Can AI Really Think? What Today's Artificial Intelligence Can and Cannot Do

Do that billions of times and something remarkable happens. To predict well, the model has to absorb grammar, facts, writing styles and the typical structure of arguments and explanations. Since 2017, most of these models have used an architecture called the transformer, introduced in a research paper titled Attention Is All You Need, which lets the model weigh how every word in a passage relates to every other word.

After this initial training, developers fine-tune the model, often using feedback from human reviewers, so that it answers questions helpfully and follows instructions rather than simply continuing text.

Why it can seem to reason

Ask such a model a maths word problem and it may lay out the steps and reach the right answer. That looks like reasoning, and in a practical sense it partly is: the model has learned patterns of step-by-step problem solving from the text it was trained on, and newer systems are trained to "think" through intermediate steps before answering, which measurably improves accuracy on many tasks.

But the reasoning is built on statistical patterns, not on understanding the world the way people do. The model has never touched a cricket ball, paid a bill or felt hungry. It knows how humans write about those things.

Why it still makes things up

Because the model's goal is to produce plausible text, it sometimes produces plausible text that is false. Researchers call these errors hallucinations. A chatbot may invent a book that does not exist, attribute a quote to the wrong person or cite a court judgment that was never delivered, all in a confident tone.

Hallucinations are more likely when:

• the question is about something rare, recent or local, such as a small-town event;
• the answer requires an exact figure or citation;
• the question contains a false assumption that the model goes along with.

Some products reduce the problem by letting the model search the web or a company's documents and quote them, but none eliminate it.

Intelligence, prediction and consciousness

These words get mixed up in public debate.

• Prediction is what the model does mechanically.
• Intelligence, in the sense of solving varied problems well, is something these systems show to a real and growing degree, though unevenly.
• Consciousness, having experiences and feelings, is a different matter entirely. There is no scientific evidence that today's AI systems are conscious, and their fluent talk about feelings is learned from human writing.

A chatbot saying "I understand how you feel" is producing the kind of sentence people write in that situation. It is not reporting an inner state.

What this means for you

AI can be an excellent assistant for drafting, summarising, translating, brainstorming and explaining. Treat it like a capable intern who never admits uncertainty:

1. Check facts that matter, especially numbers, names, laws and medical information.
2. Ask for sources and open them yourself.
3. Don't share sensitive personal data such as Aadhaar or bank details with chatbots.
4. Use your judgement for decisions about health, money and law.

The more clearly we see AI as a powerful pattern-learning tool rather than a digital mind, the better we can use it. Next, see how these tools are already being used across the country in how AI is changing everyday life in India.