Staying sharp with AI

In a 2023 study, 27 radiologists read breast scans with AI suggestions. Some suggestions were wrong on purpose. When the AI was wrong, even the most experienced group fell from 82% right to 45.5%.

A steady tool makes it worse. In a 1993 flight-simulator study, some people watched automation that was always equally reliable. They caught far fewer of its failures than people whose automation kept changing. AI also writes smooth, sure text. But smooth is not the same as true (When AI makes things up).

In a 2021 study, asking people to decide before they saw the AI's answer cut over-reliance. A confidence score helps only if it is honest: "90% sure" should be right about 90% of the time. A chatbot's sure tone is not a score at all. In a 2020 study, scores alone did not make the human-plus-AI team more accurate. And an AI can't flag a problem it was not built to look for.

In a trial Anthropic published in January 2026, 52 mostly junior software engineers learned a new Python library (a ready-made package of code). Some used an AI assistant and some did not. On a quiz right after, the AI group scored 50% on average. The group that coded by hand scored 67%. The biggest gap was in finding and fixing bugs. The AI group was not clearly faster, either. Those who learned most asked follow-up questions and for explanations.

Many people talk to chatbots for company or advice. Many find it helpful. In a 4-week 2025 study by OpenAI and the MIT Media Lab, very heavy use went with more emotional dependence. People who bonded more with the chatbot were more likely to be lonely. But lonely people may also seek out chatbots.

A chatbot tends to agree with you. This is sycophancy. It feels kind, but it can make unhealthy thinking stronger. Warning signs: you prefer the bot to people, hide how much you use it, or feel worse without it.

Children and teens need extra care. See AI with children and families.

AI runs in data centers that use electricity, and water for cooling. Training a model is a huge one-time job. Using the model to answer is inference. Each answer uses a little energy, but it happens billions of times.

Estimates for one text question vary a lot, because they count different things. Some count only the AI chips. Some add the whole data center. Some also count making the hardware. So treat each number as one estimate under one method.

What adds up is heavy work at scale, like video after video or agents that run for hours. Smaller models often use much less energy. So pick the smallest model that does the job well (Choosing a model). Don't regenerate 20 times.