AI confidentiality charter: what it must contain, and how to make staff actually follow it
Most organisations now have staff pasting work into ChatGPT, Claude, Gemini or Copilot. A charter says what is allowed. This guide lists what it must contain, country by country, and explains why a charter that nothing enforces protects no one.
- A charter answers four questions: which AI tools, which data, what never leaves, and who to tell when something goes wrong.
- The core rule is simple: client, employee and partner identifiers are coded before any message reaches an AI, never pasted in plain text.
- Red lines (bank details, ID numbers, secrets) stay out even when coded; everything else can be used once it is pseudonymised.
- The applicable law depends on the country: the GDPR and the EU AI Act in Europe, the UK GDPR, sector laws in the United States, national laws in the Gulf, Africa and Asia.
- A charter that nothing enforces is a PDF. Tie each clause to a rule a computer applies, and check it every month.
Why a charter, and why now
In most organisations the question is no longer whether staff use generative AI, but which accounts they use and what they paste into them. A contract summarised in ChatGPT, a payroll file analysed in Claude, a medical report rewritten in Copilot: each one sends names, identifiers and amounts to a provider outside the organisation.
A charter fixes the rules before an incident does. It also gives your DPO, your auditors and your clients a clear answer to the question they now ask: what do you do with our data when you use AI?
The ten clauses every AI charter needs
- Scope. Who is covered (employees, partners, interns, contractors) and which activities.
- Approved tools. The AI assistants the organisation provides, under organisation accounts. Anything else falls under the personal-account rule.
- Personal accounts and phones. Whether organisation data may go through a personal ChatGPT or Claude account, or an AI app on a phone. Most charters forbid it for client and employee data.
- The golden rule. Identifiers of clients, employees and partners are coded before a message reaches an AI. Names become codes such as
NAME-001, the AI works on the codes, and the real values are put back on the organisation's computer afterwards. - Red lines. What never leaves, even coded: typically bank details and card numbers, ID document numbers, dates of birth, passwords and API keys.
- Data categories. Health data, payroll, data about minors and case files get stricter handling. Say which.
- Internal references. Case numbers, contract numbers, project code names and client names that must always be coded, because no generic tool knows them.
- Incidents. What to do when identifying data went out in plain text: tell the AI lead the same day, without fear of sanction for a report made in good faith, delete the conversation at the provider, assess the risk with the DPO.
- Training. A short induction for everyone who uses AI. In the EU, this is also how you meet the AI literacy duty of the AI Act.
- Governance and review. Who owns the charter, who approves new tools, and when it is reviewed.
The law that applies depends on the country
The clauses above are the same everywhere. The legal annex is not. A few examples:
- European Union. The GDPR (pseudonymisation, security, transfers outside the EU) and the EU AI Act, whose art. 4 on AI literacy has applied since 2 February 2025. Each member state adds its own data protection act and professional secrecy rules.
- United Kingdom. The UK GDPR and the Data Protection Act 2018, with the ICO's guidance on AI and data protection, plus professional rules such as the SRA Code of Conduct for solicitors.
- United States. No general federal privacy law: HIPAA for health, the Gramm-Leach-Bliley Act for financial institutions, the FTC Act, and state laws such as the CCPA. For lawyers, ABA Formal Opinion 512 (2024) addresses generative AI directly.
- Gulf states. The UAE's Federal Decree-Law No. 45 of 2021, with separate regimes in the DIFC and ADGM; Saudi Arabia's Personal Data Protection Law, enforced by SDAIA since September 2024; Qatar's Law No. 13 of 2016.
- Africa. Morocco's Law No. 09-08, Tunisia's Organic Law 2004-63, Nigeria's Data Protection Act 2023, South Africa's POPIA, Kenya's Data Protection Act 2019, among others.
- Asia-Pacific. India's Digital Personal Data Protection Act 2023, Singapore's PDPA, Australia's Privacy Act 1988.
Each country also has its own identifiers to protect: Social Security numbers in the United States, Emirates ID numbers in the UAE, Aadhaar and PAN in India, the NIR in France. A charter should name them, and the tools should recognise them.
Why most AI charters fail
The typical charter is written once, signed, filed, and forgotten. Staff do not reread it before pasting a contract into an AI. Nobody can tell whether it is followed. When a client asks for proof, there is none.
A charter works when each clause is tied to something a computer does:
- the golden rule is applied by a tool that codes identifiers before any message leaves, rather than by memory;
- red lines are blocked, not just forbidden;
- internal references and national ID numbers are rules the tool knows;
- a monthly report shows which computers follow the charter and which do not.
That is what Hidden Tokens does on each computer: it codes names, identifiers and amounts locally before Claude, ChatGPT, Gemini or Copilot reads them, applies the organisation's signed policy to every message, and reports coverage clause by clause.
Generate yours, free
The Hidden Tokens AI charter generator asks twelve questions and writes the charter for you, with the legal annex for more than thirty countries. It also produces the two files that turn the charter into rules: the organisation policy your admin signs, and the settings for a single computer. It runs entirely in your browser: the page cannot make any network connection, so nothing you type leaves it.
Have the result reviewed by your counsel or DPO, adopt it, and tell your staff. For the difference between coding and removing data, see pseudonymisation vs anonymisation; for the day-to-day method, how to use ChatGPT or Claude with client data.
Frequently asked questions
Is an AI charter mandatory?
Rarely by name, but the obligations behind it are. In the EU, the AI Act (art. 4) requires organisations that use AI to ensure their staff have sufficient AI literacy since 2 February 2025, and the GDPR requires appropriate security for personal data. Professional secrecy rules apply whatever tool is used. A charter is the simplest way to show how you meet them.
Should we simply ban ChatGPT and Claude?
Bans tend to move use to personal accounts and phones, where nothing is controlled. Most organisations get better protection by allowing approved tools and requiring that identifiers are coded before anything is sent.
What is the difference between an AI charter and an AI policy?
In practice, none: some organisations call it a charter, others an acceptable use policy or a code of conduct. What matters is that it is adopted formally, communicated to staff and enforced.
Who should own the charter?
Name one AI lead who answers questions and receives incident reports, with the DPO or counsel for the legal side and IT for the tools. Review it at least once a year, because tools change faster than charters.
Is the Hidden Tokens generator free?
Yes. It runs entirely in your browser, makes no network connection and needs no account. It produces the charter as a Word file, plus the rules files that Hidden Tokens can apply on each computer.