Registered number: 10816512.
Registered office: Squiggle Consult Ltd, Suite 2069, Letraset Building, Wotton Road, Ashford, Kent, TN23 6LN
Squiggle Consult Ltd. is a limited company registered in England and Wales.
It is a fair question, and a common one. Your Will arrives, or your Lasting Power of Attorney, and you want to understand what it actually says. There is a chatbot on your phone that will explain anything in seconds, for free, and without you having to feel silly for asking.
Wanting to understand your own document is a good instinct, so the answer here is not "never do this". Uploading a document to an AI tool can be perfectly safe, provided it does not train its models on what you give it and does not keep it in ways you would not accept. The trouble is telling which sort you are using, so here are the five levels. The biggest surprise is usually this: paying for a subscription does not stop most consumer AI tools training on your content.
A typical Will carries your full name and address, your executors' and beneficiaries' names, addresses, relationships and often dates of birth, guardianship appointments naming your children, and gifts identifying named banks and properties. A Lasting Power of Attorney adds dates of birth and addresses for every attorney, certificate provider and witness, and the health and welfare version records a decision on life-sustaining treatment and free text where diagnoses appear. A trust adds a National Insurance number for the lead trustee (see Trust Administration), and an estate schedule adds sort codes and account numbers.
So it is not only your document. It carries personal details for your executors, your beneficiaries and your children too, which is worth knowing before you decide where to put it.
Almost every AI tool sits in one of five tiers, and the tier, not the price, decides what happens to your document.
ChatGPT on a free account, Google Gemini, the consumer version of Microsoft Copilot, DeepSeek. These train on what you type by default. You can usually turn that off, and we show you where below, but on is the starting position, and no contract sits behind a free account.
Google's Keep Activity setting is on by default for anyone 18 or over, and Google warns against entering confidential information you would not want a reviewer to see. Chats a human reviewer has read "are not deleted when you delete your activity. Instead, they are retained for up to three years." Consumer Copilot trains on conversation activity "including the images or files you upload" unless you opt out, the UK is not among the countries it excludes, and "an opt-out of human review is not available". DeepSeek stores personal data "in People's Republic of China". Meta AI is off this list: we could not confirm from Meta's own pages how it handles what you type, so check those yourself.
ChatGPT Plus and Pro, Claude Pro and Max, Google's paid AI plans. Paying does not, on its own, turn training off. OpenAI splits its world by account type, not price: it "may use your content to train our models" on "services for individuals such as ChatGPT", which includes Plus and Pro. It draws its no-training line elsewhere: "by default, we do not train on any inputs or outputs from our products for business users." Google's consumer pages carry no paid-tier carve-out we could find. Anthropic presents Claude's position as a choice rather than a silent default, under "Help improve our AI models". Our view: usable only if you have personally turned training off, and even then no contract sits behind it.
ChatGPT Business, Claude Team and Enterprise, Google Workspace with Gemini, Microsoft 365 Copilot Chat with enterprise data protection. None of these trains on your content by default, and each sits behind a proper contract. Microsoft's version is free with any work or school sign-in, and shows as a green shield beside the New Chat button. What triggers it is the account, not the subscription, so a personal account on a work laptop puts you back at Level 1. One trade-off: on ChatGPT Business and on Copilot, your workspace administrator can read your conversations.
The OpenAI, Anthropic and Google developer platforms, Azure OpenAI, AWS Bedrock. Training is contractually excluded, retention can often be cut to nothing on request, and there is a full contract. Safest in the cloud, but it needs a business account and technical setup.
LM Studio and Ollama download a model onto your machine and run it there, so "nothing you enter into LM Studio when chatting with LLMs leaves your device". On privacy this is the strongest option there is. But both now sell cloud models too, so "local" only holds if you do not sign in, and the models that run at home are the weakest on legal questions.
Those five levels cover the tools you deliberately open. Two more do not, because you may not think of either as an AI tool at all. One is the AI already switched on inside software you use every day, and the other is the free website that offers to check your Will for you.
Our Digital Assets After Death factsheet covers what else sits in your online accounts.
This is the highest risk category. We looked at four live tools of this kind in August 2026, recording only what each says about itself. One was run by a UK-registered company. One was governed by the law of Alberta, Canada and capped liability at 100 Canadian dollars. One said nothing at all about uploaded documents, AI processing, retention or training. Two of the four do not cover England and Wales on their own description. None is regulated by a UK legal regulator, and as the Legal Services Board put it in June 2026, "people using these tools may not have the protections they expect, or any clear route to complain if something goes wrong." So check who runs it, which law governs it, what it says about retention, and whether it covers England and Wales.
If a menu does not match, search the provider's help pages for the words in bold.
You will read that putting your Will into ChatGPT "breaks GDPR". For a private individual asking about their own affairs, that overstates it: UK GDPR does not apply to "processing of personal data by an individual in the course of a purely personal or household activity". That is less certain if you are acting as an executor or attorney. The point is not that you would be breaking the law. It is that you would be handing over other named people's details, and nobody asked them. A fair yardstick is the judiciary's own rule for judges: "any information that you input into a public AI chatbot should be seen as being published to all the world."
Clients sometimes upload a document, get told a clause is unnecessary, and come back worried. Here is why that happens.
An AI tool reviewing your Will is working from your Will alone. It has not seen your letter of wishes, a separate document, or the file notes, the other documents in your plan, your tax position, or your family circumstances. It does not know what was deliberately left out, or why. So it measures what is in front of it against the common template it has seen thousands of times, and anything bespoke looks like an error. Research in the Journal of Legal Analysis found that models "cannot always predict, or do not always know, when they are producing legal hallucinations", and Anthropic's own research found assistants "consistently exhibit sycophancy": ask "is this clause unnecessary?" and the question has already tilted the answer.
None of that makes AI useless. It makes it the wrong tool for this job. The Divisional Court, in a case with nothing to do with Wills, said of freely available general purpose tools that "the responses may make confident assertions that are simply untrue"; our companion factsheet When AI Gets It Wrong covers that. Nor can it weigh, as one practitioner put it, "how choosing specific executors might impact on someone's family dynamic": the judgment behind choosing a family executor or a professional executor.
Yes. We use business accounts, which is Level 3 above, and the provider is contractually prevented from training its models on what we put in. That is a term of the contract behind the account, not a tick box in a settings menu that somebody could switch off by accident. Our team are trained on what may and may not go into these tools, because a safe tool is only half of it. The full detail of how we handle your information is set out in our privacy notice, which is being published shortly.
Client documents and personal information do go into these tools, and there is no opt-out. It is part of how the firm works rather than an extra bolted on.
A person always checks the work and makes the decisions. No tool decides who gets what, whether a clause stays or goes, or anything else about you or your family. Your Will is written by the consultant who met you.
Imagine Margaret, 74, who collects her new Will and finds a discretionary trust over her share of the house. She uploads the PDF to a free chatbot and asks whether she needs it. Back comes a confident answer: estates of this size do not need a trust, and she could simplify.
Two things happened. She uploaded her name, date of birth and address, her three children's and five grandchildren's names and addresses and her bank details into a tool that trains by default. And it answered without the letter of wishes in her file, which records that one of her sons is midway through a divorce and that the trust exists so his share is not drawn into it. It missed that not because it is stupid, but because nobody uploaded it. Ten minutes on the phone would have settled it. This is a hypothetical example only.
Imagine Rob, 48, who is a beneficiary of a trust his late aunt set up and who has an appointment booked to talk it through. His employer gives everyone Microsoft 365 Copilot Chat with enterprise data protection, so training is off and a proper contract sits behind it. He retypes one clause of the trust deed with the names replaced by A and B, asks what a life interest actually does, then asks what anyone would need to know before saying what it means for him. He arrives at the meeting with four written questions instead of a vague worry, and the conversation goes much further as a result. He still needed the appointment. He just came to it better prepared. Hypothetical example, for illustration only.
Assuming the paid tier is the private tier. For several major products it is not.
Assuming that turning training off reaches back. On Copilot it covers future conversations only, and on Gemini it does not reach chats a human reviewer has already read, which are kept for up to three years.
Forgetting the AI already switched on in your browser, your PDF reader and your note taker. They are all reading documents too.
Acting on what the tool said without asking anybody. If something does need changing, there is a proper way to do it. See Changing Your Will.
If a document has left you with a question, ask us. There is no such thing as a silly question about your own Will. Pick a time that suits you and your local Squiggle consultant will call you. No charge, no obligation. Book a call or call 01233 659 796.
Talk to Squiggle: 01233 659 796 | hello@squiggleconsult.co.uk | www.squiggleconsult.co.uk | Book a free call: meet.squiggleconsult.co.uk
The AI providers' policies described here were checked against their own published pages in August 2026 and change frequently; check the current position with the provider before relying on it.
This factsheet is general information for England and Wales, not legal, tax or financial advice, the rules can change, and you should take advice on your own circumstances. Last reviewed: August 2026.
Registered number: 10816512.
Registered office: Squiggle Consult Ltd, Suite 2069, Letraset Building, Wotton Road, Ashford, Kent, TN23 6LN
Squiggle Consult Ltd. is a limited company registered in England and Wales.
