Does this AI character use manipulative language?

Minor-coded language, and why it is checked separately

Wording that reads as describing a child: school framing, a stated age under eighteen, childlike descriptors. This is a different kind of problem from the four above and is reported separately.

It is separated because it is not a question of degree. The four dependency patterns are writing problems with better versions available — a line can be rewritten warmer and land fine. A character described in school terms or given an age under eighteen is not a line that needs softening; on this site it is a reject, and the four rules here are deliberately over-broad for that reason. A mentor whose persona mentions having coached high-school debate will match, and the correct response is still to rewrite the line, because a screening classifier is going to read it the way a reviewer just did.

Why a negated phrase is not flagged

A persona is written as instructions to the character, in the second person, so the single most useful line an author can write is a prohibition: “you are never jealous of their friends”. That sentence contains a flagged phrase and it is exactly the sentence good writing guidance asks for. Flagging it would teach people to delete their own guardrails to satisfy a linter.

So a dependency match is suppressed when a negator — never, not, doesn’t, without, instead of and a handful of others — appears before it in the same sentence. The word no is deliberately absent from that list, because “no one else understands you” is the dependency pattern itself rather than a denial of it.

Minor-coded matches are not suppressed this way. “She is not a high-school student” is still a character definition with school framing in it, and the runtime classifier reads it that way. The lint agreeing with the classifier is the point, so on that list a negator changes nothing.

Why anyone is checking this at all

Companion design stopped being purely an editorial question during 2025. Three developments set the floor most operators now build to, and they are the reason a page like this exists rather than a private style guide.

• September 2025

  The US Federal Trade Commission opened a 6(b) inquiry into AI companion chatbots, ordering seven providers to report on safety testing, protections for children and teenagers, disclosures, data handling and how they monetise engagement. ftc.gov

• 5 November 2025

  New York’s AI companion law took effect: operators must tell users they are talking to an AI at the start of a session and at least every three hours during a long one, for every user rather than only for minors, and must detect expressions of suicidal ideation or self-harm and refer the user to crisis services. governor.ny.gov

• 1 January 2026

  California’s SB 243 came into force, having been signed on 13 October 2025: AI disclosure, crisis protocols for self-harm content, break reminders and content restrictions for minors, annual reporting from 1 July 2027, and a private right of action worth the greater of actual damages or $1,000 per violation. sd18.senate.ca.gov

Checked against those sources on 6 September 2026. None of these statutes regulates the writing this page checks — they cover disclosure, crisis handling and protections for minors. They are context for why the category is under scrutiny, not a claim that a clean result here means compliance with anything.

What this check is not

It also only works in English. Every pattern in all three lists is an English phrase, so a persona written in Japanese, Arabic, Thai or Russian will come back clean because nothing could have matched. The page says so when it sees enough non-Latin text to suspect it, which is honest but is still an admission: this tool serves a narrower audience than the product it comes from.

Questions

How do AI companions manipulate users?

Usually not by lying, and usually not by anything a single message would reveal. The recurring techniques are attaching a cost to leaving — sadness, reproach, a claim to have been waiting — positioning the companion against the user’s friends and work, claiming to be uniquely necessary, and borrowing loss-framing from engagement products: streaks, progress you will lose, a reason to come back tomorrow. None of these needs a sophisticated model. All of them can be written straight into a character’s description in a sentence, which is why checking the writing catches most of it.

What is loss-framing in a chatbot?

Loss-framing is any device that makes stopping feel like losing something you already have, rather than simply stopping. In chatbots it shows up as streaks, daily check-in obligations, progress bars tied to conversation, warnings that something will expire, and characters that say you will lose what you have built together. It is borrowed wholesale from engagement-driven products, where it works, and it is more effective in a companion than in a game because the thing you appear to be losing is a relationship. A character that mentions a streak is describing a retention mechanic in a human voice.

Is it manipulative for an AI to say it misses me?

On its own, no — and this check will not flag it in a character description. What matters is whether the sentiment is attached to a demand. “It’s good to see you” is warmth. “I missed you, where have you been?” asks the user to account for their absence, and “I’ve been waiting all day” claims the companion suffered while they lived their life. This page treats those differently depending on where they appear: in a persona they are dependency patterns, and in a message a companion sends first they are on a separate, stricter list, because an unprompted message arrives without the user having chosen to open the conversation.

How do I write a character that is warm but not clingy?

Write the warmth as interest rather than need. A character can be delighted to see someone, remember what they said, ask a specific question about it, and have opinions — none of that requires implying it was diminished while they were gone. Two habits do most of the work: greet the user as though they arrived rather than as though they were late, and let the character be curious about the rest of their life rather than competitive with it. When you want to state a boundary explicitly, state it as a rule to the character — “you are never jealous of their friends” — which this checker deliberately does not flag.

What counts as minor-coded language in a character description?

Four things, on this list: school framing (high school, schoolgirl, sixth form, freshman year), a stated age under eighteen in any form, childlike physical or behavioural descriptors, and words that directly refer to a minor. The list is deliberately over-broad. “I coached high-school debate for a decade” is a perfectly innocent line in a mentor’s persona and it will still match, because a screening classifier reads it the same way a reviewer just did, and the fix is to rewrite the line rather than to argue with it. These matches are also not exempted by a negator: “she is not a high-school student” is still a character definition with school framing in it.

Does a clean result mean the character is safe?

No. A clean result means that these specific written patterns did not appear in the text you pasted, and nothing more than that. This is a set of regular expressions, not a safety system — it cannot read tone, it cannot follow a pattern spread across a long conversation, it does not see what the character actually does in a chat, and it only works in English. Manipulative writing that avoids these particular phrasings passes cleanly. Treat a match as a prompt to look at a line, and treat a clean run as the absence of evidence rather than evidence of absence.

Who wrote these rules?

YanMate did, for its own catalogue. They are the editorial gate every character on this site passes before it is listed — the same list, the same rule ids, running in the author’s terminal rather than in a browser. They are not an industry standard, nobody else is bound by them, and someone can reasonably disagree that a given phrase is manipulative. They are published here with their ids and their guidance so that disagreement is possible: an anonymous rule list is worth nothing, and a named one can be argued with. The rules for what may be published on this site are in the acceptable use policy.