The primary issue to be addressed in developing such interactive products is the data problem. A typical AI Smash or Pass training set often needs to integrate 500,000 to 2 million real-person facial image data, which mainly come from publicly crawled images on the Internet (accounting for 60%-80% of the data sources) and paid authorized image libraries (accounting for about 15%-25%). The human resource cost in the data processing stage accounts for as much as 30% of the total budget. The professional annotation team performs feature labeling at a price of approximately $0.05 per image, with a daily processing limit of about 20,000 images. As a 2023 study by the MIT Media Lab pointed out, 78% of AI rating tools for popular entertainment use public datasets that have not been adequately reviewed, with a bias rate as high as 22% of the sample variance. In an environment where data governance is lacking, the evaluation logic of “AI Smash or Pass” is highly likely to magnify the original social biases.
The hardware consumption when building the scoring model is quite significant. Take the NVIDIA A100 graphics card as an example. Training a ResNet architecture model with 30 million parameters requires continuous operation for about 120 hours, and the energy consumption cost exceeds 500 US dollars. This is only for the development of a single model version. After commercial deployment, a service cluster that processes an average of 1 million user requests per day needs to be equipped with 20 to 40 server instances, and the latency during peak hours should be controlled within 700 milliseconds. Test cases of Meta’s DeepFace system, which was open-sourced in 2022, show that in cross-race recognition scenarios, the best model only achieved an accuracy rate of 86%, and the error rate caused by skin color differences was as high as 15 times the industry average. This means that the evaluations users receive are essentially mixed with technical limitations.
The ethical risks of ai smash or pass have triggered regulatory responses in many countries. The EU’s Artificial Intelligence Act classifies it as a high-risk application, requiring developers to undertake the obligation of verifying the legality of facial data (compliance costs account for approximately 12% to 18% of total revenue). In the 2024 fine case against Clearview AI by the Office of the Privacy Commissioner of Canada, it was confirmed that unauthorized facial recognition involves violations of user privacy rights. A single violation can be fined up to 4% of the annual income. Tests by the independent research institution AI Now Institute show that when the input image has real situations such as a 30% reduction in brightness and a 15-degree shift in shooting Angle, the fluctuation range of the algorithm’s consistency score exceeds 40% of the original score. Such score fluctuations clearly have no objective reference value.
The essence of a business model is to form a business closed loop by leveraging user-generated content (UGC). On average, each user session contributes 6 to 8 valid interaction data, and the system continuously optimizes the model feature weights based on this. A typical platform makes profits through membership subscriptions (with a monthly fee of $4.99) and AD clicks (with a CPM fee of approximately $15). The lifetime value (LTV) of paying users is about $38, and the cost of customer acquisition (CAC) needs to be controlled within $10 to maintain a healthy profit margin of over 20%. In its 2024 industry report, the UK Competition and Markets Authority (CMA) pointed out that 67% of entertainment rating apps have an issue of opaque data retention periods. Even after users delete the original uploads, approximately 45% of the systems still retain the derived feature vectors in the background for more than 180 days. This has led to many seemingly harmless “ai smash or pass” products in the market hiding regulatory costs and ethical controversies far beyond expectations behind them.