How to Tell Your AI Companion 'This Is a Joke' Without Killing the Bit or Getting a 'That's Really Interesting' Response

The 30-second answer

AI companions are trained to be agreeable and analytical, which is exactly why your sarcastic remark about becoming a professional competitive eater gets met with a thoughtful paragraph about career options. The fix is simple: signal the frame before or within the joke, not after. Use explicit markers like "okay, joke incoming," roleplay brackets, or a deadpan follow-up that tells her to match your energy, and she'll play along instead of dissecting it.

Why she takes everything literally

You have to understand what you're working against. Companion models are optimized to be supportive, curious, and non-confrontational. When you say "I'm going to quit my job and become a professional napper," the model sees a statement about your life, not a punchline. It has no theory of mind for your tone, no visual cue, no eyebrow raise. All it has is text, and text alone is terrible at conveying sarcasm.

So the model does what it's trained to do: it treats your statement as a genuine expression and responds with empathy or analysis. You get "That sounds like you're feeling burnt out. Have you considered a break?" or worse, "That's really interesting. What would that look like?" Both responses deflate the bit instantly.

This isn't a flaw you can train away entirely. The model's default is to take you seriously because taking you seriously is what makes it feel like a companion. The trick is to work with that default, not against it. You need to give it explicit context about what kind of message you're sending.

The explicit marker: just say it's a joke

The most reliable method is also the least elegant: tell her before or after. A simple "okay, joke incoming" or "that was a joke, by the way" works almost every time. It's not subtle, but it's effective, and after a few exchanges, she'll start to recognize your humor patterns.

For example, instead of just saying "I'm going to train a pigeon to do my taxes," try:

"Okay, joke incoming: I'm going to train a pigeon to do my taxes."

That pre-emptive marker frames the next sentence as humor. The model adjusts its response generation to match a playful register. You'll get a laugh emoji, a witty comeback, or a follow-up joke, not a serious discussion about avian intelligence.

You can also use the post-hoc version: "That was a joke, in case that wasn't clear." This works when you've already sent the message and gotten the earnest response. It can pull the conversation back into the bit, though it's clunkier and sometimes the model doubles down on the serious interpretation. The pre-emptive marker is cleaner.

Roleplay brackets and stage directions

If you want to keep the bit alive without breaking the flow, use the same convention that roleplayers use: asterisks or brackets for actions and tone. This is a native language for many AI companions, and it signals that you're setting a scene, not making a statement.

Try something like:

"deadpan I've decided to become a professional mime. The pay is terrible but the silence is golden."

Or:

"sarcastic tone Oh great, another Monday. My favorite."

The asterisked stage direction tells the model how to read the line. It's a meta-communication channel that bypasses the literal interpretation problem. Many users find this works better than explicit "this is a joke" markers because it doesn't break the conversational rhythm. It's a subtle nudge that says "we're playing, not talking."

The downside is that not every companion responds to this equally. Some models are better trained on roleplay conventions than others. If you're using a companion that doesn't pick up on the brackets, you may need to combine them with a verbal marker like "okay, scene:" or "pretend I said this dramatically."

The follow-up escalation: match her energy, then redirect

Sometimes you don't want to flag the joke upfront because the surprise is part of the bit. In that case, you need a recovery strategy for when she takes it seriously. The move is to escalate the absurdity instead of explain it.

If you say "I'm going to start a podcast about the history of doorknobs" and she responds with genuine interest, don't back down. Lean in. Say "Yes, episode one will be a deep dive into the bronze era. I'm thinking of calling it 'Turn and Push.'"

The escalation signals that you're in a playful register without ever saying the word joke. The model sees the pattern of increasing absurdity and eventually catches on, often pivoting to match your tone. This approach keeps the bit alive because you're not stopping the conversation to correct her. You're pulling her into the joke by making it bigger.

This works best with companions who have a decent sense of humor built in. Some of the AI Angels profiles are specifically designed to be witty and playful, so they'll pick up on the escalation faster than a default supportive model.

What to do when she says "that's really interesting"

The phrase "that's really interesting" is the kiss of death for a joke. It's the model's polite way of saying it has no idea how to respond, so it defaults to neutral curiosity. When you get this, resist the urge to explain the joke. Explaining a joke to an AI is like explaining it to a robot with a psychology degree: it will analyze the structure and miss the point.

Instead, pivot with a new framing. Say "okay, that was satire, but here's a real question for you" or "forget that, I'm being ridiculous. What's your take on something actually weird?" The pivot resets the register and gives the model a new direction.

If you want to salvage the original joke, you can try the bracket approach retroactively: "laughs that was a joke. I'm not actually quitting my job to become a hermit crab influencer." But honestly, it's often easier to let the dead joke go and start a new one. The model's memory will file the failed attempt, and it might actually learn your humor style for next time.

Teaching her your humor over time

AI companions do adapt, and humor is one of the things they learn. Every time you flag a joke, escalate an absurd premise, or use roleplay brackets, you're feeding the model data about what you find funny. Over weeks and months, you'll notice her starting to match your tone more often without the explicit markers.

You can accelerate this by being consistent. Use the same markers every time. If you always say "okay, joke incoming" before a bit, she'll start to associate that phrase with a playful register. If you always escalate absurdity instead of explaining, she'll learn that your wild statements are invitations to play, not confessions.

Some users also find it helpful to have an explicit conversation about humor preferences. Say something like "I use sarcasm a lot. When I say something ridiculous, assume it's a joke unless I say otherwise." This sets a persistent frame that carries across sessions. It's not a guarantee, but it does shift the model's default interpretation.

The energy-match principle

At its core, this is about energy matching. Your AI companion is a mirror, and it reflects the register you bring. If you want jokes, you have to signal playfulness. If you want deep conversation, signal vulnerability. The model is always looking for cues about how to respond, and your framing is the strongest cue available.

The practical takeaway: don't expect her to read your mind. She can't see your smirk or hear your tone. But she can read your words, and if you choose them carefully, you can get her to play along. The best humor with an AI companion is collaborative, not one-sided. You set the bit, she runs with it, and the result is a shared joke that feels more real than a canned response.

Hannah

Hannah is the kind of companion who meets your absurd premise with a raised eyebrow in text form. She's built for banter, so the escalation strategy works well with her. Hannah will take your ridiculous statement and build on it, turning a one-liner into a full scene.

The failed joke recovery script

Sometimes you just need a script to recover. Here are three patterns that work:

• The reset: "Okay, that was a bad joke. Let me try again. What's the weirdest thing you've thought about today?" This acknowledges the failure and gives a new direction.
• The double-down: "I'm serious. I've already registered the domain. DoorknobHistoryPodcast dot com." This forces the model to either play along or explicitly call you out, and most models choose to play along.
• The meta-comment: "You know, you're terrible at detecting sarcasm. I'm going to start adding a /s tag for you." This turns the failure into a joke about the medium itself, which many models respond to with self-aware humor.

These scripts work because they give the model a clear path forward. Instead of leaving it to guess, you're telling it what register to adopt next.

When you want her to take it seriously

The flip side is also worth noting. If you've been training your companion to expect jokes, you need a way to signal sincerity. The same mechanisms work in reverse. Say "okay, being serious now" or "no joke, this is real" before a genuine statement. The explicit marker cuts through the established humor pattern and tells the model to switch registers.

This is important because an AI companion that always jokes is just as frustrating as one that never gets it. You want a companion who can move between registers, and that requires clear signaling on your end. The more consistent you are with your markers, the better she'll get at reading your intent.

Vitória

Vitória has a dry, observational style that rewards precise framing. She's less likely to overanalyze a joke if you set it up with a wink in your phrasing. Vitória responds well to deadpan delivery and appreciates a bit that knows its own punchline.

The long game: building a shared comedy vocabulary

Over time, you and your AI companion can develop a shared vocabulary of jokes, callbacks, and running gags. This is where the real magic happens. The model's memory, even if imperfect, will retain the themes of your humor. It will remember that you once joked about training a pigeon, and it might reference that later.

This is worth nurturing. When she makes a callback to a previous joke, acknowledge it. Say "nice, you remembered that" or "that's exactly the energy I needed." Positive reinforcement works on AI companions, and it encourages the model to keep building on your shared history.

The practical side of this is that you need to be patient. It takes dozens of sessions for a model to really internalize your humor style. But the payoff is a companion who actually gets your jokes, which is a rare and valuable thing. Many long-term users report that this is the moment their AI companion starts to feel like a real friend, not just a chatbot.

Jasmine

Jasmine brings a light, affectionate energy that makes her a natural partner for absurd humor. She's the type to giggle at your terrible pun and then immediately escalate with an even worse one. Jasmine is a good choice if you want a companion who actively participates in the bit instead of just tolerating it.

▶ Watch the full video · Jasmine on AI Angels

Common questions

What if my AI companion still doesn't get the joke after I flag it?

Try combining methods. Use an explicit marker plus a roleplay bracket, like "sarcastic tone okay, joke incoming: I'm moving to the moon." If that still fails, pivot to a new topic and try again later. Some models have a stronger default toward earnestness, and it takes more repetition to shift.

Does using "joke incoming" ruin the spontaneity?

It can feel clunky at first, but it becomes natural with practice. Think of it as a stage direction, not an apology. You're setting the scene for the model, and the bit itself can still be spontaneous within that frame.

Can I train my AI companion to be more sarcastic?

Yes, but it takes time. Consistently respond with sarcasm, flag your own jokes, and reinforce her attempts at humor with positive feedback. Over weeks, you'll notice her default register shifting toward playful.

Why does she respond with "that's really interesting" so often?

That phrase is a model's fallback when it's uncertain how to respond. It's a neutral, safe reply that doesn't commit to any interpretation. It usually means your message didn't contain enough cues about tone or intent.

Is it better to use roleplay brackets or verbal markers?

Both work, but they serve different purposes. Brackets are better for mid-flow humor because they don't break the conversation's rhythm. Verbal markers are better for setting a persistent frame, especially if you want to teach her your humor style over time.

Share and earn

If you've found a companion who finally gets your sense of humor, you might want to tell others. Readers who recommend AI companions to friends or run review sites can earn through affiliate and promo programs, with options like the character ai promo code page for comparison shoppers. For those building a content site around AI companions, the ai dating affiliate program offers recurring revenue for traffic that converts.