How well can GPT-5 draft patents? A real-world experiment

OpenAI introduced GPT-5 on August 7, 2025, calling it "a significant leap in intelligence over all our previous models." It's trained to reason more effectively, follow complex instructions, and adapt to specialized domains — in other words, the kind of upgrades that sound promising for a detail-heavy, language-precise task like patent drafting.

So, we decided to put it to the test. Could GPT-5 turn an inventor's technical disclosure into a patent draft that's accurate, compliant, and worth building on? Or will it trip over the same hurdles as its predecessors?

Our invention: the magnetic tessellating umbrella

Here at Patentext, there's one made-up invention that shows up in our screenshots, our product walkthroughs, and now, in this blog: the famed magnetic tessellating umbrella. It's our go-to fictional example for testing because it's silly but still technical enough to mimic the kind of real-world inventions our users work with.

This is the plain-English disclosure we fed ChatGPT:

Imagine standing in a rainy crowd where everyone has an umbrella. Normally, there are gaps between umbrellas where the rain still gets through, and the edges bump into each other awkwardly. The magnetic tessellating umbrella solves that by giving each umbrella a perfectly shaped canopy (triangle, square, or hexagon) and embedding magnets along the edges. These magnets snap together with the umbrellas next to them, forming a continuous, watertight "umbrella wall."

The magnets can spin, so they always connect with the right polarity. And if it's windy or you want a stronger connection, you can use optional clips at the edges to lock umbrellas together. You can also set the umbrella on a tripod base so you don't have to hold it, and the pole can extend higher so people can walk underneath.

Our rubric for scoring GPT-5's patent drafting skills

We didn't just want to see what GPT-5 spit out — we wanted to measure how well it actually did the job of a patent drafter.

I'm a patent agent with 10+ years of drafting experience, having worked at both major law firms and boutique firms. I've drafted patents for Airbnb, Facebook, and dozens of startups. I'll be scoring each output on a 1 to 5 scale:

• 5 = Excellent – On par with a competent human drafter.
• 4 = Strong – Minor flaws that wouldn't meaningfully harm a filing.
• 3 = Mixed – Some good elements, but also notable gaps or errors.
• 2 = Weak – Significant issues that would require major rewriting.
• 1 = Poor – Fundamentally unusable without starting over.

How well can GPT-5 draft patents?

To keep this test fair and reproducible, we'll be using the magnetic tessellating umbrella invention for every run. For each test, we'll be feeding GPT-5 only the plain-English disclosure — not pre-written claims, legal language, or section headings.

Test 1: Can ChatGPT write broad independent claims?

Goal: See if GPT-5 can take a layperson description and produce a single, legally structured claim that captures the core inventive concept without unnecessary limitations.

Prompt: You are a U.S. patent attorney. Write an independent claim that balances broad scope with high patentability for the following invention. Follow proper U.S. claim format. Avoid unnecessary limitations.

An umbrella system comprising:
a canopy having a planform shape configured to tessellate with canopies of adjacent umbrellas without leaving gaps;
a plurality of magnetic coupling elements disposed along edges of the canopy, the magnetic coupling elements being configured to magnetically engage with complementary magnetic coupling elements of an adjacent umbrella; and
a mounting structure coupled to the canopy, the mounting structure including a support pole,
wherein the magnetic coupling elements are mounted to allow relative rotation about at least one axis to align magnetic polarities for engagement,
and wherein engagement of adjacent umbrellas forms a substantially continuous weather-resistant barrier.

My feedback:

• Overly narrow language: Phrases like "without leaving gaps" or "substantially continuous weather-resistant barrier" can create loopholes. If there's even a small gap in a competitor's umbrella, they might escape infringement.
• Awkward structure: The magnets are introduced before the structural elements they attach to, forcing the claim to loop back awkwardly via "wherein" clauses. Canopy → mounting structure → magnets would read more cleanly.
• Weaker phrasing: Using "having" instead of stronger patent terms like "comprising" can make the claim feel less precise.
• Extra filler words: "Being configured" is wordy. "Configured to" means the same thing.
• Unnecessary quantity limits: "Plurality" means "more than one." If an infringer only uses one magnetic coupling element, they wouldn't infringe under this wording.

Score: 2/5. GPT-5 captured the general idea, but unnecessary limitations, suboptimal language, and awkward structure weaken the claim's enforceability.

Test 2: Can ChatGPT write dependent claims?

Goal: Evaluate GPT-5's ability to add refinements and variations that meaningfully narrow the scope while staying consistent with the disclosure.

Prompt: You are a U.S. patent attorney. Write 8 to 10 dependent claims for the following invention, adding narrower features, variants, and refinements described in the disclosure.

My feedback:

• Formality issues: Claim 3 has an "antecedent basis" error — a term is used without being properly introduced first.
• Problematic use of "or": While technically allowed, "or" is not recommended because it reduces patentability — the examiner only has to find prior art for one of the alternatives. Here, it doesn't add meaningful breadth, so it just weakens the claim.
• Overcomplication: The independent claim already introduced a "mounting structure" to which the magnets could be "rotatably attached." Instead, GPT-5 created a brand-new "rotatable housing" in Claim 4, adding complexity without real value — and it's redundant because rotation was already addressed in the main claim.
• Lack of novelty: Claims 7 to 10 describe very basic umbrella features (telescoping pole, height adjustment, tripod base) that are not new, interesting, or patentable.

Score: 2.5/5. Structurally valid, but several claims were redundant and unnecessarily complicated. The dependent claims taken together don't provide much in the way of backup positions for prosecution.

Test 3: Can ChatGPT write a detailed description for a dependent claim?

Goal: Assess whether GPT-5 can expand one dependent claim into a thorough, compliant Detailed Description section that provides sufficient written description and enablement.

Prompt: You are a U.S. patent attorney. Below is a dependent claim. Write a Detailed Description section for a U.S. utility patent application that supports this claim. The description should explain the feature in depth, provide example embodiments, and stay consistent with the overall invention.

My feedback:

• Impractical fastening examples: Listing rivets, brackets, and clamps could technically work, but it's clunky and impractical for this product.
• Questionable rotation description: "Rotate about an axis perpendicular to the canopy plane" makes little sense mechanically and would be hard to execute.
• Disorganized flow: The section on "dimensions and magnetic field strength" is tacked on after the materials discussion. Grouping related points together would make the description more cohesive.
• Inconsistent terminology: The draft switches between "embodiments," "implementations," and "variations" without clarifying whether they mean different things or are interchangeable.

Score: 3.5/5. A solid attempt that hits a lot of detail, but it stumbles on mechanical plausibility, structure, and consistent terminology. It feels like GPT-5 was trying to sound "patent-y" and ended up overengineering both the language and the hardware.

Test 4: Can ChatGPT write a patent abstract?

Goal: See if GPT-5 can produce a concise, compliant abstract that captures the essence of the invention in ~150 words.

Prompt: You are a U.S. patent attorney. Write a 150-word abstract for the following invention, following USPTO abstract conventions.

My feedback:

• Meets length requirement: At under 150 words, it stays within USPTO guidelines for abstracts.
• Too much detail: An abstract should give a high-level overview, but this one dives a little deeper into specifics than necessary. Keeping it broader would make it cleaner and more in line with typical patent abstracts.

Score: 4/5. A solid, compliant abstract that just needs a lighter touch on the detail to hit the ideal balance between informative and concise. For a full guide to what goes into a compliant abstract, see how to write a patent abstract.

My overall thoughts

Running this experiment made it clear that GPT-5 does have some genuine strengths. It can hit word count targets, throw in plenty of technical detail, and occasionally phrase things in a way that feels impressively "patent-y." For a broader view of where AI patent tools stand today, see the state of AI patent drafting tools in 2026.

That said, its biggest weakness in this experiment was document-level planning. Even when we asked for discrete sections, GPT-5 jumped between topics, repeated ideas, and introduced terminology inconsistently. Those problems become harder to manage when a user asks a general-purpose chatbot to draft multiple interdependent sections without a structured model of the invention.

This is also a risk in chat-based patent drafting workflows more broadly: the quality of each output depends heavily on the context supplied, the sequence of prompts, and the user's ability to detect when the model has drifted.

It is also important to understand the confidentiality implications of using a consumer AI account for invention material.

Data-use, retention, and model-training terms vary by product, plan, and user settings. Before entering confidential invention information into ChatGPT or another general-purpose model, confirm whether the account is approved for company use, whether model-improvement sharing is disabled, how long the data is retained, and who may access it.

The concern is not that every chatbot interaction automatically becomes a public disclosure. However, mishandling confidential invention information can create avoidable confidentiality, ownership, contractual, and filing risks. Companies should use approved systems and review the relevant terms before uploading sensitive technical material. For more on how Patentext handles IP data, see our security practices and our overview of AI patent drafting and confidentiality.

What this experiment says about the patent workflow

GPT-5 successfully created a compliant abstract, organized technical content, generated plausible claim language, and produced a detailed description that was often directionally helpful. The problem was that its most convincing output still required someone with patent experience to identify:

• Unnecessary claim limitations
• Weak fallback positions
• Unsupported or impractical technical details
• Inconsistent terminology
• Structural problems across sections
• Features that sounded technical but added little patent value

The right lesson is not simply that companies need a better chatbot. Rather, a strong patent application depends on the full process surrounding the draft:

1. Identifying which technical work may be worth protecting
2. Capturing the invention accurately
3. Resolving missing technical details
4. Evaluating business and portfolio relevance
5. Developing a filing and claim strategy
6. Drafting the application
7. Reviewing technical accuracy and legal structure
8. Managing the resulting filing and portfolio

General-purpose AI can assist with parts of that process, but it does not own the process. If you're an inventor or technical team evaluating where AI fits, see our guide on how to use AI for patent drafting as an inventor.

How Patentext works differently

Patentext is an end-to-end patent platform for technical companies building and managing patent portfolios. The platform helps teams identify potential inventions, capture structured technical context, evaluate what is worth protecting, and manage filing decisions. When an invention is approved to move forward, USPTO-registered patent agents use the context captured in Patentext and an AI-enabled workflow to prepare the application.

That means the drafting process does not begin with a blank chat window and a one-paragraph prompt. It begins with:

• A structured invention record
• Confirmed technical details
• Defined alternatives and embodiments
• Documented business and portfolio context
• A registered patent agent responsible for the application

Current Patentext pricing

Patentext offers a free starter plan, with paid platform plans starting at $30 per month. Service prices (which include standard USPTO filing fees) are:

• $2,500 for a provisional patent application
• $5,000 for a non-provisional patent application
• $2,000 for a continuation application
• $2,000 for an office action response

Explore the Patentext platform · See patent application services · view pricing

Final verdict: how well can GPT-5 draft a patent?

In this experiment, GPT-5 performed best on bounded, lower-strategy tasks. It produced a strong abstract, a useful first pass at detailed descriptive content, and claims that looked structurally plausible. But the more strategic the task became, the more consequential the weaknesses were.

Its independent claim was unnecessarily narrow. Its dependent claims offered weak prosecution fallback positions. Its detailed description introduced impractical mechanics and inconsistent terminology. Even where the output sounded polished, professional review revealed issues that a non-specialist could easily miss.

So, can GPT-5 draft patent content? Yes. Can it independently produce a strong, filing-ready patent application from a plain-English disclosure? This experiment suggests no.

The most useful role for GPT-5 is as an assistant within a structured patent workflow — not as the unaccountable owner of claim strategy, technical completeness, or the final application. For more on what a strong drafting process looks like, see our guide on how to develop a strong patent drafting style.

Disclaimer: This article is for informational purposes only and does not constitute legal advice. Patent laws are complex and vary by jurisdiction. For personalized guidance, consult a qualified patent attorney or registered patent agent.