Not directly. ChatGPT and Claude cannot file a patent application, because patent offices only accept applications signed and submitted by the applicant or a registered attorney or agent, through portals such as USPTO Patent Center or EPO Online Filing. What you can do from inside the chat is everything that leads up to the filing: describe the invention, turn it into a structured invention disclosure, check it for gaps and hand it to a patent attorney.

With a patent workflow connector such as Lightbringer MCP, that hand-off happens without leaving Claude or ChatGPT. You create the disclosure in conversation, submit it for attorney review, read the attorney's comments and approve changes in the same chat, and Lightbringer's registered patent attorneys draft and file the application, typically within days rather than months. The platform and the connector are free; you pay a fixed, published price per application only when you decide to file. Before you paste an unfiled invention into any assistant, check its data-training settings first. See Does pasting my invention into ChatGPT or Claude count as public disclosure?

ChatGPT can produce text that looks like a patent application, but not one you should file. A general-purpose model does not search the prior art, cannot judge how broad a claim can be before an examiner rejects it, and will invent plausible-looking citations and references. It also has no view of what your competitors have filed, which is what decides whether a claim is worth anything commercially. The typical output is either claims so broad they fail novelty or so narrow they are easy to design around.

The rules reflect this. The USPTO's April 2024 guidance allows AI-assisted drafting but holds the person who signs the filing responsible for verifying every statement in it, and says that relying on an AI tool's output does not count as a reasonable inquiry. Under the USPTO's revised inventorship guidance of November 2025, AI is a tool, not an inventor: a natural person must have conceived what is claimed. Where ChatGPT and Claude are useful is the step before drafting, turning a design document or a conversation into a structured invention disclosure that states the problem, the solution, the alternatives and why it differs from what exists. That is the input a patent attorney needs. Lightbringer uses AI for exactly this, and for first drafts, with a registered patent attorney reviewing and signing every application before it is filed.

Whether it legally counts as a public disclosure has not been tested in court, but most patent attorneys treat pasting an unfiled invention into a consumer chatbot as a risk you should not take. A public disclosure is a disclosure to anyone not bound by a duty of confidentiality, and the standard consumer terms of ChatGPT and Claude do not create such a duty towards you. On ChatGPT Free, Plus and Pro, conversations are used to train OpenAI's models by default unless you switch off Improve the model for everyone in Data Controls. On Claude Free, Pro and Max, chats are used for training when the Help improve Claude setting is on; Anthropic's policy states that raw content read through connectors and MCP servers is excluded, but anything you type or paste into the chat is not.

The consequences differ by region. Europe and most of the world apply absolute novelty under Article 54 EPC, so a disclosure that counts as public before your filing date destroys novelty, with no grace period. The US gives inventors a one-year grace period for their own disclosures under 35 U.S.C. 102(b)(1). A separate risk is trade secret status, which depends on taking reasonable steps to keep the information confidential. Safer options: use business tiers that do not train on your data by default (ChatGPT Team and Enterprise, Claude for Work and Enterprise, or the APIs), use a Temporary Chat in ChatGPT or an Incognito chat in Claude, keep the specifics vague until you have filed, or file a priority application first and then work openly. Lightbringer's own advice on its connector page is the same: check your assistant's data controls before connecting sensitive, pre-filing invention details.

You file a patent from Claude by connecting it to a patent platform with registered attorneys, then running the disclosure, review and approval steps in the chat while the attorney drafts and files. With Lightbringer MCP the sequence is:

1. Create a free account at app.lightbringer.com. An administrator turns on MCP access under the cog menu, Advanced Settings, Enable MCP access for this organization.
2. In Claude (web or desktop) open Customize, then Connectors, click Add, then Browse connectors, search for Lightbringer and click Connect. Sign in with email, Google or Microsoft, approve read and write access and select your organisation.
3. Ask Claude who you are on Lightbringer to confirm the connection works.
4. Describe the invention, or paste a design document, pull request or meeting notes. Claude runs a guided interview against your organisation's disclosure template and creates the invention disclosure.
5. Ask Claude what is missing. Lightbringer scores the draft for clarity, completeness and technical depth: is the technical problem defined, are alternative embodiments described, what will a reviewer ask. Revise in the chat.
6. Submit the disclosure for attorney review. Claude asks you to confirm before any action that creates, changes or submits data.
7. A Lightbringer patent attorney drafts the application in Lightbringer. You read the attorney's comments and redlines in Claude, reply to specific passages, and approve or request changes.
8. The attorney files the application. Lightbringer files in any jurisdiction and typically gets from disclosure to filing in days.

In Claude Code, run /plugin marketplace add lightbringer-patents/claude-plugin, then /plugin install lightbringer@lightbringer, then /mcp to authenticate. The plugin adds a skill that mines your repositories, documents and tickets for patentable problem and solution pairs and drafts the disclosure. The platform and connector are free; you pay a fixed price per application when you file. Full setup notes: Lightbringer MCP.

You file a patent from ChatGPT the same way as from Claude: connect ChatGPT to a patent platform whose attorneys do the filing, then run the disclosure, review and approval steps in the chat. With Lightbringer:

1. Sign up free at app.lightbringer.com and have an administrator enable MCP access for your organisation under Advanced Settings.
2. In ChatGPT, search for Lightbringer in the apps and plugins directory, click Install, sign in with email, Google or Microsoft, and select your organisation. In OpenAI Codex or any other MCP client, add https://mcp.lightbringer.com/mcp as a remote MCP server with OAuth.
3. Confirm the connection by asking who you are on Lightbringer.
4. Describe the invention or paste your technical notes. ChatGPT creates the invention disclosure against your template, then checks it against patent review criteria for missing detail, unclear passages and an undefined technical problem.
5. Submit the disclosure for attorney review. Write actions are confirmed with you first.
6. Read attorney comments and redlines, answer questions and approve changes from ChatGPT. Lightbringer's registered patent attorneys draft and file the application.

One difference matters on ChatGPT specifically. On Free, Plus and Pro plans, conversations train OpenAI's models by default; switch off Improve the model for everyone in Data Controls, use a Temporary Chat, or use a Team or Enterprise workspace before discussing an unfiled invention. Business plans and the API do not train on your data by default. Setup details: Lightbringer MCP.

A patent MCP connector is a server built on the Model Context Protocol (MCP), the open standard that lets AI assistants such as Claude, ChatGPT, Cursor and Windsurf call external tools, and it gives the assistant permissioned access to a patent system. There are two kinds. Patent search connectors give an assistant read access to public patent databases such as the EPO and USPTO registers, so it can look up documents instead of guessing. Patent workflow connectors, such as Lightbringer MCP, give an assistant read and write access to your own invention disclosures, attorney reviews and comments, scoped to one organisation, so it can create a disclosure, run a patent-grade review, submit it to attorneys and carry the attorney conversation back into the chat.

Search tells you what exists; a workflow connector helps you file. Two questions to ask any provider before connecting an unfiled invention: who operates the server, and what do they log. Third-party MCP servers see every tool call, a point IPKat raised in August 2026. Lightbringer operates its own endpoint at mcp.lightbringer.com, is SOC 2 Type 2 certified, limits access to the organisation you select at sign-in, and confirms every create, update, submit or comment action with you before it runs. Anything that changes data waits for your approval; reads run freely. Details: Lightbringer MCP.

ChatGPT can help you plan a prior art search, but it is not a patent database and its output is not a search. A language model has no reliable, current index of patent literature, so it fills gaps with plausible-looking patent numbers, titles and applicants that do not exist, and it cannot tell you what it missed. With web browsing switched on it can retrieve real documents, but every reference still has to be checked at the source before you rely on it. The USPTO's 2024 guidance on AI tools makes that verification the practitioner's responsibility, not the tool's.

Used well, ChatGPT or Claude is a search assistant: it can generate synonyms and alternative terms for your invention, suggest CPC and IPC classification codes to browse, draft Boolean search strings, and summarise documents you paste in. Run the actual search in Google Patents, USPTO Patent Public Search, Espacenet or WIPO PATENTSCOPE, all free. Keep the confidentiality point in mind: describing your invention in enough detail to search for it is describing your invention, so use an assistant that does not train on your prompts. Lightbringer runs prior art searching inside the platform as part of the disclosure review, before an attorney sees the case. See also What is prior art search?

In the US, no, not as a general rule. The USPTO's April 2024 guidance (89 FR 25609) states there is no general duty to inform the office that an AI tool was used to prepare an application. Existing duties still apply in full: the duty of candour, the duty to disclose information material to patentability, and the rule that a signature certifies the signer made a reasonable inquiry into everything in the paper. AI use becomes disclosable when it is material, for example if an AI system, rather than a natural person, contributed the subject matter of a claim. Such a claim is also unpatentable: the USPTO's revised inventorship guidance of 28 November 2025, which rescinded the February 2024 guidance, treats AI as a tool like software or laboratory equipment and requires that a natural person conceived each claimed invention.

Europe and the UK have no rule requiring you to declare AI drafting tools either, but the inventor must be a natural person, as the EPO's Boards of Appeal and the UK Supreme Court confirmed in the DABUS cases. The practical answer is to keep records of who conceived what, so that inventorship is defensible if challenged, and to have a registered attorney review and sign the filing, because that person carries the verification duty regardless of which tools were used. See also Can AI invent a patent, or be listed as an inventor?

The best AI for patent drafting depends on who is holding the pen. There are three categories. General-purpose assistants (ChatGPT, Claude, Gemini) are good at structuring an invention disclosure and explaining the process, but have no patent database, no attorney, and consumer data-training defaults you must switch off; none of them should produce the claims you file. Attorney-facing drafting copilots (DeepIP, Solve Intelligence, Patlytics, Qatent, Edge by Patsnap, Rowan Patents) are built for patent attorneys and in-house IP teams, drafting inside Word or a browser editor with prior art and office action support; they make a good attorney faster, but you still need to hire and pay the attorney, usually by the hour. AI-native patent firms such as Lightbringer bundle the AI drafting with registered patent attorneys who review and sign the filing, at a fixed price per application.

For a founder or engineer without an IP team, only the third category ends with a filed application, since a copilot without an attorney produces a document, not a patent. For a law firm, the second category is the relevant comparison. Whatever you choose, apply three tests: does it search the prior art rather than guess, does a named registered attorney sign what gets filed, and where does your unfiled invention data go. Lightbringer's answers are yes, yes, and a SOC 2 Type 2 certified platform with strict access controls. Pricing is published on the pricing page.

AI is replacing most of the hours in patent work, not the attorney. Drafting, prior art search, invention capture and portfolio monitoring, the tasks that used to consume weeks of billable time, are now largely done by AI. What has not changed is who is accountable. Patent offices accept applications only from the applicant or a registered attorney or agent, the USPTO's 2024 guidance makes the human signer responsible for verifying everything an AI tool produced, and only a natural person can be an inventor. Nobody can sue an algorithm for malpractice, and no algorithm can decide that a claim should be narrowed to survive examination but kept broad enough to matter commercially, which is the judgement that determines a patent's value.

What AI does replace is the business model. When drafting takes hours instead of weeks, hourly billing stops making sense, which is why AI-native firms charge a fixed price per application. Lightbringer, founded by a patent attorney with 20 years in practice, cut the time from disclosure to filing from around two months to a couple of days and cut cost by up to half, and still has a registered patent attorney review and sign every application. The shortage that matters is not of attorneys in general but of attorneys who understand quantum computing, novel materials or advanced AI architectures; AI that captures the technical detail well makes those specialists more available, not less necessary. See also Do I need a patent attorney to file?