Why AI gives average answers, and what makes it smarter

1The loudest voice

First, a confession on behalf of humans: we do it too. Kabir tries the trick on five teammates under the shelter. "Pick a number between 1 and 10." Three of them say seven.

That's no fluke. In a famous 1976 study at Yale, 558 people were asked for the first digit that came to mind, and about 28% said 7. That's nearly three times what pure chance would give. In a 2025 YouGov poll, a third of Britons picked 7 out of 1 to 10. Psychologists even have a name for a version of this habit, the "blue seven": asked for any colour and any digit, lots of people say blue and seven.

So why 7? The Yale researchers had an explanation: we're trying to look random. 1 and 10 sit at the edges, so they feel too obvious. 5 is the middle. Even numbers feel too neat. 3 and 9 are in the 3 times table. Cross them all out, and 7 is the one left standing.

Now think about where the AI learned to talk. From us: billions of pages of human writing, full of games, quizzes and posts where people "randomly" picked 7. Remember the giant word-guesser from how LLMs work? It scores every possible next word, then rolls loaded dice to pick one.

Picture a stadium of 50,000 fans, all shouting their answer at once. You can't hear each voice. You hear the roar, and the roar is the answer most of them shout. That's what the word-guesser hears when it scores the next word: the roar of the whole internet. For "a random number between 1 and 10", the roar says 7.

But what about the loaded dice? They let a quieter voice win now and then, so sometimes you get a 3 or a 4. But the roar is so loud that it barely matters. In a 2025 study, three popular AI models picked 7 about 80% of the time. One of them, with its dice set to the wildest setting tested, said 7 in every single English answer. So the AI didn't just learn our habit. It turned it up.

So now the puzzle cracks open. The AI isn't broken. It's doing its job perfectly. Its job is to give the most likely answer, and the most likely answer is, almost by definition, the most common one. Reading the whole internet didn't make it unusual. It made it the world's leading expert on what's usual.

Wait, isn't the average of 1 to 10 actually 5.5? Yes! "Average" here means typical. Remember mean, median and mode from maths class? The AI gives the mode: the answer that turns up most often.

2Why every AI sounds alike

But it isn't only numbers. Remember the second round of training, where people grade the AI's answers like a teacher marking homework? Here's the twist: graders tend to prefer answers that sound familiar. So every round of marking nudges the AI further toward the safe, typical answer. Researchers have measured it: this kind of training makes answers noticeably less varied.

So different AIs end up sounding alike. In a 2025 study that won a best-paper award at NeurIPS, one of the biggest AI conferences, researchers asked 25 different AI models to "write a metaphor about time", 50 times each. Most of the answers crowded around just two ideas: time is a river, or time is a weaver.

You've probably met this sameness already: the essay that "delves into" every topic. A 2025 study of over 15 million research summaries in medicine and biology found sudden jumps in words like "delves", "showcasing" and "underscores". By that count, at least 13.5% of the 2024 summaries had been worked on with AI.

And it rubs off on us. In another experiment, 293 people wrote short stories, some with ideas from AI. The AI-helped stories were rated more creative and better written. But they were also more similar to each other. Each story got better. Together, they got more alike.

3Confident guesses

Here's a harder question. What happens when nobody knows the answer at all? That evening, Riya asks: "Who was player of the match in our school's inter-house final in 2019?" It answers at once: Aarav Mehta, 64 runs. There is no Aarav Mehta at her school.

So where did he come from? Hardly anyone ever wrote that match down anywhere. For this question, the stadium is nearly empty: just a murmur of possible names.

But the word-guesser can't stay silent. Its job is to pick the next word, so it picks the most plausible-sounding one: a believable name, in the same confident tone it uses for "Paris". That's what we call a hallucination. It isn't lying. It's a fluent guess sitting where knowledge should be.

So why doesn't it just say "I don't know"? Think of a multiple-choice exam with no negative marking. Leave a question blank and you get zero; guess, and you might get lucky. In 2025, researchers at OpenAI argued that AI is trained and tested in a very similar way, so guessing wins. When they asked a popular chatbot for the birthday of one of their own authors, it gave three different dates. All wrong.

Newer models are getting better at saying "I'm not sure", partly because researchers now reward it. They still guess more often than you'd like. So treat AI like a brilliant friend with a patchy memory. For facts that matter (names, dates, numbers, medicines), ask where it got them and check. Apps that can search the web, by writing an order slip for a search tool, can look facts up instead of guessing.

4What makes AI smarter

Riya's next thought is the obvious one. Next year's AI will be bigger. Won't that fix it? To see what "bigger" means, think about choosing a new laptop. You compare the specs: more storage, more memory, a faster processor, more apps. More of each means a more powerful machine. An AI has parts that work in a surprisingly similar way.

• The knobs are its storage. Everything it learned lives in those billions of knobs. But they hold notes, not photocopies. AI researcher Andrej Karpathy roughly estimated that one model squeezed about 10 terabytes of text into a 140-gigabyte file of knobs. Notes are also why it can misremember.
• Data is what it learns from: books, websites, code and conversations.
• Compute is the processor's work. There's the huge effort to train it, and a smaller effort every time it answers.
• The context is its RAM: what it can look at right now.
• Tools are its apps: the calculator, the search, the PDF maker.

How much compute are we talking about? Imagine doing a billion sums every second, without a break. The maths behind training the biggest models would still take you well over 100 million years. And the compute used to train the top models has been growing four to five times every year.

But does more of everything really help? In 2020, researchers at OpenAI found a surprisingly neat pattern. As they grew the knobs, the data and the compute, the model's mistakes at guessing the next word fell smoothly and predictably, over a range of more than ten-million-fold. Not magic, more like a recipe. These patterns are called scaling laws.

But there was a catch. In 2022, DeepMind trained over 400 models and found that most big models were like a student with a huge brain who barely studied: lots of knobs, not enough reading. So they built Chinchilla, with a quarter of the knobs of their earlier model, Gopher, but four times the reading, for the same amount of compute. The smaller model that read more won.

That's the accurate version of the laptop idea: more compute lets you build a bigger model and feed it more data, and you need both. Meta trained Llama 3 on over 15 trillion tokens, seven times Llama 2's data, and its models were still getting better when training stopped.

So do bigger models give less average answers? Not on their own. A bigger, better-fed model knows more of the quiet voices: the cricket coach, the poet, the heart surgeon. But ask it a plain question and it still answers with the roar. Hold on to that. It's the key to the end of our story.

5Better books, not just more

Feeding it more data sounds easy. But what if the data is junk? Programmers have an old saying: garbage in, garbage out. In 2023, Microsoft researchers trained a small model, phi-1, mostly on "textbook quality" text: clear, well-explained lessons and code. It trained on just 8 GPUs (the chips that do AI's maths) for 4 days, and it scored 50.6% on a standard coding test, beating coding models more than ten times its size that had read far more.

But here's a bigger worry: good human writing is running out. Researchers at Epoch AI put the stock of good-quality public human text at about 300 trillion tokens. In 2024 they projected that, if trends continue, AI training would use it all up sometime between 2026 and 2032. That window has already opened.

So why not train AI on AI's own writing? Think of a photocopy of a photocopy. Each copy loses the faintest details first. A 2024 study in Nature found the same thing with AI: train each new model on the last one's output, and the rare, unusual stuff disappears first. They called it model collapse. AI learning from AI is the average of the average.

That's why fresh, human, slightly weird writing is becoming precious. The quiet voices are what keep AI from turning into one big echo.

6Your un-average answer

So, 7 and Aarav Mehta. Ask for a random number, and fifty thousand voices shout 7. Ask about a match nobody wrote about, and the stadium is empty, so the model makes up a name rather than leave a blank. They're the same thing: the model gives you the most likely answer, whether or not anyone actually knows it.

And bigger won't change that on its own. More and better data, more knobs and more compute: that's how AI keeps getting smarter. But even the smartest AI, asked a plain question, gives you the roar. Why? Because it doesn't know you.

Watch what happens when Riya brings her own context. She asks for a chant for her team's big final. The first answer could be for any team on Earth. So she adds what only she knows: the team's name, the captain's nickname, a chant she loved last year, and a rule: Hinglish, under 20 words.

Same model, different context, a very different answer. Her notes changed which crowd the AI was listening to. The quiet voices were in there all along. Riya's context turned them up.

Here's how to get un-average answers:

• Give it your context: who it's for, what it's for, and what you already know.
• Show an example of what good looks like to you.
• Set constraints: word limits, tone, and words to avoid ("no 'delve', please").
• Ask for ten options, not one, then pick the surprising one. One study found that asking for five answers at once, with how likely each one is, made AI's poems, stories and jokes up to about twice as varied.
• Bring your taste. Choose, cut, mix and rewrite. The final call is yours.

And if you ever need a truly random number, ask the AI to use a tool, like running a tiny bit of code. A real random-number generator doesn't have a favourite.

So what does this mean for you? The average answer is now instant, free and everywhere. Everyone has it. That makes it the starting line, not the finish. What makes your work stand out is everything the AI can't find in the roar: your context, your examples, your taste.

As for Kabir's trick, it only works on plain questions. The AI knows what everyone would say. Only you know what you would say.

That night, Kabir sends Riya a reel: "10 secret prompts AI companies don't want you to know. Just ₹1,999. Offer ends tonight." Magic prompts that make the AI listen to the right voices. Do they work? One of them really did, once. It made an AI four times better at maths.