An e-shop rejected your complaint, a phone operator charged you for a service you did not want, and you open ChatGPT. Within half a minute you have a complaint that sounds confident and even contains sections. A chatbot writes very well. We write letters with a language model too. The difference is what happens before the writing and after it.
Do you know what to ask?
You type in a query and get an answer. Now the questions you have to ask yourself.
Did you take the answer as correct straight away, or do you know what to check in it?
Does the cited section exist in that form? A type of error you will not spot at first glance: "under Section 2161(1)(d) of the Civil Code". Section 2161 governs the seller’s liability for defects, so the citation sounds right. But subsection 1 ends with letter (c), and letter (d) is only in subsection 2. You will not notice; the trader will.
Did it load the current terms and conditions of that particular company? And the ones that applied on the day of the purchase, not last year’s?
What deadline is running for you, and by when does the company have to reply?
With ten further rounds of follow-up questions you will get to an excellent complaint with the chatbot. But you have to know what to ask, and spot where a fact, a deadline or a condition is missing from the answer. Anyone who can do that usually does not need a chatbot for the letter.
On top of that, there are the figures from 2025 and 2026. On 17 September 2026 OpenAI announced Astra for Law: the GPT-6 Astra model with a legal search index of over 230 million web addresses, from US law only. So far, selected law firms have been given access. According to OpenAI’s own measurement (200 legal research questions), the model was completely correct on 54% of the questions, ordinary Astra with web search on 38.7%. OpenAI itself recommends verifying the answer and the sources before use.
The independent Vals Legal Research ranking, as of 18 October 2026, takes roughly 235 research tasks from US law. A task counts only when it is solved in full. The best models, Claude Opus 5, Claude Fable 5.1 and Meta Muse Spark 1.3 Max, fully solve 55% of the tasks, GPT-6 Astra 39%.
Those are US research tasks. The LEXam study (ICLR 2026) tested models on 7,537 questions from Swiss bar exams. All models were worse in German than in English, and worse in Swiss law than in general or international law. We did not find a comparable measurement for the Czech language and Czech law.
In September 2025 the British consumer organization Which? put 40 consumer questions to six AI tools and published the results in November. ChatGPT got 64% right. Copilot claimed that a canceled flight always gives a right to full compensation, and Gemini advised withholding payment from a tradesperson.
Courts have already dealt with invented citations. In December 2025 the Constitutional Court (Ústavní soud) fined a lawyer CZK 25,000 (I. ÚS 3004/25). His constitutional complaint cited twelve decisions, a substantial part of which did not exist. According to reports on the case, the lawyer used a paid specialist legal information system with artificial intelligence. In October 2025 the Supreme Administrative Court (Nejvyšší správní soud) refused to reimburse the costs of submissions that "very probably" were produced with AI and cited decisions that did not exist or did not contain what was claimed (3 As 34/2025-80).
As of 19 October 2026, Damien Charlotin’s database records 2,202 court cases with content invented by AI. 57.5% of them concern people who represented themselves, without a lawyer. In April 2026 Sullivan & Cromwell, one of the largest American law firms, apologized to a court for a motion with invented citations. The firm’s own rule, "trust nothing, verify everything", was not followed in that submission. A court in Oregon imposed sanctions of roughly 110,000 dollars for 15 non-existent decisions and 8 invented citations, and the action was dismissed with final effect.
If a lawyer with a paid tool misses the error, a consumer with a free chatbot has an even smaller chance of spotting it.
How the letter is made with us
The form leads to the essential details: the date and the document number, the amount, what happened and what you are asking for. From an uploaded fine, decision or notice we extract the authority, the file number, the amount and the deadline. Before the letter is created, you check the summary and confirm it.
Many disputes, though, are decided by what the company wrote into its terms and conditions or price list: the deadline for returning the money, the address for a complaint, the fee for early termination of the contract. A chatbot either does not have these documents, or finds some version on the web and cannot tell whether it applied on the day of your purchase. We have the terms and conditions of a number of traders in our database, and if yours are not there, you can upload them. The letter then also relies on a point the trader committed to itself. The model does not write from memory, but from the texts of the statutes in their current wording and from the terms and conditions it receives directly in the task.
After the writing, we compare the citations in the letter with what the model actually received from the database. In the latest internal sample of 55 letters, 523 of 553 citations were based on these materials, 30 could not be verified by machine, and none lay outside them. The check watches what the letter cites from, not whether the claim succeeds. With each letter you see a list of sources.
With the letter you see the deadline that is running for you. After it is sent, we watch the deadline for a reply, and if the other side stays silent or refuses, we prepare a follow-up letter.
We compare language models with each other on real Czech complaints. We change the model that writes the letters according to the results. You will find the results in the article Which AI writes the best complaint?
Under the earlier wording of the Civil Code, it was interpreted that you can withdraw from the contract only when the same defect appears for the third time. An amendment effective from 6 January 2023 (Act No. 374/2022 Coll.) expressly states a repeated manifestation of the defect as one of the grounds for a price reduction or withdrawal (Section 2171 of the Civil Code). The claim depends on the statutory conditions and the circumstances; the number of occurrences does not decide it by itself. But the old interpretation still hangs in thousands of texts the models learned from, and a chatbot may repeat it without warning that the law has changed.
Chatbots also tend to agree with the person who is asking. In April 2025 OpenAI rolled back a ChatGPT update because, in its own words, it was excessively flattering. A chatbot can thus support even a demand the law does not provide for.
And a conversation does not watch the deadlines for you. You must complain about a phone operator’s bill no later than two months from its delivery (Section 64(8) of the Electronic Communications Act). The seller must deal with a complaint about goods and inform you of it within 30 days of the complaint being made, unless the seller agrees a longer deadline with you. A conversation with a chatbot does not track deadlines. A month later you have to explain the whole situation to it again.
OpenAI’s usage policies effective from 29 October 2025 do not allow its services to be used for individual advice that requires a license, for example legal advice, without the appropriate involvement of a licensed professional. That does not mean a ban on help with drafting a letter. But responsibility for the content of the letter lies with the person who sends it.
When a general chatbot is enough
When you need to get your bearings: what a complaint is, what the difference is between a guarantee and statutory rights arising from defective performance, how to reply politely to an email from an e-shop. Or for a request for goodwill, where you are not enforcing any claim.
If you use it for a real claim:
- Look up every section on zakonyprolidi.cz, including the subsection and the letter.
- Give it all the facts: dates, document numbers, amounts and the communication so far.
- Do not enter a birth number or account numbers into the chat. In the consumer versions of ChatGPT, conversations may by default be used to train models until you turn this off in the settings.
- Write the deadlines into your calendar.
What to do next
Describe your problem to Legamik in your own words and upload what you have: a receipt, a fine, terms and conditions. Legamik will prepare a letter from that, which you check and send yourself. It checks the citations against its database. It shows you the deadline, watches the time for a reply, and if the trader does not respond, it prepares a follow-up letter.
Legamik is not a law firm and does not provide legal advice. In complex disputes we recommend consulting a lawyer.