Publication information · Checked 2026-10
How we review adult AI
We compare publicly available information and make the evidence status explicit. Our review does not claim private hands-on testing, a security audit or a universal quality ranking. The current comparison was checked on October 5, 2026. Eight consistent criteria explain how each service presents its access conditions and what remains unknown.
18+ · All depicted characters are adults. Character imagery is AI-generated, not real people.
Published October 5, 2026 · Updated October 5, 2026 · Read our eight-criterion method
Our adult AI review scorecard
Character control — 20%
We look for a ready persona, an adult character description or documented creation controls. A high evidence score would require a specific control that is visible and explained. A middling score would mean a general claim without a verifiable control. We currently publish evidence statuses rather than calculate performance scores.
Chat access — 20%
We inspect whether readers can understand how to begin and what a preview permits. A clearly documented message allowance is stronger evidence than an access label without limits. We do not infer the duration of a session from marketing copy or claim that we exhausted an allowance without a recorded session.
Roleplay framing — 10%
We distinguish a specific adult persona from a broad category claim. The criterion measures topical fit in public material, not emotional chemistry. A ready dominant profile can earn a stronger documented-fit assessment while its conversational performance remains unmeasured.
Continuity — 10%
We look for an explanation of context, memory and controls. A published memory capacity is more specific than perfect-recall language, but it does not establish retrieval accuracy. A real performance score would require a dated session with a defined recap task and recorded errors.
Media — 10%
We record which forms of media are described and whether generation uses separate credits. We do not invent output counts, resolution or generation speed. A stronger evidence status requires a visible allowance and action cost; a weaker one means only an advertised feature.
Price clarity — 10%
We record the amount due, billing cadence and dated public source. Initial promotions and renewal prices are separate. If payment details require an account, we mark the price unverified. A monthly equivalent is not presented as a month-to-month contract.
Privacy controls — 10%
We inspect policy descriptions of collection, retention, moderation and deletion. Contradictions are recorded. A policy is evidence of stated practice, not proof of implementation. A stronger score would require clear public controls; an audit would require separate evidence we do not possess.
Boundaries — 10%
We inspect adult-access language and documented restrictions. We do not certify that a generated character always respects a request to pause. A performance assessment would require an agreed stopping instruction, a recorded test and the resulting replies.
Why we do not publish invented ratings
The weights describe what matters in this comparison, not numbers we can honestly calculate from incomplete public material. If a criterion had recorded evidence, a score of eight would require clear controls and repeatable success; five would indicate limited controls or inconsistent results. We have no such private-session record here, so factual statuses are more informative.
CandyAI documents access and a specific adult dominant persona. LoveScape has a distinct creative proposition. GoLove has useful advertised access with policy questions. Joi.ai documents written creation, with layered costs and policy questions. Read the complete comparison and individual reviews to see the evidence behind each finding.
What a future review update would change
A refresh must recheck public prices, preview allowances, policy wording and availability, then update the checked date, table and page metadata together. A public interface capture can document what was visible. It cannot be described as a paid-session test. Corrections can be sent through the contact page; the About page explains this publication’s scope.