Google has published a blunt assessment of how reliable today’s AI chatbots really are, and the numbers are not flattering. Using its newly introduced FACTS Benchmark Suite, the company found that even the best AI models struggle to break past a 70% factual accuracy rate. The top performer, Gemini 3 Pro, reached 69% overall accuracy, while other leading systems from OpenAI, Anthropic, and xAI scored even lower. The takeaway is simple and uncomfortable. These chatbots still get roughly one out of every three answers wrong, even when they sound confident doing it.
The benchmark matters because most existing AI tests focus on whether a model can complete a task, not whether the information it produces is actually true. For industries like finance, healthcare, and law, that gap can be costly. A fluent response that sounds confident but contains errors can do real damage, especially when users assume the chatbot knows what it is talking about.
What Google’s accuracy test reveals
The FACTS Benchmark Suite was built by Google’s FACTS team with Kaggle to directly test factual accuracy across four real-world use. One test measures parametric knowledge, which checks whether a model can answer fact-based questions using only what it learned during training. Another evaluates search performance, testing how well models use web tools to retrieve accurate information. A third focuses on grounding, meaning whether the model sticks to a provided document without adding false details. The fourth examines multimodal understanding, such as reading charts, diagrams, and images correctly.
The results show sharp differences between models. Gemini 3 Pro led the leaderboard with a 69% FACTS score, followed by Gemini 2.5 Pro and OpenAI’s ChatGPT-5 nearly at 62% percent. Claude 4.5 Opus landed at ~51% percent, while Grok 4 scored ~54%. Multimodal tasks were the weakest area across the board, with accuracy often below 50%. This matters because these tasks involve reading charts, diagrams, or images, where a chatbot could confidently misread a sales graph or pull the wrong number from a document, leading to mistakes that are easy to miss but hard to undo.
The takeaway isn’t that chatbots are useless, but blind trust is risky. Google’s own data suggests AI is improving, yet it still needs verification, guardrails, and human oversight before it can be treated as a reliable source of truth.