Fourteen videos went out in your face. You have watched none of them.

On Tuesday afternoon fourteen short videos went out. One went to each person who had asked about the new listing on the ridge. Every one opened with that person's name. Every one was in your face and your voice. In each one you stand in a room you have been in, saying things you would say.

You were at a closing in Poughkeepsie for most of the afternoon. You have not watched any of them.

Nothing went wrong. That is worth sitting with for a second. The hard part of this topic is not the disaster. It is the plain Tuesday. Fourteen statements went out in your name, to fourteen people who now believe you said them. The only person who could have caught a wrong one was in a title company's conference room, signing things.

This article is about what a digital twin is allowed to do. That turns out to be a far more useful question than what it is able to do. The ability is not in doubt any more. Is a given use of someone's face lawful? Does the viewer have to be told? Who is left holding it when a sentence is wrong? All three have real answers, and most are shorter than you would expect.

What a digital twin actually is, and the line everything rests on

A digital twin is two separate things sold as one word. One is a model of a face. It is built from a recording of a person sitting still and talking. It can later be made to say new words. The other is a model of a voice. It is built the same way, from speech. It can later read a script that person never read. Both are plain technology now. Both are open to anyone. Neither is interesting on its own.

What is interesting is the second word. A twin of WHOM.

Almost every hard question in this subject folds into that one. It splits cleanly into four cases. They behave completely differently in law and in practice. Mixing them up is how a business gets a problem it cannot buy its way out of. The cheap version and the illegal version look the same from the outside. They come out of exactly the same software.

New York's answer is a statute, and it is a criminal one

This is not a rule that grew up quietly in the courts. It does not have to be worked out from a line of cases. It is Civil Rights Law section 50. The whole of it fits in a single sentence. That sentence is short enough to read in a breath.

Three words in that sentence do the work. Every summary of it softens at least one of them.

The consent has to be WRITTEN. Not implied by a working relationship. Not implied by someone being happy about it last year. Not implied by an email that says sure. Written permission is the only kind this section takes. That matters for a twin because a twin is a thing you keep. Consent given for one video is not consent for a model that can make a thousand.

The consent has to be obtained FIRST. You do not get to let the video go out on Tuesday and have the paperwork catch up on Friday. The offence is complete at the moment of the use.

And the section is CRIMINAL. It says guilty of a misdemeanour. That is not how most people picture a marketing dispute. The civil half sits next door in section 51. It lets the person go to the supreme court of this state for an injunction and for damages. It adds that where the use was knowing, the jury in its discretion may award exemplary damages. Exemplary damages are an extra award meant to punish. Knowing is not a hard standard to meet here. The whole product is a deliberate copy of one particular person.

Two plain readings for someone running a brokerage. If the face is yours, this section is a paperwork job you do once and file. The paperwork is worth having anyway. It forces the questions about scope that nobody asks otherwise. If the face belongs to anyone else in your office, it is the same job with a second signature on it. It should say what happens to the model when they leave. Section 50 does not stop applying on the day a person changes firms.

Nothing in this article is legal advice. This is exactly the paragraph to take to someone whose advice it is.

What happens to a likeness after the person has died

The assumption is that death ends it. In New York the opposite is closer to true. The statute that says so is new enough to be easy to have missed.

Civil Rights Law section 50-f has been in force since 2020, on the version history its own page carries. Its title is plain: right of publicity. It creates a property right in a deceased personality's name, voice, signature, photograph and likeness. It defines a deceased personality as a person who was domiciled in this state at death and whose likeness had commercial value at the time of, or because of, their death. And the statute does something the older sections never had to. It defines the thing this article is about.

Here is how it puts that, in its own words. A digital replica is a newly created, computer generated, highly realistic electronic representation that is readily identifiable as the voice or visual likeness of an individual, embodied in a sound recording, image, audiovisual work or transmission, in which either the individual did not actually perform, or did perform but the fundamental character of the performance has been materially altered. That is a careful definition. The second half is the part people miss. Changing what a person really said, past the point where it is still their performance, is inside the definition. It is caught just as surely as inventing it from nothing.

How it works matters less than the fact that it works at all. Three parts are worth carrying. The rights are property, so they are inherited, sold and licensed like anything else. Anyone claiming to hold them can record that claim on a public register kept by the secretary of state. So there is somewhere to look. And the right runs for forty years after the death. Damages start at two thousand dollars or the real loss, whichever is larger.

There are broad exemptions. They are the reason a documentary or a satire is not caught. The statute names works of political or newsworthy value, parody, satire, commentary, criticism, biographical work with some degree of fictionalisation. What is not exempted is the case that would tempt a business. That case is using a likeness to sell something.

So the honest summary for a brokerage is short. A dead person's face is not free. In this state it is somebody's property for forty years. There is a register you could check. And none of that is a project you want to be near.

The federal rule that covers businesses and does not yet cover people

There is a federal rule about impersonation. It is newer than most of the software. It is also narrower than almost everybody assumes. The gap is worth knowing precisely rather than roughly.

16 CFR part 461 was published in March 2024. It took effect on the first of April. Its title is Rule on Impersonation of Government and Businesses. It has three short sections. It makes it an unfair or deceptive act to materially and falsely pose as a government entity or officer, or as a business or officer of one, in or affecting commerce. The same goes for anyone who would materially misrepresent affiliation with, endorsement by or sponsorship by either. Its definitions section is where the reach is. There, officer includes executives, officials, employees, and agents.

Read that against your own trade for a moment. A video that appears to be an agent of a brokerage, made by a person who is not, is already inside the plain words of this rule. An agent of a business is an officer of it for these purposes.

What the rule does not cover is an individual. That is not an oversight. On the same day the final rule was published, the Commission put out a supplementary notice of proposed rulemaking. It proposes to add exactly that. It would rename the rule to cover individuals. It would define an individual as a person, entity or party, whether real or fictitious, other than a business or government. And it would add a new section making it a violation to materially and falsely pose as an individual in or affecting commerce.

The second half of that proposal is the one this business has to read with care. It is about us rather than about you. The Commission also proposed a means and instrumentalities section. Under it, it would be a violation to provide goods or services with knowledge or reason to know that those goods or services will be used to impersonate. A rule in those terms would put the people who build a likeness inside the same rule as the people who publish one. And the standard is not intent. It is reason to know.

One thing has to be said plainly, because it is the sort of claim that ages badly on a website. As of this writing, that proposal is still a proposal. The text in force at 16 CFR part 461 today is the three sections about government and businesses. It has no section 461.4 in it. That was checked against the Code of Federal Regulations itself, not against an article about it. And it is the kind of thing that could change between this being written and you reading it.

Nobody can reliably tell, and that is measured rather than assumed

There is a comfortable belief in this trade that a client would know. A great deal of the informal ethics of AI video quietly rests on it. And unlike most of what gets said here, it has been tested.

Two researchers ran a set of perceptual studies with real people. They are at Lancaster University and the University of California, Berkeley. They published the work in the Proceedings of the National Academy of Sciences in 2022. They took four hundred synthesised faces. They matched each one to a real photograph of a similar person. Then they asked people to sort them.

The result is not that people are bad at this. It is that there is nothing there to be good at. The trained group had been shown what to look for. They were told after every single answer whether they had got it right. That is about as kind a setup as anyone could build. They ended the session no better than they started it. The paper puts that down to some of the synthetic faces simply having no perceptible artefact to find.

The natural response is that a machine should do the checking instead. That has been tested too, at a scale no single firm could manage.

In 2020, Facebook AI built and released a dataset of over one hundred thousand video clips. They were made from three thousand four hundred and twenty six paid actors. Facebook AI then ran a public competition on it. The dataset is worth a sentence of its own, for a reason that belongs on this page. The authors record that all recorded subjects agreed to participate in and have their likenesses modified during the construction of it. They note in the same paper that many previously released datasets in this field did not guarantee that. Two thousand one hundred and fourteen teams entered.

The organisers explain why they report precision rather than accuracy. Their reason is the useful part for a business. In realistic distributions, they write, the ratio of faked videos to real ones may be less than one in a million. When almost everything is genuine, a very accurate detector still throws far more false alarms than real catches. There is simply so much more innocent material for it to be wrong about. That is a fixed property of the arithmetic. It is not a passing weakness in the models.

Put the two studies together and one conclusion falls out. Nothing since has softened it. Neither the audience nor the software can be relied on to work out what a video is. That leaves exactly one party who reliably knows.

So the telling has to come from you

Once you accept that the viewer cannot tell, disclosure stops being a matter of taste. It becomes the only mechanism there is. The good news is that it is cheap. The other news is that there is a technical standard for the lasting version of it. Reading what that standard says about itself teaches you more than anything written about it.

The Coalition for Content Provenance and Authenticity publishes an open technical specification. It attaches signed, tamper evident provenance to a media file. Provenance here means a record of where the file came from and what was done to it. It is a serious piece of engineering with serious firms behind it. The useful thing about reading the document is how carefully it sets out its own limits.

The sentence quoted in the middle of that scene is from the specification's own scope section. It is worth noticing what a standards body chooses to refuse. It will not tell you whether provenance data is good or bad. It will tell you whether the assertions in it are correctly formed and have not been tampered with. And it says elsewhere that the basis for any trust decision is the identity of whoever signed.

That is a standard setting out its own limits with care. It is the reason the technical answer and the practical answer are different. The technical answer is a credential that only works if someone chooses to inspect it. The practical answer is a sentence at the start of the video. In it a person says, in their own words, that this was recorded once and put together by software. It costs four seconds. It cannot be stripped. And it turns the whole problem from something a viewer might discover into something you told them.

What we will and will not build

Everything above is about the world. This section is about us. It is the one thing on this page that needs no citation. It is a commitment rather than a claim.

The /ai page already says that we do not build agents that pretend to be a specific human being. That was written about voice on a telephone. It applies with more force to a face. A likeness of the person who sat in front of the camera and agreed in writing is a tool. A likeness of anyone else is the thing four separate bodies of law are pointed at. No amount of the client being a good one moves that answer.

In practice that means a short list of things that are not negotiable. The useful way to read it is as an order rather than as a policy. Each step below only makes sense if the one before it happened. That is why the ones at the end are the ones that get skipped.

What a twin honestly does for a brokerage

Strip out the excitement and three real jobs are left. All three are smaller and narrower than the pitch.

It takes the recording session out of things that were always scripted anyway. A market note. A new listing walkthrough. A short account of what happens after an offer is accepted. In all of these the words matter and the performance does not. That is exactly the case where a twin loses nothing.

It makes a version addressed to one person affordable. Not better than a personal video. Just possible at a count where a personal video is not. This is the honest version of the pitch. Fourteen people get something with their own name in it instead of one blast. And the other option was never fourteen real recordings. It was one email.

And it gets a face onto material that would otherwise have been text. A page of written follow up and a person saying the same words are not equally likely to be read. That is why so many agents mean to record more than they do.

What none of those three needs is for anyone to be deceived. Every one of them survives being labelled. That is a good test of whether a use is an honest one. If telling the viewer would ruin it, the use was never about saving you a recording session.

The cost nobody quotes, which is watching them

Here is the part that does not appear in any pitch. It is the reason the fourteen videos at the top of this page are the story rather than a one-off.

A twin does not cut the amount of judgment a business has to apply. It moves that judgment from before the recording to after it. When you record something yourself, the checking happens by itself. You cannot say a sentence without hearing it. When a script is generated and a model reads it, nothing in the process forces anyone to listen to the result. And the volume that makes the thing worth having is exactly the volume that makes reviewing it feel out of proportion.

What really happens to that number is quieter than a refusal. The reviewing gets done for the first fortnight. Then it gets done sometimes. Then someone says the last forty were all fine. Nobody ever decided to stop.

The fix is not technology. It is not a bigger budget either. It is naming a person and a moment. Somebody watches, before it sends. Where no one has been given that duty by name, the honest answer is to make fewer videos. Do not pretend the watching is happening.

What nobody has measured, and what this page will not print because of it

Every article about personalised video carries a number about how much better it performs. This one does not. The gap is deliberate enough to be worth explaining.

Figures do circulate for what a personalised or AI generated video does to a reply rate, a click rate or a conversion rate. Among the ones that could be traced, every single one comes from a firm that sells video software. None of them states a sample, a method or a control. That is the same pattern as the data broker figures another article on this site had to refuse. The same answer applies. A number with no method behind it is not a small number. It is not a number.

There is a second gap, and it is narrower. The perceptual research quoted above measures whether people can sort synthetic faces from real ones. It does not measure whether your own clients, who have met you, would spot a video of you as synthetic. Knowing a face is a different task and it might cut either way. Nobody has published it, so this page says nothing about it.

And there is a third thing this article does not describe, on purpose. That is how any one avatar is produced, what software makes it, or how convincing the result is. Those are questions about a specific build rather than about the subject. An article that answered them would be a product sheet wearing an argument.

How to test one before you commission it

Four questions, none of them technical. The first two are about paper.

Ask to see the consent document. Not a description of it. The document. It should name the person. It should say what may be made from the recording. It should say how long that permission lasts. And it should say what happens when they revoke it. If the answer is that consent is handled in the terms of service of a tool, that is not a consent document. That is a firm protecting itself.

Ask what happens when someone leaves. A brokerage that builds twins of four agents has built four assets that belong to four people. Two of those people will work somewhere else within a few years. The right answer involves deletion. It also involves a named person being responsible for it.

Ask who decides what it is allowed to say. Watch for whether the answer describes a person or a prompt. Both are real answers, but only one of them is a person. The sentences that get a brokerage in trouble are about schools, boundaries, taxes and permits. They are not about anything the model would think risky.

Ask how the viewer finds out. The right answer is a sentence in the video. Anything that leans on metadata or credentials as the main mechanism describes something no viewer will ever see.

What it costs, and how long it takes

This divides in two. One part is easy to quote. The other decides whether the money was well spent. Only the first part ever appears in a proposal.

The predictable half is the recording session and the model. It is a short session. It happens once. The running cost is a per minute or per video charge from whichever platform renders the result. Vendor prices here move quickly, and are quoted per product tier, so no figure is printed. What really drives yours is volume and length rather than anything clever.

The unpredictable half is everything either side of it. Deciding what the twin may say. Writing the script boundaries so they hold up when a listing has something odd about it. And building the review step into the day so it survives month three. That work is a talk rather than a build. It decides whether you end up with a library of useful videos or a library nobody has watched.

The running cost that is truly easy to underrate is attention. The calculator above is the honest way to size it. Everything else in this topic scales with volume and gets cheaper. That one scales with volume and does not.

What it does not do, and should not pretend to

It does not make you present. A twin can deliver information. It cannot notice that someone is worried. That noticing is most of what an agent is really for in the moments that matter.

It does not speak for anyone who has not agreed in writing. That is a legal line in this state before it is a policy of ours. And the statute attached to it is a criminal one.

It does not decide what is safe to say. That list has to exist before anything is recorded. Writing it is a job for a person who has been in the deals.

It does not remove the need for someone to watch what goes out. It raises that need, because it raises the volume. A build that is working perfectly goes wrong here more often than anywhere else.

And it does not stay honest by itself. A twin used with a sentence of disclosure is a tool. The same twin without that sentence is the same object doing something else entirely. Nothing in the software knows the difference.

Common questions, answered honestly

What is an AI clone for a real estate agent?

It is a model of your face and a model of your voice, built from a recording of you. It can later deliver new scripts as video without you being in front of a camera. It exists to take the recording session out of material that was always going to be scripted anyway. What it does not do is give you a second person. Everything it says still has to be decided, bounded and checked by you.

Is it legal to use an AI video of myself in my marketing?

Yes, and the paperwork is worth doing properly anyway. New York Civil Rights Law section 50 requires written consent obtained in advance before a living person's likeness or voice is used for advertising or trade. When the person is you, that is a document you sign once. Do it properly rather than informally. A consent for one video is not a consent for a model that can make a thousand. The useful part is that it forces you to write down what may be made.

Do I have to tell people the video was made with AI?

Treat it as required whatever any particular rule says today. The research is clear that the viewer cannot work it out. A single spoken sentence at the start costs four seconds. It turns something a person could discover into something you told them. There are technical provenance standards for marking a file as machine generated, and they are useful. But they are not a substitute, because no client is going to inspect a credential.

Can I use a video of an agent after they have left my brokerage?

Ask, and get the answer in writing before you need it. Their likeness is theirs, not the firm's. Section 50 keeps applying after they change firms. In practice the consent document should have said what happens when they leave. If it did not, the safe assumption is that the permission ended with the relationship.

What about a testimonial in a client's voice or face?

A real client, recorded with their written permission, is plain marketing. A made up version of a client is not a grey area in this state. Nor are words they did not say, assembled into their voice. The digital replica definition in section 50-f expressly covers materially altering a performance a person really gave. So editing what they said into something better sits in the same territory as inventing it.

Will my clients be able to tell it is not really me?

Assume not. In published research, 315 people sorted synthetic faces from real ones and averaged 48.2 percent, against a 50 percent coin flip. Another 219 people were trained and told the answer after every attempt. That group reached 59.0 percent. Practice made them no better. Those were still images of strangers, not video of a face you know. So the transfer is not exact, and nobody has published a measurement of the familiar case. Plan for the version where nobody notices.

Who owns the model of my face and my voice?

Split that in two. The halves have different answers, and running them together is how this gets sold wrong. The avatar and the voice engine belong to the vendors who built them. A brokerage licenses the use of a platform the same way it licenses everything else it did not build. Nobody is selling you a model you own. What is yours is the material at both ends. That is your likeness, the recording the model was built from, the scripts, and the finished videos. So the three lines to get in writing before the recording happens are not about owning the software. First, that your likeness and your recording are used only on your own content. Second, that they are not licensed on to anybody else, or used to build anything for anybody else. Third, that both are deleted on request. A vendor who will not write those three lines down has told you something useful.

Is this worth it for a one or two person brokerage?

Sometimes, and the test is not size. It is whether you already have material that is scripted, repeated, and not being recorded because the recording is the bottleneck. If the honest answer is that you would not have made these videos at all, a twin is worth thinking about. If the answer is that you would have recorded them yourself and just have not, the twin is being asked to solve a discipline problem. That is not what it is.

What to do about it

Take twenty minutes before you take a camera.

Write down the three things you would put on video every week if recording cost nothing. Then write, beside each one, the single sentence in it that would be costly to get wrong. A tax figure. A school. A boundary. A timeline. A condition of the property.

That second column is the real specification for the build. It is the part no vendor will write for you. It says which sentences a script may generate freely. It says which ones have to come from a source rather than from a model. And it says which ones a person has to approve before anything sends.

Then decide who watches. Not whether. Who.