helper consent
What leaves the system: the task sentence you typed
- No Phone Activity SuggestAI-generated outputchild accounts: text-helper consent
What leaves the system: the cell's size and the no-phone hour's setting
- Numerology ReadingAI-generated output
What leaves the system: full name and the computed life-path number
- Parent Prep HelperAI-generated output
What leaves the system: the conversation topic and the child's age band — no child name
- Pass The Controller HelperAI-generated output
What leaves the system: the game category being coached
- Relationship Reflection GenerateAI-generated output
What leaves the system: the reflection prompt you selected and your written answer
- Spotlight Practice ObservationAI-generated output
What leaves the system: the child's display name and the activity type
- Spotlight Year in ReviewAI-generated output
What leaves the system: the spotlight's display name and activity type
- Study Habit HelperAI-generated output
What leaves the system: the subject and the child's age band — no child name
- Vibe ExplainerAI-generated outputchild accounts: text-helper consent
What leaves the system: the family slang term and its vibe key, to rewrite per age tier
- Viral Challenge Idea GeneratorAI-generated output
What leaves the system: the cell's shape (family/friends) and a short context you provide
- Wonder Question PersonalizeAI-generated outputchild accounts: text-helper consent
What leaves the system: the wonder question template and the child's age band — not dispatched by any product surface today
- Workshop Health SummaryAI-generated output
What leaves the system: the hearth's positive observations and repair categories
- Workshop Howto ExplainerAI-generated outputchild accounts: text-helper consent
What leaves the system: the how-to's title, skill area and safety category, and the reader's persona
Runs on a schedule after you opt in (11)
You turn these on once; they then run on their own cadence until you turn them off.
- Awards Night SpeechAI-generated outputchild accounts: text-helper consent
What leaves the system: the winning award category name only — never member names, numbers or comparisons
- Birthday Year ReviewAI-generated outputchild accounts: text-helper consent
What leaves the system: the celebrant's first name plus counts of missions completed, Warmth received and journal entries this year — no entry text
- Cell Mini GameAI-generated outputchild accounts: text-helper consent
What leaves the system: the game type and a count of the cell's recent journal entries — no entry text
- Cosmic Couple CompatibilityAI-generated output
What leaves the system: both partners' first names and sun signs, and the week
- Cosmic Daily HoroscopeAI-generated output
What leaves the system: sun sign and first name
- Fairness ScorecardAI-generated outputchild accounts: text-helper consent
What leaves the system: per-member counts of assigned and completed chores for the week, keyed by first name
- Family Lore BookAI-generated outputchild accounts: text-helper consent
What leaves the system: the month, the cell name and counts of shared journal entries and completed missions — no entry text
- Mood WeatherAI-generated outputchild accounts: text-helper consent
What leaves the system: the cell's mood distribution for the last 24 hours as counts — no names
- Pulse Personalized CopyAI-generated output
What leaves the system: the re-engagement segment key only — no name, no activity
- Sunday Highlight ReelAI-generated outputchild accounts: text-helper consent
What leaves the system: counts of the week's shared entries and Warmth given — no entry text
- Weekly RecapAI-generated output
What leaves the system: the cell name and the text of the past week's shared journal entries
System checks on what you submit (8)
Not content you see — validators and safety checks that run on what is submitted.
- Coach Validate Mental Health SafeAI-generated output
What leaves the system: a draft AI coach reply being checked for crisis-unsafe wording — nothing about the user
- Coach Validate Non ClinicalAI-generated output
What leaves the system: a draft AI coach reply being checked for clinical advice — nothing about the user
- Coach Validate Positive FrameAI-generated output
What leaves the system: a draft AI coach reply being checked for a positive frame — nothing about the user
- Content Moderation Imagenothing generated for you
What leaves the system: the uploaded photo or image being safety-classified
- Content Moderation Textnothing generated for you
What leaves the system: the submitted text being safety-classified
- Sentiment ClusteringAI-generated output
What leaves the system: aggregated journal sentiment signals — not dispatched by any product surface today
- Support TriageAI-generated output
What leaves the system: the title and description of the support request you submitted
- Translation KickstartAI-generated output
What leaves the system: the English UI string being translated and the target locale — no user data
In preparation — configured, not reachable yet (8)
Listed so the configuration is public. None of these can be used today — no provider serves the capability yet, or the surface is switched off — so nothing is sent for them.
- Cell Crest GeneratorAI-generated output
Will send: the symbol prompts each member submitted for the family crest
- Couples PortraitAI-generated output
Will send: the two selfies the couple provided, the chosen art style and occasion
- Entity Thumbnail ImageAI-generated output
Will send: a family item's title and kind to illustrate — no personal data
- Family Coach Chatyou're told it's an AI
Will send: the messages you type into the family coach chat and the reply history of that session
- Relationship Coachyou're told it's an AI
Will send: the messages you type into the relationship coach and that session's history
- Selfie AvatarAI-generated output
Will send: the selfie you provided and the chosen avatar style
- Voice Memo SummaryAI-generated output
Will send: the transcript of your voice memo
- Voice Memo Transcriptionnothing generated for you
Will send: the voice memo audio you recorded
Providers and models are not published per feature; the sub-processor list records what each vendor receives. Child-consent classes are granted per class by a verified parent in the Parent Dashboard.
The features that deliberately don't
Three exclusions are architectural decisions, not gaps we intend to fill. Conflict resolution is never AI-mediated (ADR-029): the structured scaffolding — turn-taking, cooling-off, the shape of the conversation — is the thing that helps, and handing it to a model would replace the part that works. Safety-critical Workshop advice is never AI-generated (ADR-061). And Family Shield is a rule-based system, not an AI system: it evaluates conditions you configure, and its decisions are traceable to those rules rather than to a model's inference.
Adult-zone content is not general AI input
No automated content model processes Ember content without your explicit consent (ADR-120). Automated or AI processing of Ember content happens only where you opt into an AI-powered adult feature, or where child-safety moderation applies to text you submit — and in both cases only under the explicit adult-feature consent you give (GDPR Art. 9(2)(a)). The full enumeration of the boundary, and the other thirteen guarantees that enforce it, is on our zone-isolation page.
Telling AI output from human content
Our standard has two halves, and they are at different stages. Knowing when you are talking to an AI is built in: both the Family Coach and the Relationship Coach are designed to disclose that you are interacting with an AI system before your first message — and both coaches are in preparation until a conversational provider is live, so today no conversational AI is reachable at all. Marking the generated output itself is being rolled out. Our Terms of Service say AI-generated output is identified as AI-generated where it appears and carries a machine-readable record of the system that produced it. As of 20 September 2026 the app renders that mark on every AI-output surface that receives a provenance record, and our backend now attaches the record to every generated artefact; verification surface by surface is in progress, with six confirmed on device so far. We would rather date the rollout here than describe a finished control. This paragraph will say "complete" when the count is complete, and that change will be dated too.
Nothing runs that you didn't turn on
AI runs because you turned something on. Most features run when you tap them — the summary, the prompt set, the helper. A few run on a schedule once you opt in — the daily horoscope, the weekly recap, the parent-side Fairness scorecard — and keep running until you turn them off. Nothing spends your Sparks or produces content in your cell that you did not opt into. Safety moderation is the one exception, and it is not a content feature: it screens what you submit before others see it, which is a condition of publishing rather than something you invoke.
Children's accounts: AI is off until a parent turns each class on
For an account under 13, AI processing that sends the child's data to a provider is gated per class, not per feature: a photo of the child, the child's voice, a conversation with a model, and a one-shot text helper. Each class is off by default, is turned on only by a verified parent in the Parent Dashboard, can be turned off there at any time, and the gate fails closed — a class that has not been granted makes the feature simply unavailable to that child. A class that is not yet enforced is not shown as a permission the parent declined, because nothing of that kind is being sent. Withdrawing the overall parental consent withdraws every AI class at the same moment. Where an image feature is offered on a child's account, the photograph goes to the provider only under the image-of-child class, the source photo is discarded within 24 hours on our side, and no face template or faceprint is derived from it — the terms are in the Children's Privacy section of the Privacy Policy.
Who is actually behind the gateway
ConKarma routes every AI call through a vendor-neutral AI Gateway holding parallel contracts with Anthropic, OpenAI, and Google. Which of them serves a given feature is a quality decision we make per feature and can change. We name the vendor set openly; we do not currently publish a per-feature provider-and-model catalog. The canonical sub-processor list records what each vendor receives and under what transfer safeguards.
Frequently asked questions
How is this different from the AI privacy page?
That page is about data: what the AI can read, which vendors process it, and our no-training stance. This page is about disclosure: which surfaces are AI at all, which we keep deliberately human, and how you tell generated content apart. Different questions, so they get different pages.
Is there an AI reading my cell in the background?
No. There is no general-purpose chat surface inside the cell and no background AI following along. AI produces content only in features you turned on — most when you tap them, a few on a schedule you opted into, like the daily horoscope or the weekly recap. The one thing that runs without any opt-in is child-safety moderation of content you submit, which screens it before others see it.
Can I tell which text was written by AI?
Partly. Knowing you are talking to an AI is built in — both coaches are designed to disclose it before your first message, though the coaches themselves are in preparation and no conversational AI is reachable today. Telling an individual piece of generated text apart is being rolled out: as of 20 September 2026 the app marks AI output where it appears, wherever our backend has attached the provenance record our Terms of Service describe — verification across surfaces is still in progress. We'll date this answer again when it's complete.
Does AI handle conflicts, safety advice, or Family Shield?
No, and not by omission. Conflict resolution is never AI-mediated (ADR-029) because the structured scaffolding is what helps. Safety-critical Workshop advice is never AI-generated (ADR-061). Family Shield is rule-based: it evaluates conditions you configure, so its decisions trace to rules rather than to model inference.
Which AI vendors does ConKarma use?
Anthropic, OpenAI, and Google, held in parallel contracts behind a vendor-neutral AI Gateway, picked per feature on quality. We name the vendor set openly but don't currently publish which specific model serves each individual feature. The canonical sub-processor list records what each vendor receives.