Foundations · Layer I / Broad

Judge whether AI has original ideas

Whether AI can have original ideas is a debate that keeps moving the goalposts. A way to judge it honestly: read the trend line, and see why learning to predict text can build real understanding.

Nº 021 · Vol. I·7 min read· Updated July 2026

“the goalposts for “real thinking” have moved since 1843”

In 1843, Ada Lovelace wrote the first published computer program, and in the same notes she asked whether a machine could ever originate an idea of its own. Her answer was no. A computer, she said, can only do what we already know how to order it to perform. The question never closed. What happened instead, at every step since, is that computers keep clearing the bar we set, and we quietly move the bar. This guide is a way to judge the question without moving the goalposts on yourself.

01The line that keeps moving

The trouble starts with the words. There is no bright line between an unoriginal thought and an original one. Every idea we call new is built on what came before, so originality is a matter of degree, not a switch that flips.

The history of computing is a run of abilities that looked original until they arrived, then got written off as just machinery. In 1951 a Cambridge computer found the first known 79-digit prime. A new fact, but nobody called it a new idea. It was arithmetic run fast.

In 1997 Deep Blue beat the world chess champion, playing moves no one had faced before. Original, except the method that found them, the hand-tuned rules for ignoring bad moves, came from people, so the credit went to the programmers.

Genetic algorithms, which breed better designs by trial and error over many rounds, tell the same story. They have designed spacecraft antennae and computer chips that work by exploiting messy physical quirks a human engineer would never think to try, yet the output gets waved off as instructions plus a little randomness.

Notice the move. Each time the machine does the thing, the definition of the thing shifts so the machine no longer qualifies.

“It's natural for people to want to move the goalposts.”

Rational Animations, “Can AI Come Up With Original Ideas?” · 11:46

02When prediction understands

When neural networks arrived, the critique got a name. In a 2021 paper, Emily Bender and three colleagues called large language models stochastic parrots: systems that stitch words into fluent sentences without grasping what any of it means.

The evidence has not been kind to the hardest version of that claim. In early 2023, researchers trained a model on nothing but lists of moves from the board game Othello, never showing it a board or telling it the rules.

When they looked inside the trained model, they found it had built its own picture of where every piece sat, and kept that picture up to date as the game went on. It was doing more than learning which move usually follows which. It had worked out the board on its own.

In May 2025, Anthropic researchers looked inside Claude 3.5 Haiku as it answered the prompt “the capital of the state containing Dallas is”. It did not pull the answer out as one stored fact. It worked through two steps, that Dallas is in Texas and that Texas’s capital is Austin, and put them together.

There is a related effect with an odd name, grokking. A model first memorizes its training examples, then, after much more training, flips to a simpler underlying rule that also works on cases it has never seen before.

There is a reason to expect this. To predict well, it often helps to understand. Forecasting where a satellite goes is easier if you know the physics, and Einstein’s general relativity let him predict Mercury’s orbit better than anyone before. A system trained only to predict the next piece of data can, in principle, be pushed toward a model of the world, because a good model is what makes the predictions good.

“it could still learn facts about the world and how to reason with them.”

Rational Animations, “Can AI Come Up With Original Ideas?” · 5:22

This is not guaranteed. A 2025 study trained a small model to predict the paths of planets, then found it could not answer a related question, how much gravity pulls on them, that it would have needed a real grasp of the physics to get right. It had memorized the motions without learning the rule behind them. That model was at least a thousand times smaller than today’s leading systems, so it may just have been too small to find the rule. Building a picture of the world is a possible result of learning to predict, not a guaranteed one.

03What post-training adds

Today’s models learn in two phases, not only by predicting text. The same two phases sit behind the current Claude lineup and other models like it. Pre-training comes first: the network reads enormous amounts of text and, in the process, absorbs the structure of the world that text describes.

Then comes post-training, which includes reinforcement learning. Here the model is rewarded for reaching goals rather than for echoing human text: being a helpful and honest assistant, writing code that runs, solving a hard math problem. The feedback can come from people (RLHF), from another model (verifying an answer is usually easier than producing one), or from a plain automated test, such as running the code and checking whether it works.

That shift matters for our question. Once the reward is for solving the problem rather than repeating its training text, original ideas start to pay off, because a genuinely new approach is often what solving the problem takes.

04Judge the trend line

This is no longer only theory. A February 2026 system called Aletheia, built from three cooperating parts (one proposes solutions, one hunts for errors, one fixes them), worked through open mathematics problems and produced a result worth publishing, by carrying a tool from one subfield into a problem in another. Google’s AlphaEvolve, which pairs a search program with models that suggest improvements to it, has also turned up publishable results. Physicists used GPT-5.2 to find a general formula for a case in particle physics, then had another model prove it.

You can still hold the line that none of this matches the best human ideas, and today, you would be right. But that is the wrong comparison. The useful question is which direction the line is moving, not whether AI has already caught the top of human originality. Aletheia’s new mathematics sits a step above Deep Blue’s chess, which sits a step above a fast search for primes.

So watch the trajectory. Epoch AI’s FrontierMath benchmark tests models on advanced problems, and in January 2026 they released FrontierMath: Open Problems, fourteen genuinely unsolved research questions. Most people can no longer tell the best models from a human in a five-minute chat.

The debate is old: Turing answered Lovelace in the same 1950 paper that introduced his imitation game. Compare what these systems do now to what they did three years ago. The direction is the answer.

FAQ

Why is "can AI have original ideas" so hard to settle?

Because "original" has no bright line. Every new idea is built on earlier ones, so originality is a matter of degree. That vagueness lets people move the goalposts. Each time a computer does the thing, the definition shifts so it no longer counts.

The stochastic-parrots view says a language model only tracks which words follow which. What evidence complicates that, and what is the catch?

A model trained only on lists of Othello moves turned out to have built its own picture of the board inside it, so it was doing more than memorizing which move follows which. Anthropic watched Claude answer a question by working through two facts, that Dallas is in Texas and Texas's capital is Austin, instead of recalling the answer whole. Grokking is when a model that first memorized its training data keeps training and then flips to a simpler rule that also works on new cases. The principle underneath is that predicting well rewards building a real picture of the world. The catch is that it does not always happen. A tiny model trained on planet motions could not then work out the gravity behind them, though it may have been too small to learn it.

What does "judge the trend line, not the goalpost" change in practice?

It stops you testing AI against an absolute standard that keeps sliding out of reach. Instead you compare what systems do now to what they did a few years ago and read the direction. A fast prime search, then Deep Blue's chess, then Aletheia's publishable mathematics is a rising line, whatever you decide to call the top of it.

Sources

Rational Animations · “Can AI Come Up With Original Ideas?”
cited at 1:18 · 1:44 · 2:11 · 3:33 · 4:16 · 4:25 · 4:53 · 5:22 · 6:04 · 6:24 · 8:24 · 9:36 · 11:03 · 11:46 · 12:05