Your voice, your accounts, your home: privacy checks before AI training work

The work has moved closer to you

Early AI training work was text on a screen: rate this answer, write a better one. A growing share of the listings on this board now ask for something personal as the raw material. The model needs to hear a real voice, act inside a real inbox, watch a real desktop or see real hands folding real laundry, and the cheapest way to get that is to use yours.

The work can still be legitimate, but it changes what you are agreeing to. When the deliverable is a recording of your voice or a session inside your own accounts, you need to know what you are handing over, to whom, and for how long, as well as what it pays. Most listings say what it pays and leave the rest unanswered.

Below are the four kinds of personal input that current listings ask for, with real examples from this site, followed by the questions to settle before you start.

1. Your voice, licensed for cloning

Mercor runs a family of voice-cloning listings, one per accent or language, such as the German, Swiss German and UK English versions. The shape is the same across them: record a session of about four hours, which a client uses to clone your voice for its internal customer-experience AI agent. Pay is $50 to $100 an hour (up to $150 for Swiss German), so roughly $200 to $400 for a session.

That is a one-off payment for something that can keep working after you stop. Our write-up of the German listing notes that the ad limits how the voice will be used but leaves open how long the licence lasts and whether you can withdraw it. With voice work, those two gaps matter most.

Check before you record:

• How long can the client use the clone, and is there an end date?
• Can you revoke consent, and what happens to the model if you do?
• Is use limited to the stated purpose (an internal agent), or can it be sold on or used publicly?
• If you work professionally as a voice actor, narrator or dubbing artist, does your agency or union agreement say anything about AI cloning? Some do, and a four-hour session is not worth breaching one.

Voice recording for speech datasets (reading prompts so a model learns to recognise speech, as in micro1's audio-recording listings) is a different and lower-stakes thing from cloning. The listing usually tells you which one it is; if it does not, ask.

2. Your own accounts and apps

Some evaluation work only makes sense on a real person's digital life, so the listing asks you to bring yours.

micro1's Gmail and Google Calendar AI Assistant Evaluator asks for daily personal use of Gmail and Google Calendar, plus an iPhone with iMessage and a Facebook or Instagram account, and involves handing everyday email, scheduling and booking tasks to an AI assistant. The ad does not explain why those particular accounts are needed. Invisible Technologies' AI Personalization Evaluation Expert listing requires "a certain threshold of recent activity across a specific combination of commonly used digital services", verified during qualification, and has you chat with two AI assistants that can draw on "information or activity available through your connected services". Mercor's Personalized Life Assistant Expert is lighter: it asks for heavy personal use of AI assistants and judges answers to personal requests about food, health, careers and learning.

When a job connects an AI system to your real accounts, two things can leave your control: the contents of those accounts (other people's emails, your calendar, your messages) and whatever the evaluation data records about your usage.

Check before you connect anything:

• Are test accounts provided, or must you use your own? If your own, can you create a fresh account for the project instead?
• What exactly will the assistant be able to read, and does access end when the project does? Revoke any connection yourself afterwards; do not rely on it expiring.
• Do the tasks involve contacting real people? Other people in your inbox did not agree to be training data.
• Could a task cost money? Bookings and orders made during testing can carry real charges or cancellation fees. The Gmail listing above does not say who covers them.

3. Your screen

Computer-use models learn from recordings of people using software. micro1's GitHub Specialist has you carry out GitHub workflows while recording your screen, and Mercor's Application Users family (PyCharm, Visual Studio Code and others) asks for real work inside a named application on your own Mac, captured at more than 2.5 megapixels.

A screen recording captures everything on the screen, not only the task: notifications, open tabs, file names, the corner of a password manager, an access token in a terminal.

Check before you press record:

• Use a separate user account or a clean desktop for the work, with notifications off.
• Never record inside real client or employer repositories, or with production credentials loaded. Many employment contracts forbid it outright.
• For work that requires your own licence (Mercor's listings say it provides no licence, VM or hardware), confirm the licence terms allow this kind of use.
• Ask whether recordings are reviewed by people as well as used for training, and how long they are kept.

4. Your home and your body

Robotics models need to see how people move. micro1 has several listings for first-person video of everyday tasks, including its US at-home video recorder and a Spanish-language version, Graba Videos desde Casa, which has you wear a smartphone on a head strap and capture video plus the phone's motion-sensor data while both hands stay in frame. That listing also requires identity verification before a contract is issued.

First-person footage of your home can show far more than your hands: family members, the layout of rooms, documents on a counter, the view from a window, a street sign.

Check before you film:

• Clear the frame of other people, mail, screens and anything with an address or name on it.
• Ask whether faces or voices that appear by accident are removed, and how.
• Confirm where footage is stored and who can see it, given that identity documents are also on file.

Questions that apply to all four

Whatever the input, these are the questions that matter, and most listings answer none of them:

1. Who is the end client? Listings usually say "a leading AI lab" or "a customer". You may not get a name, but you can ask what kind of organisation it is.
2. What exactly is kept, and for how long? Recordings, clones, account data and your identity documents are all different answers.
3. Can you withdraw, and what happens then? For a voice clone this is the most important question of all.
4. Is use limited to training, or can it be used in a product? A voice clone used in a live customer-service agent is a product use.
5. Where is it processed? If you are in the EU or UK, data protection law gives you rights over personal data, including your voice and image. A company processing it should be able to tell you its legal basis and how to exercise those rights.

If a recruiter cannot or will not answer these, that tells you something. Legitimate platforms usually have a data-handling or privacy page; Invisible, for example, states on its Meridial FAQ that it operates under SOC 2, HIPAA and GDPR compliance. A certification is about how data is secured, not about what the client may do with a clone of your voice, so it does not replace the questions above.

Your employer's material and your contract

The four sections above are about your own data. Expert evaluation work raises a second question: whose material you bring to it.

If you have a job, a professional licence or confidential clients, read your contract before taking outside AI training work. Look for clauses on outside work or moonlighting, on conflicts of interest, and on who owns what you create. Some ownership clauses are worded broadly enough to reach work done in your own time. People in regulated professions, government roles or client-facing consulting may need permission first. If the wording is unclear, ask your employer or a lawyer rather than guess.

Then keep the two jobs apart. Never paste an employer's or client's code, documents, research or data into a task, a prompt or an application sample, and never use patient or customer records, even with names removed. When a task or work sample needs a realistic example, write your own or use public material such as an open-source project, a published paper or a textbook problem. The screen-recording rule above (no real client repositories, no production credentials) is one case of the same principle.

Is it worth it?

Often, yes. Several of these listings pay well for little time, need no professional credential, and are honest about what they are. Decide on them with the same clarity you would bring to signing a model release or a voice contract, because that is closer to what they are.

The more of you that is in the data, the more you should know about where the data goes: a rating task needs no questions, and a clone of your voice needs all of them.

For the general checks that separate real platforms from imitations, see is AI training work legitimate, and how to spot the fakes. For payment terms, see how you actually get paid.

Sources

Every listing detail above is taken from the listings on this site as read on 24 to 26 September 2026, which in turn quote the platforms' own ads. The SOC 2, HIPAA and GDPR statement comes from Meridial's public FAQ, read on 26 September 2026. This guide is not legal advice. We are not affiliated with any of these platforms and we do not process applications.