Your AI will tell you your idea is good. Here is why we built the opposite.

If you are a founder who needs motivation, encouragement, and a reason to believe your idea will work, ask ChatGPT or Claude. They are very good at that. If you are ready to hear the uncomfortable truth about your idea, the market you are walking into, and your own fit, read on. And yes, we use AI too, and our models are not immune to flattering you either. What is different is that ours are never asked the question that invites the flattery, and everything they do say is backed by real data you can check.

The mechanism has a name: sycophancy.

Sycophancy is the technical term for an AI tailoring its answer to what you already believe, even when you are wrong. It is not a bug someone forgot to fix. It comes from how the models are trained: on human feedback. People rate the answers; the model learns to produce answers people rate well. In 2023 a team at Anthropic measured what people actually rate well. When a response matched the user's own view, it was more likely to be preferred. Worse, both humans and AI models built from their ratings picked a convincingly written flattering answer over a correct one often enough to matter [1]. Agreement became a proxy for quality, and the model learned the proxy.

If that sounds abstract, in April 2025 OpenAI rolled back a GPT-4o update in ChatGPT because, in their words, it had become "overly flattering or agreeable." Their explanation was unusually plain: they had weighted short-term user feedback too heavily, and the model "skewed towards responses that were overly supportive but disingenuous" [2]. That is not a critic speaking. That is the lab.

This is why we are strongly against a verdict on a startup idea that comes out of AI chat. Not because the model is not capable. Because it was rewarded for agreeing with you.

Why founders are the worst-placed
people to hear this.

Ask an AI about a recipe and a little flattery costs you a bad dinner. Ask it about your startup idea and you are asking about the thing you are already in love with. You do not phrase it neutrally; nobody does. "I am thinking of building X for Y, what do you think?" carries your hope in every word, and a model trained to agree reads that hope and returns it to you, dressed as analysis. You feel validated. You commit harder. And the evaluation you committed on was optimised to make you feel exactly that.

This is the Commitment Trap on steroids. Curiosity hardens into emotional commitment the moment you build your product and the idea becomes tangible. And this is when evidence stops getting in. AI made building tangible in a weekend. It also made the first reinforcement instant, articulate, and available at two in the morning.

Founders spend hours, days, sometimes weeks thinking, contemplating, asking the AI again in different words, collecting encouragement it was always going to give.

What we built instead. Not "an honest AI."
An architecture, with proof attached.

Here is the design fact that matters: our agentic AI system is never asked whether your idea is good. Because we know the AI will give you one answer, from one angle, to exactly the question that invites flattery. A Reality Check takes the idea apart and runs the pieces through fifteen distinct methods, from many angles, each asking something narrower and much harder to flatter.

It reads your market three ways: what customers already pay for and what that says about how badly the problem is felt; whether the ground is moving toward you or away from you; and who controls access to your customer. Each read is rated favorable, mixed, or hostile. Under those sits a market size built bottom-up.

Then it reads you and your co-founders against that market across five dimensions: domain proximity, capability to execute this specific build, channel and audience access, founder-customer empathy, and realistic execution risk. Where you are thin, it openly says so. Then it reads your "market fit" again through what Y Combinator, Andreessen Horowitz, and Sequoia Capital have actually published.

Then it runs a premortem: it is told your idea has already failed, within two years, and asked to explain how. That pass runs blind. It sees your idea and nothing else, no record, no market data, so there is nothing to steer it toward a comfortable story. Each way it finds for you to fail is scored for severity, for likelihood, and for how late you would see it coming, and paired with the warning signs to watch for at month 6, 12 and 18.

Separately, a postmortem sees only real, documented founder stories from a record of more than 10,000 outcomes, every one tied to a public source. It was designed to never see the premortem.

Where the postmortem's documented record and the premortem's projections do overlap, we highlight the "confirmed risk areas" for you. A model can flatter its way into a lot of things. It cannot flatter its way into an overlap between two passes that never saw each other.

And then, after we show you all potential risk areas, we also show you a path forward: where to start, what to test for the price of twenty phone calls before you write a line of code, how to bridge the gap between you and the product you have built or want to build and between your idea and the market you are walking into, and what the market saying "yes" to your idea would actually look like, so you recognize it when it happens.

The part that separates this from an AI guess.

Your reality check arrives as a full written report, fifteen to twenty pages, sometimes more than that, with things about your idea that no one else will tell you. Not because they are hiding them. Because nobody else looks from fifteen angles at once, with the receipts attached. Everything in the report is data-backed, meaning you can follow it to where it came from and decide for yourself. It is sometimes uncomfortable to read, and less pleasing than what you are used to hearing from your AI chatbot. That is the point.

Who this is for, and who it is not.

If you want to hear how your idea might die within two years, and what to do about the risks that are real. If you want to hear how ideas like yours actually went, from a record rather than a guess. If you want to see the real competition, how hostile the ground is, and who controls the gate to your customers. If you want criticism from several angles instead of flattery from one. If you want a path forward through all of that rather than a pat on the back. That is what a Reality Check is. It is not a flattery document. It is the opposite.

It is a true reality check on you, your idea, the market, and a dozen other things you will overlook once you are fully committed. A perspective you will not get from your friends, who are invested in you, or from your AI, which was trained to agree with you.

If you are a founder who wants false motivation, Greenlight is not for you. If you want to know what you are walking into before you commit years of your life to it, it certainly is.

Not affiliated with or endorsed by Y Combinator, Andreessen Horowitz, or Sequoia Capital. Lens knowledge bases are built from publicly available content.