How We Use AI (Honestly)

In plain English. No legal jargon. Last updated 5 June 2026.

What is AI doing on Fightingback

When you generate a letter:

• We send the questions you answered to Anthropic's Claude Sonnet 4.6, along with a tool-specific prompt that tells the model what legislation to cite and what tone to use.
• The model writes the letter. We do not edit it (beyond automated safety checks). What you see in the preview is what the model produced.
• We log the call to our own database so we can monitor cost, quality and abuse. The log includes: input length, output length, model used, our internal user ID and timestamp. We do not log credit-card numbers, passwords or anything Anthropic would not also see.

When you upload to Decoded:

• We extract the text from the document (no images go to the model, just text).
• We send the extracted text to Claude Sonnet 4.6 along with a Decoded-specific prompt that asks the model to explain the document in plain English.
• The explanation appears in your preview.

For some lightweight steps (routing requests, repairing malformed responses, generating short summaries) we use a cheaper, faster model - Anthropic's Claude Haiku 4.5.

Where the AI runs

Anthropic's servers are in the United States. When we send your data to them, it crosses the border. We rely on standard contractual clauses approved by the UK Information Commissioner's Office for the transfer.

Anthropic's data processing terms are public at anthropic.com/legal/dpa. The headline point: Anthropic does not use API customer data (your data) to train its models. That is a contractual commitment.

AI can make mistakes

The model is good. Very good. But it is not perfect. Common AI failure modes:

• Wrong section number. The model might cite a slightly wrong sub-section. The general rule is right; the specific reference might need correction.
• Hallucination. Rare with our prompts, but the model might mention a case name that does not exist. If you do not recognise a citation, look it up before sending.
• Missing your facts. If you forgot to mention something important, the model cannot include it. Add it in the edit.
• Tone. The model defaults to firm-but-polite. If you want angrier, gentler, more lawyerly - rewrite.

This is why we always show you a preview and never auto-send.

The Fightingback chatbot

The little chat widget you see on the homepage and inside the app is an AI-powered helper. It does two things: it answers questions about how Fightingback works, and it routes you to the right one of our 26 tools based on what you describe.

It does NOT:

• Give you tailored legal advice on your specific situation.
• Draft letters for you (the tools do that).
• Replace a solicitor, Citizens Advice or any other professional support.

When you chat with the bot, the messages you send go to Anthropic's Claude model (same as for letter generation) along with a system prompt that tells the bot what Fightingback is, what tools exist, and what it should never do.

Safeguarding constraints we have built into the chatbot

The bot will not engage with topics outside its remit. Specifically, the system prompt instructs it to refuse and signpost away from:

• Anything outside UK consumer rights. Off-topic asks (recipes, jokes, code help, roleplay, general chit-chat) are politely declined and the user is redirected to the tool that fits their actual fight, or to Citizens Advice if they have a legal question we do not cover.
• Crisis topics. Self-harm, suicide, eating disorders and similar topics are not engaged. Instead the bot is instructed to suggest the user contact the Samaritans (116 123) or another appropriate UK crisis service.
• Reports of criminal abuse. Reports of sexual assault, domestic abuse or other criminal harm are not handled by our consumer tools. The bot directs the user to the police (999 or 101) and to specialist support services (Refuge, NSPCC, Rape Crisis England & Wales) rather than attempting to draft a "letter".
• Personal venting unrelated to a legal claim. If someone is upset but has no consumer-rights matter to act on, the bot will not roleplay as a counsellor. It signposts to Samaritans or MIND.
• Medical or mental health advice. The bot is explicitly told not to give any. It signposts to NHS 111, GPs, or MIND as appropriate.
• Anything legal we are not built for. Criminal defence, immigration, family law, divorce, child custody, personal injury, conveyancing, will-writing. The bot will refuse and recommend Citizens Advice or a solicitor.

See our Disclaimer page for the full list of crisis and support helplines we point to.

Our "learning" feature

We use a small text embedding model from OpenAI (text-embedding-3-small) to power a feature we call learning: when you generate a new letter, we look for similar successful letters in our history and use them to guide the AI toward your situation.

OpenAI receives a short summary of the letter category (a few hundred characters). They do not receive personal details like names, addresses or PCN numbers. The embedding is a list of numbers that has no human-readable meaning.

OpenAI does not use this data to train their models either.

What you can do if you do not want AI processing

Fightingback cannot work without AI processing. The letters are AI-generated. If you would prefer a human-drafted letter, Citizens Advice and many high-street solicitors offer that service, often free for low-income users.

If you are happy with AI processing for the letter itself but want to take what we wrote and use it as a starting point for your own version, you absolutely can. That is what the edit and copy buttons are for.

Changes to our AI stack

If we swap models or providers in a way that changes what data goes where, we will update this page and email existing customers. The version of this policy in force is shown at the bottom.