Can AI Discover Something Genuinely New Without Human Examples?

Yes, AI can discover something genuinely new without direct human examples, but with important caveats about what “new� means. In fields like mathematics, materials science, and biology, AI systems have produced results that were not explicitly taught or shown by humans, though they still rely on human-defined goals, data, and evaluation criteria. This article explains how such discoveries happen, what limits remain, and why the answer depends on how you define “genuinely new.�

Formula

The calculator above gives a rough “novelty score� based on two factors: the number of training examples (t) and the number of model parameters (d). The formula is:

Novelty = (d / (t × 100)) + (t / 1000)

Where:

• t = number of training examples (unitless count)
• d = number of model parameters (unitless count)

This is not a real scientific metric—it’s a heuristic to illustrate that novelty tends to increase with model complexity relative to data, but also with more data (up to a point). The first term captures how much “room� the model has to find patterns beyond the data, and the second term captures that more data can help. The score is unitless and meant for conceptual comparison, not rigorous measurement.

Worked Example

Using the default values from the calculator: t = 1000 examples, d = 1,000,000 parameters.

1. First term: d / (t × 100) = 1,000,000 / (1000 × 100) = 1,000,000 / 100,000 = 10
2. Second term: t / 1000 = 1000 / 1000 = 1
3. Novelty = 10 + 1 = 11

So the novelty score is 11.00. This suggests that with a large model and relatively few examples, the potential for novel output is high—but remember, this is a simplified illustration.

What “Genuinely New� Means in Practice

When people ask if AI can discover something new, they usually mean one of three things:

• New to the AI – the AI finds a pattern it wasn’t explicitly trained on.
• New to humans – the result is not known to the researchers or the field.
• New in the world – the result is a first-ever discovery, never before recorded.

AI has achieved the second and third in several domains, but always within a framework set by humans. For example, in 2022, DeepMind’s AlphaTensor discovered a faster matrix multiplication algorithm—a mathematical procedure that was not in its training data. The AI was given the goal “find a faster way to multiply matrices� and a set of rules, but the exact algorithm it found was new to humans. Similarly, in materials science, AI has predicted new crystal structures that were later synthesized in labs. These are genuine discoveries, but they rely on human-defined objectives and evaluation functions.

How AI Discovers Without Examples

There are several techniques that allow AI to go beyond its training data:

• Reinforcement learning: The AI explores a space of possibilities and receives rewards for achieving a goal. AlphaGo’s move 37 in game 2 against Lee Sedol is a famous example—it was a move that no human had played in thousands of years of Go history.
• Generative models: Models like GANs or diffusion models can create new images, music, or molecules that are similar to training data but not identical. The novelty comes from interpolation and recombination.
• Search with evaluation: The AI searches a large space (e.g., all possible mathematical proofs or chemical compounds) and uses a learned or programmed evaluator to score candidates. This is how AlphaTensor worked—it searched for matrix multiplication algorithms and evaluated them for speed.
• Active learning: The AI chooses which data to collect next, potentially discovering new regions of the input space that humans hadn’t considered.

In all cases, the AI is not “thinking� in a human sense; it is optimizing a mathematical objective. The novelty emerges from the search process, not from spontaneous insight.

Common Misconceptions

One common misconception is that AI can generate truly novel ideas without any human input at all. In reality, every AI system is trained on human data, whether that’s text, images, or scientific measurements. Even AlphaTensor’s discovery was based on the rules of matrix multiplication, which are human-defined. The AI didn’t invent a new branch of mathematics; it found a clever trick within an existing framework.

Another misconception is that “new� means “unpredictable.� But AI discoveries are often predictable in hindsight—once you see the result, it makes sense. The novelty is in the search, not in the outcome being random.

Finally, some people think that AI will soon discover things without any human supervision. But every AI system has a goal that a human sets. Even “open-ended� systems like OpenAI’s GPT-4 are designed to follow instructions and produce text that humans find useful. The AI doesn’t decide what is interesting or important; that judgment is built into the training objective.

Examples of AI Discoveries

Here are a few documented cases where AI produced results that were new to humans:

• AlphaTensor (2022): Found a way to multiply 4×4 matrices using 47 multiplications instead of the previous best 49. This was a new mathematical result, verified by humans.
• AlphaFold (2020): Predicted protein structures for hundreds of thousands of proteins, many of which had not been experimentally determined. The predictions were accurate enough to be used as a substitute for lab work in many cases.
• AI-designed drugs: In 2019, an AI system (called “Bossâ€� by Insilico Medicine) designed a novel drug candidate for fibrosis in about 21 days. The molecule was new, though it was later synthesized and tested.
• Game strategies: AlphaGo’s move 37 and later AlphaZero’s strategies in chess and shogi were not in any human database. They were discovered through self-play, where the AI played against itself millions of times.

These examples show that AI can indeed discover things that humans hadn’t found, but the discoveries are always in domains where the rules are clear and the objective is measurable.

Limits and Challenges

Despite these successes, there are significant limits to AI discovery:

• Goal specification: Humans must define what “goodâ€� means. For scientific discovery, this is often a proxy like accuracy or speed, which may not capture the full complexity of the problem.
• Data availability: AI needs data to learn from. In fields with sparse data (e.g., rare diseases), discovery is harder.
• Interpretability: AI often finds solutions that work but are hard to understand. A new matrix multiplication algorithm is a sequence of steps, but explaining why it works is not trivial.
• Evaluation: How do you know if a discovery is “genuinely newâ€�? It requires comparing against all known results, which is often impossible in large search spaces.
• Bias and noise: AI can discover false patterns that are artifacts of the training data, not real phenomena.

These challenges mean that AI discovery is not automatic; it requires careful human oversight and validation.

Frequently Asked Questions

Can AI discover something without any data?

No, AI needs some form of data or environment to interact with. Even reinforcement learning requires a simulation or real-world environment with defined rules. Without data, the AI has nothing to learn from or optimize against.

Is AI discovery the same as human creativity?

No, AI discovery is based on optimization and search, not on subjective experience or emotion. Human creativity often involves intuition, emotion, and cultural context, which AI does not have. However, the output can be similar in novelty and usefulness.

Can AI discover a new law of physics?

Potentially, yes, if you define “law� as a compact mathematical relationship. AI has been used to rediscover known laws (like Kepler’s third law) from data, and it could find new ones if given the right data and objective. But the AI wouldn’t “understand� the law; it would just find the equation.

Will AI replace human scientists?

Not entirely. AI can accelerate discovery by searching large spaces and finding patterns, but humans are still needed to set goals, interpret results, and decide what is important. The role of the scientist may shift from doing routine analysis to directing AI and validating its findings.

How do we know an AI discovery is correct?

AI discoveries must be validated by other methods—experiments, proofs, or independent calculations. For example, AlphaFold’s predictions were compared to experimental structures, and AlphaTensor’s algorithm was verified by running it. Human experts are always involved in the final check.

What is the difference between “new� and “novel� in AI?

“New� usually means not seen before, while “novel� implies a degree of originality or creativity. In AI, a result can be new (not in the training data) but not novel (it’s a trivial combination). True novelty requires a significant departure from existing patterns, which is harder to achieve and measure.

Related

• Quiet Quitting Rules: Why People Do It
• Recover Deleted Pictures on Nikon Camera
• How to Make a Facebook Page: Complete Guide

in t | Training examples | - | 1000 | 1
in d | Model parameters | - | 1000000 | 1
out novelty | Novelty score | - | (d / (t * 100)) + (t / 1000) | 2