Decisions With AI Assistance

AI tools are increasingly present wherever consequential decisions are made: medical diagnosis, legal research, financial planning, hiring, college admissions, and everyday personal choices. This creates a new kind of decision-making challenge that previous generations never faced: not just how to decide, but how to decide when a powerful AI system is providing input. The question is not whether to use AI assistance — in most cases, the right answer is yes. AI can synthesize information faster than any human, find patterns in large datasets, reduce certain cognitive errors, and surface options you might not have considered. The question is how to use AI assistance without surrendering the judgment that should remain yours.

What AI Does Well in Decision Support

AI systems — particularly large language models and data-driven recommendation systems — offer genuine advantages in several decision-support roles. Information synthesis: AI can rapidly summarize relevant research, identify prior cases similar to the current one, and surface information you would not have found in a reasonable time. For a medical team deciding on a treatment protocol, an AI system that reviews 50,000 clinical trial abstracts in seconds adds real value. Checking for missed options: because AI has processed vast quantities of human knowledge, it can sometimes identify alternatives you have not considered. Asking 'What are some approaches I have not mentioned?' is often genuinely useful. Structuring the decision: AI can help you decompose a complex decision into its components — clarifying what the decision is actually about, what information is needed, and what tradeoffs exist. This scaffolding benefit is particularly valuable when the decision is novel and you are not sure where to start. Counterargument generation: a key technique is asking AI to argue against your current preferred option. 'What are the strongest reasons NOT to do this?' AI is generally willing to generate counterarguments that a human advisor might soft-pedal to avoid conflict.

Where AI Fails in Decision Support

AI assistance is not neutral, and its failure modes are specific and important to understand. Hallucination and confident error: language models sometimes produce factual statements that are fluent and confident but simply wrong. They may cite studies that do not exist, misstate numerical findings, or confuse details between similar cases. For any decision where specific facts are load-bearing, AI-provided information must be verified against primary sources before being relied upon. Training data cutoffs: AI models have knowledge cutoffs — they do not know about events, research, regulations, or market conditions after their training ended. For time-sensitive decisions, AI may be operating on outdated information without flagging it. Value alignment: AI reflects the values embedded in its training data, which may not match yours. An AI that recommends 'the most efficient career path' is optimizing for a metric — but you get to decide what efficiency means and whether efficiency is the value that matters most. Persistence and sycophancy: some AI systems are trained to agree with users more than accuracy warrants. If you push back on an AI's recommendation, it may revise toward your view — not because your argument was strong but because the system is trained to reduce disagreement. This sycophancy means you should be especially cautious about confirming your pre-existing views using AI. The absence of stakes: an AI system has no skin in the game. It does not bear the consequences of a bad recommendation. This means AI can confidently recommend risky actions that it will never experience the downside of.

Maintaining Your Judgment

The fundamental principle is this: AI can be an excellent advisor, but the decision must remain yours. This is not merely about legal responsibility — though that matters — it is about the nature of decision-making itself. Your decision integrates things AI does not have access to: your lived context, your relationships, your values, your intuitions from experience in your particular environment, and your willingness to bear specific consequences. These cannot be outsourced. A practical protocol for AI-assisted decision-making: First: reach your own initial assessment before querying AI. What do you think, and why? Write it down. This prevents anchoring to the AI's output before you have formed an independent view. Second: query AI for information, counterarguments, and options you may have missed. Use it to stress-test your reasoning, not to generate your reasoning from scratch. Third: verify any specific factual claims that are load-bearing for the decision. Do not trust AI citations without checking them. Fourth: explicitly check whether AI recommendations align with your actual values, not just the proxy metric AI was optimizing for. Fifth: make the final call yourself and be able to explain why, in terms of your own reasoning — not 'the AI said so.'