AI in its present form has limitations. By definition, AI powered computer programmes itself and learns as it goes along. In the process, it creates a database of information, which it uses to generate additional computer programming code as it keeps on learning more without any human assistance.

All these programmes require huge carefully categorised data to be smart. If the data is carelessly characterised, the machine arrives at the wrong conclusions. An object can be misinterpreted by an image recognition programme, say a slight change in image makes the computer perceive a vehicle as a rhino. AI thus is ineffective if there is no carefully categorised data.

To make driverless cars successful, video footage of actual surroundings on the roads is gathered and is labelled. To deal with certain situations, unfortunately data is not yet available, say crash data. Data from other sources is collected when it is absent. Such dummy data is useful in some areas, but is hazardous in some situations.

When there are accidents, there is no adequate data about them, and hence predictive models built here are neural. There are the safety concerns, and these halt the progress of the autonomous or driverless cars. It is difficult for such cars to navigate a K-turn.

Computers cannot be always better than us in making decisions. They are good at maths and calculations. Still regression models the computers use to make predictions or recognise pattern are very old. A group of birds on the roads makes autonomous cars brake or swerve, whereas the normal car drivers would continue driving slowly as they know birds will fly off.

Autonomous cars work ideally in controlled situations.