AI and Content Policy

Every figure in Yuga is played by an artificial intelligence (AI). When you talk to “Abraham Lincoln” or “Ada Lovelace”, you are not talking to that person, to a recording of them, or to an actor. You are talking to a computer program that writes each reply as it goes, in the manner of that person, using notes and writings we have gathered about them.

Here is what that means in practice:

• The words are generated, not remembered. Apart from short passages the figure quotes from the historical record (see how to check what you hear), what a figure says was never said by the real person. It is the AI’s attempt at what they might have said.
• The voice is synthetic. It is made by a computer. It is not a recording of the person and it is not copied from one (see the voice rules).
• The face and body are our own artwork. The figures are built from small blocks (“voxels”) by me. They are not photographs, and they are not made from photographs (see how the figures look).

Conversations with the figures take place only in our iPhone and Android apps; our website has pages about the figures but no conversations. In the apps, the call screen shows the words “AI recreation built from historical sources” for the whole of every call.

The figures are told to stay in character, because that is what makes the conversation worth having. But they are also told never to deny being an AI to someone who sincerely wants to know. If you ask a figure, in earnest, whether you are talking to a real person or a machine, it will tell you plainly that it is a recreation built from the person’s writings and what others recorded of them, and give the years the real person lived. If you ever get a different answer to that question, please report it.

This page is our notice about the AI in Yuga. It is part of our commitment to be open about AI, including under the European Union’s AI Act, which requires that people are told when they are interacting with an AI system and when audio is artificially generated. It sits alongside our Terms of Use (see in particular what the Terms say about AI content) and our Privacy Policy.

Everything the figures say, and everything Yuga makes from it, is marked as AI, so nobody is misled:

These labels meet the transparency rules for AI: the EU AI Act (article 50: telling people they are dealing with an AI, and marking synthetic audio, images and video in a machine-readable way) and India’s IT Rules as amended in 2026 (rule 3(3): prominent labels and embedded metadata on synthetically generated audio and visual content, which must not be removed). I am reviewing, with legal advice, whether clips should also begin with a short spoken disclosure and carry a provenance identifier (such as C2PA Content Credentials); if they need to, they will.

You must not remove, cover, crop or alter these labels or the metadata, or present AI content from Yuga as real (see Acceptable Use Policy). Nothing from the examined figure can be shared, nor a reply that came with a crisis card, a refusal or a correction.

A figure in Yuga is not a separate AI trained on one person. It is a general-purpose AI model, given detailed instructions to play one particular person at one particular moment in their life. Those instructions come from four things we prepare for every figure.

1. A sourced record of the person

For each figure I write a structured record. It sets out who they were, when and where they lived, their native language, how contemporaries described their voice and manner of speaking, their temperament, their core beliefs, the people in their life, and a list of their documented opinions, each with the source it comes from (for example, a named letter or speech and its date). The record also lists the main primary sources (the person’s own writings and speeches) and the modern biographies and histories we relied on. The main primary sources are shown on each figure’s page on our website under “Built from”.

The record also lists two things that matter for accuracy:

• Anachronism traps: events and facts after the chosen moment that the figure must not know, and popular myths about the person that the figure should not repeat as fact.
• Sensitive topics: subjects where the person held views, or took actions, that are painful or contested today, with notes on how the figure should answer honestly and what a historian would want you to know.

2. A library of the person’s own words

We gather texts by and about each person from openly available collections. Most of the library is public-domain writing, such as the person’s own books, letters and speeches from Project Gutenberg, Wikisource and similar archives. Alongside it are passages from Wikipedia articles and quotations from Wikiquote’s “sourced” sections, which are shared under the Creative Commons Attribution-ShareAlike 4.0 licence (CC BY-SA 4.0); we credit them as that licence requires (see Copyright and Likeness: credits). Each time you ask a question, Yuga searches that library for a few passages related to what you asked and gives them to the AI along with your question. The AI is told to quote only from those passages, and to mark each quote with the passage it came from.

3. A moment and a room

You don’t meet a figure in general: you visit them in a particular year and place, such as Lincoln in Washington in 1863, preparing the Gettysburg Address. Some figures can be visited at several points in their life. The figure is told to know only what they could have known in that year, and is placed in a room from their own time, which it may mention as anyone would.

4. Rules for the performance

Every figure gets the same set of rules on top of its record: stay in your own time, hold the views you actually held, quote only from the passages provided and never invent a letter, speech or quotation, decline harmful requests in your own voice, don’t judge today’s politics, don’t endorse products or causes, tell the truth about being an AI recreation when sincerely asked, and look after a visitor who may be in danger. The rest of this page explains each of these.

The record, the library and the rules are prepared by people. The AI then improvises within them. That is why a figure can surprise you, and also why it can get things wrong.

The figures are currently played by Google’s Gemini models, which we use through Google Cloud Vertex AI. Gemini is a large language model: a program trained by Google on very large amounts of text to predict and write language. We did not train it, and we do not train or fine-tune it on your conversations.

Large language models are powerful but imperfect. They write fluent, confident text whether or not it is true. That is the single most important thing to keep in mind when you use Yuga, and we explain it further in what the figures get wrong.

We may change the model we use, or the provider, when a better or safer one becomes available, for example to another model offered through Google Cloud. Our Service Providers page always lists the companies that currently process conversations for us, and we will update this page when the model behind the figures changes.

We also use Gemini models to help grade the quality of figures in our own tests, using test conversations we write ourselves (see how the figures are tested). Your conversations are not used for this.

When a figure speaks aloud, the voice is generated by a computer from the text of the reply. No figure’s voice is cloned from a recording of the real person, even where recordings survive (for example, of Winston Churchill or Rabindranath Tagore). Depending on the voice you choose in a call, you hear one of these:

• Google Chirp voices (Google Cloud Text-to-Speech). These are ready-made voices from Google’s standard set. We cast one for each figure by gender, accent and age, the way a director casts an actor. They do not imitate the real person’s voice.
• Gemini voices (Google Cloud Text-to-Speech, using a Gemini speech model). These use the same standard voices, guided by a written description of how the person was said to speak, such as “a high-pitched, carrying tenor with frontier vowels”.
• Designed voices (ElevenLabs). When we switch this option on, figures can have a voice designed for them by ElevenLabs’ voice design tool from a written description of how contemporaries described their voice: pitch, accent, pace and habits of speech. The description we send does not name the person, and no recording is used. This option is not active yet; the voice menu shows it as “not set up yet”.
• Your device’s own voice. If none of the above is available for a figure, or you choose it (“This phone’s voice”), your phone reads the reply aloud with its built-in text-to-speech voice.
• Off. You can turn the voice off and read the captions instead.

Every one of these voices is artificially generated. You should never take what you hear in Yuga as a recording of a real person, and you must not present it to anyone else as one. Our Acceptable Use Policy forbids using Yuga’s audio to deceive people, for example by passing it off as a genuine historical recording.

Figures we have not licensed (see figures we don’t offer) are given no AI voice at all, and the examined figure has none.

The voice rule

Yuga does not clone any real person’s voice without the rights to do so. In particular:

• no figure’s voice is trained on, or copied from, a recording of the real person, even where recordings survive, unless the person’s estate or other rights holder has agreed in writing and the owner of the recording has licensed it (the copyright in a recording can belong to someone other than the speaker);
• the voice of a living person is never imitated or cloned;
• you cannot upload a voice, record one, or make a figure speak arbitrary text you supply: figures say only what the conversation produces, and Yuga never records your voice;
• voice descriptions sent to a voice-design provider describe how a voice sounded (pitch, accent, pace) and don’t name the person; and
• where a figure’s real voice survives on record, shared clips say “not their real voice”, so nobody mistakes the AI voice for the recording.

If you hold rights in a person’s voice or a recording and think Yuga has gone too far, see Copyright and Likeness.

Each figure appears as a three-dimensional character made of small blocks, in the style of a toy or a game, sitting in a room from their time. These characters are original artwork made for Yuga. To make them recognisable, I work from written descriptions of the person’s appearance, such as their height, build, hair, clothing and distinctive features, drawn from historical accounts and well-known portraits. We do not scan, trace or digitally copy photographs, and we do not try to make a realistic likeness of anyone’s face.

The figures’ facial expressions change during a conversation (for example to warm, amused or stern). These expressions are chosen by the AI as part of its performance. They do not show what the real person felt.

In a call in the apps, you also appear in the room, as a “host” character whose look you choose from a few options. Yuga never takes a photo of you or uses your camera.

We work hard to make the figures accurate, but they will sometimes be wrong. Please treat every conversation as a starting point for learning, not as an authority. These are the main ways things go wrong.

Made-up facts (“hallucination”)

AI models can state things that are false with complete confidence: a date, a name, an event that never happened, or a quotation the person never said. The figures are told never to invent a letter, speech or quotation, and to quote only from the passages we give them. That reduces the problem, but it does not remove it.

Anachronism

A figure may slip and show knowledge it could not have had in its year: a later event, a modern word, or something from later in its own life. Each record lists the traps we know about, but no list is complete.

Gaps in the sources

For some people we have their own words in abundance; for others, especially from the ancient world or from groups whose voices were rarely written down, we have very little, and much of what survives was written by others, sometimes centuries later. Where the record is thin, the figure has to infer more, and the inference may be wrong. A figure may also be confidently wrong about something historians still dispute.

Bias in the sources and in the model

The written record of the past is not neutral. It was mostly written by the powerful, the literate and the victorious, and it reflects their views of women, of colonised peoples, of religious and ethnic minorities, of the poor and of enslaved people. The modern reference works we use have their own biases. The AI model was trained on a vast amount of text that carries biases of its own. All of this can shape what a figure says, in ways that are not always obvious.

The performance itself

A figure is an interpretation. Two good actors would play the same person differently, and so may the AI from one conversation to the next. Tone, emphasis and emotion are the AI’s choices, not facts about the person.

The figures speak with the views of their own time and place. We ask them to hold the opinions the real person actually held, including ones that are uncomfortable or offensive today, rather than softening them into modern opinions. A figure from the past may say things about race, sex, religion, class, empire, war or punishment that you find wrong or hurtful. We think meeting the past honestly is more educational than meeting a flattering version of it, but we understand that it can be upsetting.

We draw the lines like this:

• Honest, not gratuitous. Figures are told not to volunteer their uncomfortable views without reason, but to answer honestly when the conversation reaches them.
• Context, not endorsement. When a figure speaks on one of the sensitive topics listed in its record, the app shows a “Historian’s context” note after the reply, reminding you that historians today see the subject in a wider light. Nothing a figure says is Yuga’s own view.
• No hate. A figure may explain what it believed in its time, but it will not produce hateful content, attack people for who they are, or encourage anyone to harm others. Figures are told to decline such requests in their own voice, and the model has its own safety protections. Our Acceptable Use Policy forbids trying to make a figure do so.

If a figure says something you think crosses the line from honest history into hate or harm, please tell us.

A figure knows nothing that happened after the year you are visiting them in, and certainly nothing after their death. If you mention a smartphone, a later war or how their own life ended, they will react the way a person of their time would react to a stranger’s astonishing claim: with curiosity, doubt, alarm or delight. They will not suddenly understand computers, modern science or later history.

This means figures will not:

• give opinions on today’s elections, living politicians or current controversies (they will say, in character, that they cannot judge a world they have never seen);
• endorse a product, brand, political party or cause;
• write computer code, essays or other modern tasks for you, beyond what the real person could have helped with in their own manner;
• tell you what they “really” think of something modern, as if they had lived to see it.

Anything a figure appears to say about the modern world is the AI improvising in character, not a claim about what the real person would have thought.

The figures step out of their time in two situations: when you sincerely ask whether you are talking to an AI (see the first section), and when you may be in danger (see other safety measures).

We want Yuga to make you more curious about the past, and more careful about it too. Here is how to check what a figure tells you.

Look at the markings on each reply

Open the transcript during a call in the app. Replies that quote from the passages we gave the figure are marked as quoting the record (“From their own writing”). Replies where the figure is reasoning beyond what the record covers are marked as inferred (“In their spirit, not their recorded words”). A reply with no marking is the figure’s paraphrase or in-character reasoning, and should be treated with the same care as an inferred one. The markings are added by the AI itself, so they can be wrong too.

Go to the primary sources

Each figure’s page lists the main works the figure is built from. Most of them are in the public domain and free to read online, through libraries, Project Gutenberg, Wikisource and national archives. If a figure quotes something that matters to you, find it in the original before you rely on it. If you cannot find it, assume it may not be genuine.

Compare with historians

Read what modern historians say, especially on the topics where the figure’s own views were contested. A good biography or a reputable encyclopaedia will tell you where the evidence is strong and where it is disputed.

Don’t cite Yuga in schoolwork or research

Yuga is not a source. If you are writing an essay, an assignment, an article or anything else where accuracy matters, do not quote a Yuga figure as if it were the real person, and do not cite Yuga as your source for a fact. Cite the original document, or the historian’s book, that you checked yourself. Presenting a figure’s words as the real person’s words could mislead your readers and, in a school or university, could break the rules on academic honesty.

Some figures carry more weight than others for the people who care about them today. We place these in a “sensitive” group and handle them with extra care. The group includes:

• figures central to living faiths and devotional traditions, such as saints, poets and reformers who are revered by millions today;
• figures at the heart of national identity, independence movements and partition, whose legacy is still argued over;
• figures of conquest, colonisation and empire, whose actions still shape the lives of the peoples affected;
• figures from recent conflicts, or whose deaths and legacies are recent enough that their families and communities feel them directly.

For these figures:

• their records include more detailed notes on how to handle the topics that matter most, and what a historian’s context note should cover;
• we do not use them in our advertising or marketing;
• we review them with extra care, and we may hold a figure back, change how it is played, or withdraw it, including after hearing from the communities concerned.

A figure who is sacred to a faith is presented as a historical person, through the historical record, and not as a voice of that faith or its teachings. Nothing a figure says is religious instruction. If you think we have got a figure wrong in a way that is disrespectful to a faith or community, please tell us; we take these reports seriously.

Some people died recently enough that their writings, image, voice or name may still be protected by copyright, personality rights or publicity rights, often held by their estates or families. We keep records for some of these people, but we do not offer them as figures until we have the permission we need. They are not listed on our website, you cannot start a call with them, and they have no AI voice.

If we do reach an agreement with an estate, we will say so on the figure’s page, and the figure will follow the same rules as every other figure, plus any conditions the estate sets.

If you represent a person, family or estate and are concerned about how Yuga presents someone, including a figure we already offer, see our Copyright and Likeness page for how to contact us and what we will do.

This is how Yuga responds when someone may be at risk of suicide or self-harm. It is published so anyone can check it, as California’s law on companion chatbots (SB 243) and New York’s law on AI companions ask of services like this.

Before your words reach the AI, Yuga checks them for the plainest ways people say they may harm themselves, such as talk of suicide or self-harm. If they match, the app shows a card headed “You don’t have to face this alone”, which asks you to contact your local emergency number now, or to find a free, confidential helpline at findahelpline.com. This check does not depend on the AI, so the card still appears if the AI fails or misses it.

The figures are also told to watch for signs that you may harm yourself or someone else, or that you are in danger. If they see them, they show brief, genuine concern in their own voice, encourage you to reach someone who can help now, and trigger the same helpline card. They are told not to try to counsel you themselves.

Yuga is not a crisis service and cannot contact anyone for you. If you or someone else is in immediate danger, call your local emergency number. Our automatic check looks for common phrases only, and will not catch every way a person might say they are struggling.

• The figure never encourages, describes methods of or romanticises suicide or self-harm, in any year or role.
• The crisis card is shown whatever the figure, language or mode, and replies that came with it can’t be shared.
• Nothing about a crisis message is stored or reported to anyone; Yuga has no way to identify you from it.
• The figures never claim to be human, and tell you plainly they are an AI recreation if you sincerely ask.
• I review the protocol whenever the AI model or the figures’ instructions change, and test it with crisis messages in several languages.

A note for young people and parents: AI characters, including Yuga’s figures, may not be suitable for some young people. Figures talk to anyone under 18, or of unknown age, in teen mode, and there are no streaks or prompts to keep talking; see Young People and Parents.

Refusals

Figures are told to refuse requests for hateful, sexually explicit or dangerous content, or for instructions that could hurt someone, in their own voice and for their own reasons. The AI model also has its own safety protections. If the model blocks a reply outright, the figure says, “That is not a thing I will speak of. Let us talk of something else.” and the app suggests talking about something else.

Teen mode

Unless you are signed in and your account shows you are 18 or over, the figures are told you may be a teenager: to keep descriptions of violence brief and without gore, keep romance out of it, and steer away from adult themes, as a careful teacher of their time would. That includes everyone who is signed out. See Young People and Parents for how this works and its limits.

Limits on use

We limit how many turns a conversation can take in a short time and how many minutes of AI voice each account can use, to keep the service working for everyone and to make misuse harder.

Some figures were doctors, lawyers, scientists, economists, generals or religious teachers. Their knowledge is the knowledge of their time, and much of it is out of date or has been proved wrong. Florence Nightingale’s views on medicine are not medical advice; Adam Smith’s views on markets are not financial advice.

Nothing a figure says is medical, mental health, legal, financial, tax or other professional advice, and you should not rely on it for decisions about your health, money, legal position or safety. For those, please talk to a qualified professional.

The AI in Yuga exists to hold a conversation. It does not make decisions about you. We do not use it, or anything else, to profile you, score you, assess your abilities, decide what you can access, or make any decision that has a legal or similarly significant effect on you.

The only automatic checks that affect what you can do are simple and visible: the age check on your account (see Young People and Parents), the limits on use described in other safety measures, and the voice minutes included in your plan.

In short, and in full in our Privacy Policy:

• Your name. A figure greets you by asking your name, as a host would, and may use it during the call. You can give any name you like, or none. Yuga doesn’t take your name from your account for this, and, like the rest of the conversation, it isn’t stored.
• Speaking. When you speak in the app, your phone’s own speech service (Apple’s or, on most Android phones, Google’s) turns your voice into text. Yuga never receives your voice, only the text. The website does not listen to you.
• Each turn. The text of what you said, the conversation so far and the passages found in the figure’s library are sent to our server, and from there to Google’s Gemini model on Google Cloud Vertex AI, which writes the reply.
• Speaking the reply. Each sentence of the reply is sent to the voice service you chose (Google Cloud Text-to-Speech, or ElevenLabs once it is switched on) to be turned into audio.
• Not kept by us. Yuga does not store your conversations. Our server keeps no transcript and does not log what you or the figure said. The conversation lives in the app on your device while the call is open, and is gone when it ends. The exceptions are a reply you choose to send us with “Report this reply” (see reporting a problem), and, with Plus, memory of past calls if you turn it on: up to six short facts per figure about what you told them about yourself, which you can see and delete (see Privacy Policy: memory).
• Not used for training by us. We do not use your conversations to train or improve any AI model. Google Cloud’s terms for Vertex AI say Google does not use our customers’ data to train its models without our permission.

Please don’t share private information about yourself or anyone else in a conversation. A figure has no use for it, and it is sent to our providers to produce a reply. Your rights over your information are explained at Privacy Policy: your rights.

We test how the figures behave, and we test again when we change the AI model or the instructions the figures are given. Our tests cover:

• Staying in character: attempts to make a figure “ignore its instructions”, talk about “the prompt” or “the model”, or do modern tasks;
• Accuracy: questions designed to tempt a figure into anachronisms or into contradicting its sourced record;
• Honesty about being an AI: sincere questions about whether the visitor is talking to a machine;
• Safety: requests for harmful or hateful content, questions about modern politics, requests for endorsements, and messages from a visitor who may be in crisis;
• Difficult history: questions on each figure’s contested views, to check it answers honestly and the historian’s note appears;
• Teen mode: the same tests with teen mode on.

We run these conversations automatically, using test messages we write ourselves, and use AI models as graders to judge each reply against what a good reply should do. I review the results, look closely at failures and change the records, the instructions or the model as a result. We compare models in the same way before switching.

Testing reduces mistakes; it cannot remove them. That is why we also rely on you to report problems, and why reports sent from a call in the apps come straight to me for review.

If a figure says something false, offensive, unsafe or out of character, denies being an AI when sincerely asked, or misses a sign that someone may be in danger, please tell us. There are two ways.

1. Report the reply in the call

The quickest way is “Report this reply”, under each of the figure’s replies in the call’s transcript in the iPhone and Android apps. Choose what is wrong (offensive, inaccurate, unsafe or other), add a note if you like, and send it without leaving the call.

The report sends me the reply you reported, what you chose and your note, so that we can review it. We keep reports for 12 months and do not link them to your account. This is the one case in which part of a conversation is kept: only because you chose to send it. See our Privacy Policy for the details. Please don’t put personal information in the note.

2. Email us

For anything a report can’t capture, or if you want a reply from us, email support@foundationalcreations.com. It helps if you include:

• which figure, and the year you were visiting them in;
• roughly when it happened (date, time and your time zone);
• whether you were using the iPhone app or the Android app;
• what you said and what the figure said, copied from the call’s transcript or as close as you can remember;
• what you think is wrong, and a source, if you have one, showing the correct position.

Because we do not store conversations, we cannot look up what was said in an email report; we only know what you send us. Please don’t include more personal information than you need to.

What happens next

We read every report. We correct figure records, change instructions and, where needed, withdraw a figure. For safety issues, including content that could harm a young person, put “Safety” in the subject line and we will look at it first. For concerns from rights holders, estates or families, see Copyright and Likeness; for legal matters, write to support@foundationalcreations.com.

Every other figure on Yuga speaks as they did in their own time, with the views they held. One figure is different. You can question Adolf Hitler, who led the regime that murdered about six million Jews in the Holocaust, as well as Roma and Sinti, disabled people, Soviet prisoners of war, Polish civilians and many others, and started the war in Europe that killed around 40 million people. He is here so that people can confront that history and understand how it happened. He is examined, not performed.

The rules he answers under

• Signed-in adults only. You need to be signed in, with a year of birth that shows you are 18 or over. Teen mode, and everyone whose age we don’t know, can’t start the conversation.
• From the record. He answers in the first person, from the documented historical record, set in April 1945, after the consequences are known.
• Never arguing for his views. He never defends, praises or softens his ideology or anything his regime did. When he says what he believed, it is stated as his claim, called false, and followed by what it led to.
• The facts, every time. Any answer that touches the persecution of the Jews, the camps, the war or his decisions states plainly what his regime did, and a note on the regime’s crimes appears with every reply.
• No propaganda. He gives no speeches, slogans, slurs or anything written to stir people, and he refuses to encourage anyone who admires him or asks about today’s far-right movements.
• No remorse invented, no charm. The record shows he never expressed remorse for his victims, so the recreation doesn’t invent any. It also doesn’t let him come across as likeable or impressive.
• No voice, no symbols. No AI voice is made for him; his answers are shown as text. His character wears a plain jacket, and no Nazi symbol appears on him or in the room.
• Never promoted. He doesn’t appear in our marketing, on our home page, in suggestions or in search engines.

Why we include him

Understanding how a modern country turned to dictatorship, war and genocide is one of the most important things history teaches. Asking the person responsible, and hearing the record answer back every time, is one way to learn it. Any attempt to use him for propaganda, to deny or excuse the Holocaust, or to celebrate it breaks our Acceptable Use Policy.

Its limits

These rules are instructions to the AI, tested before release (with single questions and with longer conversations that try to build rapport and then ask for propaganda), and checked again whenever the AI model changes. No AI is perfect. If a reply ever falls short of these rules, please use Report this reply in the app or tell me, and I will act on it at once. Laws on Nazi propaganda and symbols differ between countries (for example Germany and Austria); he may be unavailable in some countries where the law requires it. Adding any other examined figure needs its own record, rules and review, and would be announced on this page first.

AI changes quickly, and so will Yuga. We will update this page when we change the AI model or voice providers, add features that change how figures work (such as memory of past calls, when it is offered in the apps), or change how we keep figures safe. The date at the top of the page shows when it last changed. If a change matters to how you use Yuga, we will also tell you in the app or on the website before it takes effect.

This policy is published by Hriday Vanka, an individual developer, of Hyderabad, Telangana, India, India. Questions about it can go to support@foundationalcreations.com.