1. Check the Source and Its Motives
Before you get drawn in by a new AI's amazing capabilities, ask a simple question: who made this and why? The AI race is dominated by a few major players: OpenAI (creators of GPT models), Google (behind Gemini), and Anthropic (which makes Claude). Each
has different priorities. OpenAI often pushes the boundaries of capability first, Google leverages its vast ecosystem of products like Search and Android, and Anthropic has built its brand on a 'safety-first' approach. Understanding the company’s business model—whether it’s selling API access, integrating AI into enterprise software, or driving ad revenue—gives you context for its strengths and potential blind spots.
2. Test Its Reasoning, Not Just Its Memory
The earliest chatbots were good at recalling facts they were trained on. The next generation of models, like the much-anticipated GPT-5, are being judged on their ability to reason through multi-step problems. Don't just ask it for the capital of France. Give it a complex, real-world scenario from your own work or life. For example, ask it to plan a multi-city trip on a tight budget with specific constraints, or to analyze a dense document and identify the three most critical risks. A truly advanced model won’t just retrieve information; it will synthesise it, apply logic, and help you solve a problem that doesn’t have a simple answer.
3. Probe for Bias and Inconsistency
All AI models are trained on vast datasets of human-generated text and images, and they inevitably inherit the biases contained within that data. A critical part of evaluating any new model is to test its fairness and consistency. Ask it the same question phrased in different ways. Prompt it on sensitive or nuanced social topics. Does it provide balanced perspectives, or does it lean towards a specific viewpoint? Does it refuse to answer reasonably safe questions? A model's inconsistency or hidden biases can be a major risk, especially if you're relying on it for important tasks at work or for decision-making.
4. Verify, Then Trust (A Little)
One of the biggest flaws of early AI models was their tendency to "hallucinate"—to confidently state incorrect information as fact. While newer models are getting much better at accuracy and citing sources, you should never blindly trust an AI's output. A good habit is to treat every answer as a first draft from a very fast, sometimes unreliable assistant. When it provides a fact, ask for the source. If it gives you a summary of an event, cross-reference it with a quick search. The goal isn't to catch the AI in a lie, but to build a workflow where human verification is a natural and necessary step. The best AI users are sceptics.
5. Assess Its Practical, Real-World Utility
Beyond the impressive benchmarks and technical demos, the most important question is: does this tool actually help you? A model might be able to write a sonnet and code an application, but if your main need is summarizing long email chains, its performance on that specific task is all that matters. Try to integrate the new tool into one of your daily or weekly tasks. Did it save you time? Did it reduce stress? Did it produce a better result than you could have achieved on your own? The best AI model for you isn't necessarily the one with the highest benchmark score, but the one that provides tangible value for your specific needs.
