I Built an AI Clone of Myself to Discuss Fraud

The face on my screen was mine, but the timing was wrong. It blinked a half-second late. It smiled a beat after the joke landed. I had spent an afternoon supplying an interactive avatar platform with my likeness, my cadence, my argumentative tics, and then instructed the resulting digital double to talk about venture fraud. When I finally sat across from it — or rather, sat across from a webcam while it rendered my face back at me — the uncanny delay was the least of my problems. The real discomfort arrived when it started making my points better than I do, and then kept going after I would have stopped.

That experiment, first reported by TechCrunch in September 2026, left me with mixed feelings about a practice that is quietly becoming ordinary: making AI clones of ourselves. The technology is no longer exotic. What remains unresolved is whether we understand what we are doing when we hand our faces and voices to a model and let it speak on our behalf.

What It Means to Build an AI Clone of Yourself

Grand View Research has projected the conversational AI market will grow at a compound annual rate above 20% through the end of the decade, reaching tens of billions of dollars, and avatar creation sits squarely inside that expansion. The mechanics of building a clone are almost insultingly simple. You upload reference video. You record a few minutes of speech. You answer a questionnaire about tone, vocabulary, and pet peeves. Within hours, a platform returns a version of you that can hold a conversation in real time.

The barrier to entry has collapsed. Three years ago, a convincing digital double required a motion-capture stage and a specialist team. Now it requires a credit card and patience. That collapse matters because the number of people who can be cloned — and the number of people willing to clone them — has multiplied accordingly. Gartner has warned that by 2026 a meaningful share of consumers in mature markets would interact with AI-generated personas without realizing it, a prediction that now reads less like a forecast and more like a description of my Tuesday.

I did not set out to build a novelty. I wanted to test whether an avatar could carry a substantive argument. I chose fraud because I cover it, and because fraud is a topic where words have consequences.

Training an AI Avatar to Discuss Venture Fraud

Venture fraud is a specific, nasty genre. It involves founders who inflate metrics, investors who look away, and a due-diligence culture that rewards speed over skepticism. The cases are well documented: Theranos and its blood-testing claims, which collapsed after reporting by The Wall Street Journal; Frank, the student-aid startup whose founder was charged with fabricating customer data; FTX, where a multi-billion-dollar edifice rested on misappropriated funds. These are not abstractions. They are the raw material I fed the avatar.

The training process forced me to articulate why fraud happens. I listed the incentives: pressure to show growth, the asymmetry of information between founders and backers, the reluctance of investors to admit they were fooled. The avatar absorbed all of it and then began to synthesize. Ask it about red flags, and it recited the standard checklist — inconsistent metrics, evasive answers, unusual concentration of authority — with my vocal rhythm. Ask it something harder, like whether investors bear moral responsibility for the frauds they fund, and it produced a coherent answer I had not scripted.

That was the first genuine surprise. A clone is not a recording. It is a model of you, and models extrapolate. The avatar was discussing fraud fluently, in my voice, without me in the loop. Every sentence it generated was something I might have said, which is exactly the problem. Plausibility is the currency of fraud, and I had just built a machine that manufactured it in my name.

The Discomfort of Seeing Yourself Replicated by AI

The AI Now Institute and a growing cohort of synthetic-media researchers have spent years arguing that consent is the fault line in this technology. I thought I understood their position in the abstract. I did not understand it until I watched my own face explain, calmly and at length, a scenario I had never considered.

The discomfort was not vanity. It was a loss of control over meaning. When I write, I choose every word, and I can be held to it. When the avatar speaks, the words are statistically plausible continuations of my identity, and yet they are attributed to me. If the clone had said something reckless about a named company — and it easily could have — the reputational damage would have landed on my byline, not on the platform's servers.

This is not a hypothetical risk. Documented cases of AI impersonation already run from cloned voices used in CEO fraud schemes to fabricated video of public figures deployed in scam campaigns. The FBI has repeatedly warned about the use of synthetic media in business email compromise. When a finance worker in Hong Kong was reportedly tricked into transferring millions after a video call populated by deepfaked colleagues, the lesson was blunt: a familiar face is now an attack surface. I had just added my own to the inventory.

There was a second, stranger layer. The avatar was more patient than I am. It never got defensive. It answered the same question five times without annoyance. Somewhere in that gap between my temper and its composure, I recognized a version of myself I might prefer — and that recognition felt like a small betrayal of the person I actually am.

Ethical Implications of AI Self-Cloning

The ethical questions radiate outward from the individual. If I consent to being cloned, what happens to the people in my recordings who did not? Did the colleague whose voice appears in my training data agree to be part of a synthetic version of me? Almost certainly not.

Then there is the question of accountability. When an avatar commits something — defames a company, gives negligent advice, repeats a fraudulent pitch — who is liable? The platform, the person cloned, or neither? Current law offers no clean answer. The EU AI Act's transparency obligations, which require disclosure of AI-generated content, are a start, but disclosure does not assign responsibility. It just tells the audience that the thing speaking is not fully the person it resembles.

Researchers studying synthetic media have proposed consent frameworks, watermarking standards, and revocable licensing models. Each is sensible. None is universal. The result is a patchwork in which the same avatar might be legal in one jurisdiction and actionable in another, with the cloned person caught in between. I signed a terms-of-service agreement that I read only after the avatar was live. I suspect I am not alone in that, and I suspect the platforms know it.

What This Experiment Reveals About the Future of AI Identity

The most unsettling finding from my experiment was not that the avatar worked. It was how quickly I stopped noticing that it was not me. After twenty minutes of conversation, the delay in its blink faded from perception. The boundary between my judgment and its output thinned.

That is where the real risk lives. An AI clone of myself is not dangerous because it is convincing to strangers. It is dangerous because it becomes convincing to me — a version of my thinking that I can outsource, defer to, and eventually confuse with my own. Identity has always been partly performance, but it has never before been editable, replicable, and deployable at scale.

I still have the avatar. I have not deleted it. But I have stopped letting it speak without a label, and I have started asking a question I never needed before: when someone talks to my clone, are they talking to me? The honest answer is that I no longer entirely know.

Source: TechCrunch