Where does my text go?
The short answer: it depends on the tool. Most tools send your text to one AI company. Multi-model chat sends it to six. Here’s what happens, who can see it and how to remove it.
What happens when you paste something
- Your text travels over an encrypted connection (HTTPS) to our servers, which run on Vercel in the US.
- Our code sends the text to the AI company or companies that tool uses, also over HTTPS.
- Code checks what comes back, and you get the result.
- Depending on the tool, we either discard the text or store it for the time shown below.
Stored text sits in our database (Supabase, in the US), which encrypts everything at rest with AES-256. Verify text and Help Center Check source documents are encrypted a second time by our own code before they’re saved (AES-256-GCM). Multi-model chat text and Help Center Check’s saved answers and findings have only the database’s encryption.
How long we keep it
The same table as our privacy policy. A daily job at 05:50 UTC deletes what has expired.
| Tool | What you paste | Results |
|---|---|---|
| Offer Check (free tool and Pro) | Not kept. The text and the terms are checked and discarded; the claims check runs without storing anything. | Not kept. One row of counts per scan (the light, the tactic names, the cost), no text. |
| Verify free tool (/verify, no account) | One day, encrypted, then deleted by the daily purge. | One day, with the text. A share link keeps a copy with personal data masked until the link expires, 90 days. |
| Verify with an account (API, MCP server, batch reports, hosted Verify Gate) | Encrypted for your plan’s window, then deleted: 7 days on Free, 30 on Builder, 90 on Team. An encrypted list of the claims found in an answer is cached for 24 hours. | The same window as the text. |
| Help Center Check free tool (/help-center-check/try) | One day to run the check, encrypted, then deleted. A draft check ("check a draft before you publish") keeps none of the text. | One day, with the text. |
| Help Center Check workspace (trial, founding plan) | Each captured version of your sources, encrypted, for as long as the workspace exists. | Every check, finding and alert, for as long as the workspace exists. You can export it all; email us to delete the workspace and everything in it. |
| Multi-model chat (/query) | Your questions, in your history, until you delete them there or we delete your account. | The answers and a receipt page anyone with its link can open, until you delete the question or we delete your account. |
| Private check, on any tool | Not kept by us. Sent only to Anthropic to run the check; Anthropic deletes it within 30 days, or keeps it up to 2 years if flagged for a policy violation. | Not kept. One anonymous counter row: when, which tool, the cost. |
| Private check, “No AI at all” | Not kept, and it never leaves our servers. | Not kept. One anonymous counter row: when, which tool. |
The evidence ledger and research copies are on the full table. AI companies keep their own copies for their own periods, listed further down.
Who can see it
- You. Your results and your history are tied to your account.
- Anyone with a link. Each multi-model chat answer has a receipt page. It isn’t listed anywhere and search engines are told to skip it, but anyone who has the link can open it, including your question. A Verify share link shows a masked copy for 90 days. Only share links you’re happy to make public.
- Xyntherium. The owner has admin access to the service. Support tickets can be read there. Verify text and Help Center Check source documents are decrypted only by the running service. Multi-model chat text and Help Center Check’s saved answers and findings are readable in the database. Our access policy limits when anyone may read customer content, and admin reads are logged.
- Our infrastructure providers. Vercel runs the servers and Supabase runs the database. Neither is meant to read your text, but both handle it.
- The AI companies. Each one receives the text its step needs, and handles it under its own terms, below.
Which AI companies process it
| Tool | Sent to | Only if you ask |
|---|---|---|
| Offer Check | Anthropic | “Ask all six” also sends the text to OpenAI, xAI, Perplexity, Google, DeepSeek. |
| Help Center Check, free tool | Anthropic | None |
| Help Center Check, paid workspace | Anthropic | None |
| Verify, free tool and demo | Anthropic | None |
| Verify, account and API | Anthropic | None |
| MCP server | Anthropic | The cross_check tool (paid plans) also sends the question to OpenAI, xAI, Perplexity, Google, DeepSeek. |
| Multi-model chat | Anthropic, OpenAI, xAI, Perplexity, Google, DeepSeek | None |
Multi-model chat also uses OpenAI to compare the answers and Anthropic or OpenAI to write the combined answer, whichever your plan uses.
Which AI companies see my text
What each company says it does with text sent through its API. Those are their terms, not ours, and we don’t hold a zero-data-retention agreement with any of them.
| Company | Built in | Where it runs | Used for training? | How long they keep it | Policy |
|---|---|---|---|---|---|
| Anthropic | United States | Stored in the US. By default Anthropic may route traffic to the US, Europe, Asia and Australia. | No. API inputs and outputs aren’t used for training by default. | Deleted within 30 days. If a request is flagged for a Usage Policy violation: up to 2 years, and safety scores up to 7 years. | Anthropic: is my data used for training?Anthropic: how long data is stored |
| OpenAI | United States | OpenAI’s default. We haven’t pinned a region. | No. API data isn’t used for training unless the customer opts in. | Abuse-monitoring logs, which may include prompts and responses, for up to 30 days. | OpenAI: your data |
| xAI | United States | xAI says its default endpoint doesn’t guarantee a processing region. We use the default endpoint. | No. API inputs and outputs aren’t used for training without permission. | Requests and responses stored encrypted for 30 days, then deleted. | xAI: security FAQ |
| PerplexityPerplexity’s legal page says Sonar is based on Meta’s Llama 3.1 70B and fine-tuned by Perplexity; its 2025 post says Llama 3.3 70B. Neither names sonar-pro, the model we call. | United States | Perplexity doesn’t publish a region for this API. | No. Perplexity’s API terms say customer content isn’t used to train AI models. | No period published. Perplexity says zero data retention isn’t the default. | Perplexity: API privacy and securityPerplexity: API terms |
| United States | Google says data may be stored transiently or cached in any country where it or its agents have facilities. | Depends on the billing tier. Paid: not used to improve products. Unpaid: used to improve products, and people may review it. | 55 days for abuse monitoring. | Google: Gemini API additional termsGoogle: abuse monitoring | |
| DeepSeek | China | China. DeepSeek’s privacy policy says it processes and stores personal data in the People’s Republic of China. | No promise. DeepSeek’s developer terms don’t say API inputs are excluded from training, and its privacy policy lists training as a use of personal data. | No period published for API inputs. Its privacy policy says it keeps input data for as long as there’s an account. | DeepSeek: Open Platform termsDeepSeek: privacy policy |
- Google’s unpaid tier allows training and human review. Treat Gemini as the unpaid tier until the paid tier is confirmed.
- DeepSeek makes no no-training promise for API text, publishes no retention period and stores data in China.
Choosing providers doesn’t make Xyntherium approved for government, classified or patient data. Follow your organization’s rules.
How to exclude an AI company
- In multi-model chat and “Ask all six”: open Choose AIs and untick any company, or press “US-developed models only.” Agreement is then counted over the AIs you selected.
- As your default: Account, AI providers.
- For your whole organization (Team): an admin sets the list there too, and members can’t select the others.
- On the API and MCP: send
providers_allow,providers_denyorproviders_preset: "us_developed". If nothing allowed remains, you get a 400 with the reason.
Offer Check, Verify and Help Center Check run on Claude, the model they were measured on. If you exclude it, they tell you before anything is sent, and you can use “No AI at all” instead. Your choice is enforced on our servers, where each outgoing call is checked. Results show who was used: “Checked with: Claude, ChatGPT, Gemini.”
Private check
Tick “🔒 Private check” next to the button. It’s off by default and you tick it each time. There are two levels.
| Level | Where your text goes | What’s kept | What it can’t do |
|---|---|---|---|
| Private check | Only to Anthropic, to run the check. Multi-model chat uses that one provider instead of six. | On our servers: an anonymous counter (the time, the tool, that it was private, the cost) to hold limits and spend. No account, no address, no text. | No share links, no history, no saved result, no research use. If you close the page, the result is gone. |
| No AI at all | Nowhere. It runs on our servers and goes to no AI company. | The same anonymous counter. | Catches less: no pressure-tactic detection, no web sources and no links. It only runs our code-only checks, and it isn’t offered in multi-model chat. |
We use Anthropic because its API terms don’t allow training by default and its 30-day retention is among the shortest published. That isn’t zero: Anthropic deletes API inputs and outputs within 30 days, and keeps them up to 2 years if they’re flagged for a Usage Policy violation. We hold no zero-data-retention agreement. Anthropic: how long data is stored.
Private check isn’t available for batch checks, image checks or a Help Center Check workspace, because those keep your text so they can do their work. Verify’s private API checks are limited per day and don’t count toward your monthly allowance.
We don’t train on your data unless you opt in
We don’t train AI models on what you send us. If you tick the research box on a check, we keep a masked copy of a check that found a problem, to study where AI answers go wrong. That’s off by default, it covers one check, and even then we don’t use it to train models. Research use, in full.
That promise is ours. Each AI company’s own policy is in the table above.
How to delete your data
- Multi-model chat: delete any question in History. Its answers and its receipt page go with it.
- Verify and the free tools: text is deleted automatically when its window ends.
- Anything else, or your whole account: email admin@truelevelmarketing.com from the address on your account. We answer within 30 days.
Deleting from Xyntherium doesn’t reach copies an AI company has already kept. Those end on the company’s own schedule.
What to avoid
- Don’t paste passwords, API keys or government ID numbers into any tool.
- Don’t put confidential documents into multi-model chat in its normal mode: six AI providers process it, one of them in China.
- If you’re under rules about client, patient, government or classified data, follow those rules. Nothing here makes Xyntherium approved for that data.
Questions: admin@truelevelmarketing.com. See also our privacy policy and security page.