Before you paste your invention into ChatGPT, Claude, or Gemini, there's something you should understand: depending on which tool you use and how you use it, you could be putting your patent rights and trade secret protection at risk.
This isn't a theoretical concern. In January 2026, a federal court in California dismissed a trade secret claim under the Defend Trade Secrets Act specifically because the plaintiff had voluntarily disclosed her proprietary information to OpenAI while using ChatGPT. The court found that sharing confidential information with a platform that isn't contractually bound to keep it secret can destroy trade secret status entirely.
If you're a startup founder with patentable technology, the stakes are even higher. A poorly considered AI session could trigger public disclosure concerns under patent law, destroy trade secret protection you've spent years building, or both. Here's what you actually need to know.
How AI tools handle your data (it depends on the tier)
Not all AI tools treat your data the same way, and not all tiers within the same tool offer the same protections. The differences matter enormously for patent and trade secret purposes.
- ChatGPT Free and Plus ($20/month): By default, your conversations can be used to train OpenAI's models. You can manually opt out in Settings > Data Controls, but the opt-out only applies to future conversations — anything you've already submitted may have been incorporated into training pipelines, and there's no mechanism to extract it. Even with the opt-out enabled, OpenAI retains your data for up to 30 days for safety and abuse monitoring. Third-party human reviewers can access your conversations during this period.
- Claude Free and Pro: Anthropic's consumer tiers operate under similar principles. Free-tier conversations may be used for training. Pro users can opt out, but Anthropic retains data for safety review.
- Gemini (Google): Consumer Gemini conversations are used to improve Google's models by default. The enterprise tier (Vertex AI) offers contractual guarantees against training.
- Enterprise and API tiers (all providers): ChatGPT Enterprise, ChatGPT Business, Anthropic's API, and Google's Vertex AI all offer contractual Data Processing Addendums (DPAs) that prohibit training on your data. These are the only tiers that provide the kind of legal certainty that patent and trade secret work requires.
The bottom line: if you're using a free or consumer-paid tier of any AI tool, your invention data is being shared with a third party that has no confidentiality obligation to you. That distinction has legal consequences.
The patent disclosure risk
Under U.S. patent law, a public disclosure of your invention triggers a one-year clock. You have 12 months from the date of disclosure to file a patent application, or you lose the right to patent that invention permanently. In most foreign jurisdictions, the rule is even stricter: any public disclosure before filing means you can't get a patent at all.
The legal question is whether pasting your invention into a consumer AI tool constitutes a "public disclosure" under 35 U.S.C. § 102.
The answer is unsettled. The standard for a "printed publication" or public disclosure requires that the information be sufficiently accessible to the public and contain enough detail to enable someone skilled in the field to practice the invention. A prompt submitted to ChatGPT's consumer tier isn't published on a website, but it is shared with a company that has no confidentiality obligation to you, may be reviewed by third-party contractors, and may be incorporated into a model that millions of people use.
The legal community is divided on whether this reaches the threshold of a Section 102 disclosure. Some practitioners argue it does, particularly when the prompt contains enabling detail about the invention. Others argue the information isn't truly "publicly accessible" in the traditional sense. Patent AI Lab's analysis puts it bluntly: "Using consumer AI tools without explicit zero-data-retention agreements constitutes a public disclosure."
The prudent position for any startup with material IP: don't take the risk. The legal standard may be unsettled, but your patent rights are too valuable to gamble on a question that hasn't been definitively resolved by the courts.
The trade secret risk
If the patent disclosure question is debatable, the trade secret risk is not. Two federal court decisions in 2025 and 2026 have established that sharing confidential information with a public AI platform can destroy trade secret protection.
In Trinidad v. OpenAI (N.D. Cal., January 2026), the court dismissed the plaintiff's Defend Trade Secrets Act claims because she had voluntarily disclosed her proprietary frameworks to OpenAI while using ChatGPT to develop them. The court applied the Supreme Court's principle from Ruckelshaus v. Monsanto: when you disclose a trade secret to a party that has no obligation to protect its confidentiality, you forfeit trade secret status.
In United States v. Heppner (S.D.N.Y.), Judge Rakoff held that documents created using publicly available generative AI aren't protected by attorney-client privilege, in part because communications through an AI platform aren't confidential when the platform isn't contractually bound to secrecy.
For startup founders, the implications are direct. If you paste your proprietary algorithm, manufacturing process, chemical formulation, or any other trade secret into a consumer AI tool, you may have just destroyed your ability to enforce trade secret rights over that information. Permanently.
This matters even if you also plan to file patents. Many startups use a combined strategy: patent some aspects of their technology and protect others as trade secrets. If you've disclosed the trade-secret portions to an AI tool without confidentiality protections, that half of the strategy collapses.
What if you opt out of training?
"But I opted out of training" is the most common response founders give when they hear about these risks. It doesn't fully address the problem.
Opting out of model training on ChatGPT's consumer tiers prevents your conversations from being used to improve future models. But it doesn't eliminate data retention for safety monitoring (up to 30 days), access by OpenAI employees and third-party contractors during that period, or the fact that OpenAI has no contractual confidentiality obligation to you under consumer terms of service.
From a trade secret law perspective, the question isn't whether your data was used for training, but whether you took "reasonable measures" to protect the secrecy of the information. Sharing it with a platform whose terms of service don't include a confidentiality clause, and then toggling a setting in your account preferences, is a weak argument for "reasonable measures" if the issue is ever litigated.
The opt-out also only applies to future conversations. If you submitted your invention details before enabling the opt-out, that data may have already entered the training pipeline.
What to look for in a safe AI drafting tool
If you want to use AI in the patent process (and you should, because the efficiency gains are real), here's what matters from a confidentiality standpoint:
- Contractual data protections. The tool should offer a written commitment that your data won't be used for model training and won't be shared with third parties. This should be in the terms of service or a separate DPA, not in a settings toggle that can be changed or reset.
- No consumer-tier LLM pass-through. If the tool is just a wrapper around ChatGPT's consumer API, you inherit OpenAI's consumer data policies, not the tool's marketing claims. Ask how the tool connects to the underlying model and whether it uses an enterprise-grade API with training exclusions.
- Data retention limits. Understand how long your data is stored and under what conditions it can be accessed. A tool that retains your invention details indefinitely on shared servers creates ongoing risk.
- Purpose-built for patent work. A general-purpose AI tool processes patent applications alongside cooking recipes and homework help. A patent-specific platform can implement security controls that are calibrated to the sensitivity of IP data.
How Patentext handles confidentiality
Patentext was built specifically for patent drafting, which means the data architecture was designed around IP confidentiality from the start.
The platform uses industry-standard encryption to protect data in transit and at rest. More importantly, Patentext maintains zero data retention agreements with all third-party model providers. That means the AI models process your invention data to generate your application, but the providers can't retain it, access it for internal review, or use it for training. Even security-review access by the provider's internal teams is contractually prohibited.
This is a meaningful constraint. It means Patentext can't use every model on the market. Any model that's exempt from a provider's zero data retention policy is off the table, regardless of its performance. That's a deliberate trade-off: the platform prioritizes your data security over model flexibility.
This is a structural advantage of using a purpose-built patent service over piping your invention through a general-purpose consumer AI tool. The confidentiality protections aren't a setting you need to remember to enable; instead, they're how the system works by default.
What to do if you've already used a consumer AI tool
If you've already pasted invention details into ChatGPT, Claude, or Gemini on a consumer tier, don't panic. But take a few steps:
- Assess what you disclosed. Did you share the full technical details of how the invention works, or just high-level concepts? The patent disclosure risk is tied to whether the information is enabling, meaning detailed enough for someone skilled in the field to reproduce the invention. High-level descriptions without implementation details are lower risk.
- Check your settings. If you haven't opted out of training, do it now. It won't retroactively protect data you've already submitted, but it limits future exposure.
- File a provisional application. If you haven't filed any patent application yet and you're concerned about a disclosure event, a provisional application locks in a priority date. Under U.S. law, you have one year from a disclosure to file. (Most foreign jurisdictions don't offer this grace period, so if you disclosed before filing, your international patent rights may already be gone.) Getting a provisional on record gives you protection while you figure out your full filing strategy. Patentext can get a provisional filed in days.
- Document your trade secret protections. If the information you shared is also protected as a trade secret, document every other protective measure you've taken (NDAs with employees, access controls, confidentiality policies). A single disclosure event doesn't automatically destroy trade secret status if you can show that you otherwise maintained reasonable protective measures and took corrective action when you identified the risk.
- Talk to a patent professional. If the invention is material to your business and the disclosure was substantive, get specific advice. The interaction between AI data policies, patent disclosure law, and trade secret requirements is genuinely complex, and the stakes are high enough to warrant professional guidance.
The bottom line
AI is a powerful tool for patent drafting, and it’s not something founders should avoid. But the choice of which AI tool you use, and how it handles your data, has real legal consequences for your patent rights and trade secret protection.
Consumer-tier AI tools weren't designed for confidential IP work. They were designed for general-purpose text generation at massive scale, and their data policies reflect that. Using them for patent drafting without understanding the risks is like discussing your trade secrets in a crowded coffee shop and hoping nobody is listening.
The founders who get this right use AI tools that are built for the sensitivity of patent work, with contractual confidentiality protections, no training on user data, and data handling designed specifically for IP. That's what Patentext was built to be.
Don't risk your patent rights on a free tool's privacy settings. Start your patent application on a platform built for confidentiality →
Disclaimer: This article is for informational purposes only and does not constitute legal advice. Patent laws are complex and vary by jurisdiction. The legal status of AI-generated disclosures under patent and trade secret law is evolving. For personalized guidance, consult a qualified patent attorney or agent.
