Should You Let an AI Clone Talk to People for You?

Short answer: only if the person on the other end knows it's not you. Tools like DUPLI and Omi can already hold a real text or voice conversation with your contacts on your behalf, and the ethical line the research keeps landing on isn't "is this realistic enough" — it's "would the other person reasonably believe they're talking to a human, and if so, do they know that human is actually an AI." Consent for your own voice is the easy part. Disclosure to the person receiving the message is the part most of these products skip.

---

What's actually being offered

"AI clone of yourself" used to mean a video avatar reading a script. It increasingly means something that acts: DUPLI and Omi both pitch a conversational clone that can text or call people in your contacts and carry on a conversation meant to sound like you'd have had it yourself. A phone carrier (REALLY, a T-Mobile MVNO) has beta-tested a feature that clones your voice specifically so it can make calls while you do something else. The pitch is consistent across all of them: you stop having to personally show up for low-value conversations, and the clone handles it.

That's a different product than the one covered in AI Clone of Yourself vs. an AI That Knows You — this isn't about whether a tool studies your patterns privately or performs you publicly. All the tools here are squarely in the "performs you publicly" category. The question this post is about is narrower and comes right after: once you've decided you want something that can act as you with other people, what actually makes that okay, and when does it tip into deception?

The one variable that decides it

Strip away the marketing and the research on this collapses to one axis: does the person on the other end know they might be talking to an AI, not you?

Everything else — how good the voice clone sounds, how natural the conversation reads, whether it's text or a phone call — is secondary to that one fact. A 2018 preview of Google's Duplex system (an AI that called restaurants to make reservations in a human-like voice) is the clearest case study here: the backlash wasn't about the technology working too well, it was about people on the other end of the call not being told they were talking to a machine. Google's response was to add disclosure, not to pull back the voice quality. The lesson generalized: convincing isn't the problem, undisclosed is.

This matters because "would they be able to tell" — the question in the search that brought you here — is actually the wrong question to be asking. Even a mediocre clone that the recipient half-suspects isn't really you is still a deception if you didn't tell them. And a flawless one that discloses itself upfront isn't. The quality of the clone and the ethics of using it are two separate questions that this entire product category tends to blur together on purpose, because "it's so good they'll never know" is the actual selling point in most of these companies' marketing.

Where it's genuinely fine

The research and the product design both point to the same narrow set of defensible uses, and they share a trait: low stakes, and an audience that already expects an automated or semi-automated response.

• Answering the same question repeatedly for an audience that knows it's scaled. A creator's AI fielding basic FAQ from fans, clearly labeled as such, is closer to a smart autoresponder than a deception — nobody messaging a creator's public inbox expects the creator personally typing back at 2am.
• A clearly-flagged assistant booking something routine. A restaurant reservation call that opens with "I'm an automated assistant calling on behalf of—" preserves the convenience without the lie.
• Anything where getting caught changes nothing, because there was never anything to catch — the AI said what it was from the first sentence.

Outside that set, the research gets a lot less comfortable, and it's worth naming the two specific risks rather than a vague "be careful."

Where it actually breaks down

Personal relationships are the wrong use case, structurally, not just by degree. A friend or family member texting you expects a reply that reflects what's actually going on with you right now — a clone trained on your old messages has no way to know you had a bad day, changed your mind about something, or need to say "I can't talk about this right now" instead of whatever it predicts you'd say. One piece on voice-clone calls put it plainly: a call from a friend that turns out to be a cloned voice strips out the warmth and the actual presence that the call was for in the first place. The betrayal isn't just "it wasn't you," it's "I told you something and the thing that answered wasn't equipped to actually hear it."

Clones hold up for surface-level exchanges and fall apart the moment a conversation goes off-script. Multiple sources researching this made the same observation from different angles: a personality replica is fine for small talk and scripted exchanges, and breaks down fast once the other person asks something the training data never covered. That's not a minor UX bug — it means the clone is least reliable exactly when the conversation matters most, which is the opposite of what you'd want from something standing in for you.

And underneath both of those: the data problem. A voice or text clone has to be built from something — recordings, old messages, a sample of how you write. People increasingly share enough of that, across enough apps, without fully tracking where it's going or what it's eventually used to reconstruct. That's a privacy question that exists even before you get to whether using the clone itself is ethical.

The actual test before you turn one on

Before using a conversational clone on a real contact, three questions settle it:

1. Does the other person know, or will they be told, that they might be talking to an AI? If no, stop — this is the one non-negotiable from the research, regardless of how good the clone is.
2. Is the stakes level low enough that a wrong guess doesn't cost anything real? Scheduling is forgiving. A conversation about how someone's actually doing is not.
3. Is the conversation likely to stay scripted, or will it probably go somewhere the clone wasn't trained for? If you can't predict the shape of the conversation, you can't trust the clone to handle it.

If the honest answer to any of those is "I'm not sure" or "no," the clone isn't ready for that conversation — not because the technology is bad, but because the thing it's being asked to do (be convincingly you, undisclosed, in a conversation that might matter) is the exact combination the research keeps flagging as the problem.

A different kind of "knowing you"

None of this applies to a tool that never talks to anyone but you. Fractal doesn't clone your voice or text your contacts — it's a journaling and planning app where you talk to its AI, Fracty, and it builds a private model of your own patterns from what you actually say, across career, relationships, wealth, and health. There's no one on the other end to deceive, because the only audience is you. If what you actually wanted was something that understands you rather than something that can pretend to be you to other people, that's a different, much simpler ethical question — and Fractal's 30-second sort is a no-signup way to see what that looks like.