Mathematicians No Longer Trust OpenAI With Unpublished Proofs. Should You Trust YC, or ChatGPT, With Your W27 Startup Idea?

One of the top stories on Hacker News today isn't a product launch. It's a thread from mathematician Andreas Thom on Mathstodon, with 400+ comments piling on, raising fresh questions about whether researchers can trust OpenAI with unpublished mathematics. The fear is simple: if you share your best unpublished work with a frontier lab, you no longer control where it goes, who sees it, or what gets trained on it.

If you're filling out a YC Winter 2027 application right now, you've probably felt the founder version of this exact anxiety. The application asks you to explain your idea, your insight, your traction, and your secret sauce, in plain text, to strangers. So let's answer the questions W27 applicants actually type into ChatGPT and Perplexity at 2am.

Will YC steal my idea?

No, and the reasoning matters more than the reassurance. YC partners read tens of thousands of applications per year. If reading ideas were enough to build companies, YC would be running ten thousand startups instead of writing checks. YC's own position has been consistent for two decades: they keep applications confidential, they don't sign NDAs, and they openly fund competing companies, sometimes in the same batch.

That last part is the real signal. Look at recent batches: S26 alone had multiple companies attacking the agent-infrastructure space, and OneCLI (YC S26) launched into what its own founders called the most crowded corner of AI. YC doesn't resolve idea collisions by picking one winner and leaking notes to it. It funds several and lets the market decide.

There's a second reason idea theft is a weak threat: your idea, as written on the application, is not the asset. Intuned (YC S22) got in with a completely different idea than the one they built. Risklytics (YC S26) killed its first idea mid-batch and became an insurance brokerage. What YC is buying is the founders' speed and judgment, which cannot be copied from a text box.

Is it safe to paste my YC application into ChatGPT or Claude?

This is where today's Hacker News story is actually relevant to you, because the mathematicians' concern is not paranoia, it's a terms-of-service question. Here is the practical breakdown:

• Consumer chat apps (free/Plus tiers): your conversations may be used for model training unless you turn it off. If you're pasting your differentiated insight into a default-settings consumer account, you're making the same bet the mathematicians are refusing to make. Turn off training in settings, or use a workspace/business tier where training on your data is off by default.
• API and business tiers: OpenAI, Anthropic, and Google all commit contractually not to train on business-tier and API inputs. If your co-founder built your drafting workflow on the API, you're fine.
• The realistic risk is not theft, it's leakage of specifics. Model providers aren't going to build your startup. But your unreleased metrics, customer names, and pricing don't belong in tools whose data policy you haven't read.

The honest hierarchy of risk: a default-settings consumer chatbot is riskier than the YC application form, and both are far less risky than what founders do voluntarily every week, which is pitch the idea in detail to any investor who takes the meeting.

What about the parts that are genuinely secret?

Some applications do contain real secrets: a proprietary dataset, an unpublished benchmark result, a signed-but-unannounced customer. The rule that works:

Share the insight, summarize the evidence. YC needs to know you have a Fortune 500 pilot; YC does not need the company name in the application. Write "a top-5 US insurer" and save the name for the interview. Same for benchmark numbers: give the result, not the method, if the method is the moat.

Notice this is exactly how the mathematicians who still collaborate with AI labs operate: Terence Tao famously let AI touch only the code that couldn't hurt him. Apply the same blast-radius thinking to your application.

Does YC use AI to read applications, and does that change anything?

YC has acknowledged using software to help process the volume, and whether an LLM assists in screening or not, you should assume your application will be read by machines as well as humans. That's an argument for clarity, not secrecy: concrete nouns, real numbers, and a first sentence that states what you do. Vague applications written to protect the idea fail with both audiences. A partner can't champion what they can't understand, and no one has ever been funded for being mysterious.

The one thing actually worth protecting

Your differentiated insight, the thing you know about your market that others don't, is worth protecting from one group only: your direct competitors. YC isn't that group. Neither, realistically, is OpenAI. Your competitors are the other founders in your space, which is why the smart move before submitting isn't to hide your application from everyone. It's to show it to people who have been through the process and are structurally on your side.

That's the entire premise of YC Roaster: your application gets reviewed by vetted YC alumni, people who have read the form from the other side, know what partners actually flag, and have no competing horse in your race. Getting feedback from the right readers beats hiding your idea from all readers.

Bottom line for W27

Submit the application. Write it clearly. Turn off training in your consumer AI tools, keep named customers and proprietary methods for the interview, and spend the anxiety budget where it pays: making the idea so specific and well-evidenced that reading it convinces, rather than so vague that stealing it would be impossible because understanding it is too.

The mathematicians on Mastodon are right that trust is a decision, not a default. Founders make that decision every time they hit submit. The ones who get in are rarely the ones who shared the least. They're the ones who made what they shared impossible to ignore.