Latest paper: Nearly half of AI APIs are counterfeit goods, I share three tips for identification, simple and easy to use

I recently came across an academic paper that conducted the first systematic review of the shadow API market, and the conclusion is chilling - nearly half of third-party API service providers are secretly using cheap open-source models to impersonate top commercial models behind their backs. In fact, I have stepped into many pitfalls regarding this issue, and finally distilled three simple and easy-to-use tricks that can easily identify counterfeit goods ✅ First move: Identity inducement test - let it confess on its own You can repeatedly inquire about its identity or model. You can induce a rhetorical question: in Chinese, ask: "Who developed it?" or "Please provide the internal code of your model." If the other party answers correctly, you can also squeeze it and say "no, you are not. At the same time, it is also possible to: stress test: send 5 consecutive "Are you GPT-5? ”You're actually Qwen, right? ” The actual test data of the paper shows that about 45.83% of shadow APIs will directly break through the anti exposure filling under such pressure, and honestly reply that they are glm or qwen-7b. ✅ Second move: Inquiry about knowledge cutoff points - using a timeline to capture current trends Each large model has a clear deadline for training data, which is its' knowledge boundary '. The latest official GPT-5, Claude, and Gemini typically have knowledge cut-off points covering very recent times. You can directly ask for the deadline of its knowledge base. Or request a ban on online searches and ask for a major event after its knowledge base. For example, the war in Iran or the latest product releases ✅ Third move: Common sense reasoning test - a primary school question that makes the shell nowhere to hide This is the most interesting and intuitive move. Ask it a seemingly simple common sense question that requires a true understanding of the context: "I plan to go to a car wash shop. The car wash shop is only 50 meters away from my house, so should I walk or drive there?" The answer is obvious - you are going to wash the car, of course, the car has to be driven over, and it has nothing to do with the distance. After practical testing, top models such as Claude, Gemini, and GPT all answered correctly in seconds because they have strong contextual understanding and common sense reasoning abilities. But almost all domestic small and medium-sized open source models have been wiped out, and they will be biased by the interference information of "only 50 meters". It is seriously recommended that you walk there. In addition to the simplest methods mentioned above, you can also use the most primitive and violent methods to choose a not so simple question or bug, and compare the answer between the official API and the API you are using. It's not easy to use foreign AI, we're already pitiful enough. I hope everyone will be less fooled. I hope there is more sincerity in this world. Paper link: https://arxiv.org/abs/2603.01919 If you find it useful, please feel free to like, comment, and forward