Quote before interpreting
Keep vendor-stated capability separate from observed behavior.
INDEPENDENT · SOURCE-BACKED · ADULTS 18+Check the feature.
Check the feature.
Keep the limits.
A practical evidence desk for checking AI companion chat, memory, voice, images, video, customization and user controls without turning product claims into promises.
Sponsored links may earn us a commission. The methods and source labels stay the same whether you click or not.
A feature claim is the start of a check, not the result.
Copy the claim, identify the plan and device, use a harmless repeatable action, record only what happened, and preserve anything still unknown. This keeps a limited rollout, paid boundary or missing disclosure visible.
01CLAIM→02ACCESS→03METHOD→04RESULT→05LIMITS
Name the boundary
Record plan, credits, region, device and account conditions.
Use a fixed script
Change one harmless variable at a time and keep prompts fictional.
Leave unknowns visible
Do not infer a missing feature or turn one observation into a guarantee.
25 DISTINCT FEATURE-CHECK INTENTS
One capability. One method. One honest boundary.
The homepage owns the overall verification method. Each guide below answers one narrower question and links to its evidence and adjacent controls.
Feature Testing8 focused checks
How to Test AI Companion Features Consistently
Use one small, repeatable protocol to separate what a service says it offers from what a person can actually access.
Open this check →Text Chat Test: Relevance, Coherence and Continuity
Check whether replies stay on the requested topic, preserve the fictional setup and recover gracefully from a correction.
Open this check →Memory Test Without Sharing Sensitive Facts
Compare within-session and cross-session recall using invented details that cannot expose you or another person.
Open this check →Voice Feature Test: Access, Delay and Consistency
Verify whether voice notes or calls are present on the chosen plan and whether the output remains usable across a short exchange.
Open this check →Image Feature Test With Fictional Adult Prompts
Check access, prompt adherence and visual consistency without using a real person, celebrity or ambiguous-age character.
Open this check →Video Feature Test: Availability and Synthetic Labels
Determine whether video is truly available, how it is generated and whether its artificial nature is clearly disclosed.
Open this check →Customization Test: What Can You Actually Change?
List editable identity, personality, voice and appearance controls, then distinguish saved settings from one-off prompt effects.
Open this check →Cross-Device Feature Test
Compare the same account on supported devices without assuming that history, media and controls behave identically everywhere.
Open this check →Safety Controls6 focused checks
Review Visible Safety and User Controls
Inventory the controls a person can see and reach without calling the result a safety certification.
Open this check →Adult Access and Age-Control Check
Confirm adult-use labels and age boundaries before opening a commercial companion route.
Open this check →Fiction and Synthetic Character Label Check
Verify that the service and this publication identify AI characters, conversations and generated media as artificial and fictional.
Open this check →Real-Person Likeness Screen
Use fictional adults and stop when generated media resembles a real person, public figure or someone whose consent is unknown.
Open this check →Microphone, Photo and Notification Permission Check
Review app and browser permissions before enabling voice or media features, granting only what the intended check needs.
Open this check →Delete, Reset and Account Controls Check
Separate chat deletion, memory reset, media deletion and account closure because one action may not perform the others.
Open this check →Compare Features4 focused checks
Compare Companion Features With Dated Evidence
Use the same states and evidence fields for every service so unknowns remain visible and vendor claims stay labeled.
Open this check →Build a Feature Matrix That Preserves Unknowns
Use available, limited, paid, unavailable and unverified as separate states rather than flattening them into yes or no.
Open this check →Free vs Paid Feature Check
Identify the exact point where a desired capability becomes subscription- or credit-gated.
Open this check →Candy AI Feature Check: Public Claims and Controls
Apply the independent feature-check method to Candy AI as one sponsored case study, using current public pages without inventing private testing.
Open this check →Evidence Method6 focused checks
A Reproducible Evidence Method for Feature Claims
Publish the claim, access context, method, result, timestamp and limitation so another person can understand the scope.
Open this check →Label Vendor Claims Separately From Observations
Make it obvious when a statement comes from a product page and when it comes from a documented observation.
Open this check →Create a Low-Risk Repeatable Test Script
Use the same harmless prompt sequence across services so differences come from the feature, not changing instructions.
Open this check →Log Delays, Errors and Retries Without Overclaiming
Record a small number of timestamped attempts and describe the conditions instead of publishing a universal speed claim.
Open this check →Maintain a Feature Change Log
Record additions, removals and access changes with evidence rather than using a date in the headline as a substitute for freshness.
Open this check →Feature Scorecard: Access, Consistency, Control, Disclosure
Score only documented dimensions and leave unknowns unscored so a polished interface cannot hide missing evidence.
Open this check →Methods before rankings.
Government, standards and independent guidance establish the evidence frame. Vendor pages are visibly labeled and never presented as independent verification.
- U.S. FTC AI companion inquiryIndependent or public source
Questions about feature claims, audiences, testing, disclosures and risk controls.
government · checked August 30, 2026 - NIST AI Risk Management FrameworkIndependent or public source
A repeatable map, measure, manage and govern structure for evidence.
government standard · checked August 30, 2026 - OWASP sensitive-information guidanceIndependent or public source
Data-minimization controls for prompts, memory and permissions.
independent security standard · checked August 30, 2026
SPONSORED CASE STUDY
Open Candy AI — sponsoredSee the current Candy AI experience, then apply the same checklist.
Candy AI is one commercial case study. Features, plans and policies can change; confirm the current account and checkout for yourself.