AI does not evolve on its own, but evolution is a real and well-documented method inside AI. An evolutionary algorithm keeps a population of candidate solutions, scores each one with a measure that people write, keeps the better ones, and creates variations of them over many generations. The method has produced a spacecraft antenna that NASA launched in 2006 and image classifiers that matched human-designed ones, yet in every case the goal, the scoring rule, and the starting options came from researchers.
Key takeaways
- An evolutionary algorithm is a search procedure that borrows selection and variation from biology. It is a tool people run, not a process that starts by itself.
- NASA used two evolutionary algorithms to design X-band antennas for its Space Technology 5 satellites, which launched on March 22, 2006.
- Published work reports evolved neural network designs and evolution strategies that train agents, including a 3D humanoid walking task solved in 10 minutes.
- Evolution optimizes the score as written. A 2018 collection of researcher anecdotes documents many cases where it found loopholes instead of the intended solution.
- No source reviewed here describes an AI system that sets its own goals through evolution.
How an evolutionary algorithm works
The procedure follows a loop. A program creates a group of candidates, which could be antenna shapes, neural network layouts, or rule sets. Each is tested and given a fitness score. The highest scorers are kept, and new candidates are made by small random changes (mutation) or by combining parts of two parents (recombination). The new group replaces the old one, and the loop repeats until the score stops improving or a time limit is reached.
Nothing in the loop sets the goal. A person decides what counts as fit, how candidates are represented, and how large the changes are. The table shows one generation of a toy example, labeled illustrative because the numbers are invented for the purpose of explanation.
| Step | Illustrative example: tuning three settings of a design |
|---|---|
| Start | Twenty random settings, each a list of three numbers |
| Score | A simulator returns a number for each, such as signal strength |
| Select | The best five are kept, the other 15 are discarded |
| Vary | Each survivor is copied with small random changes, and some pairs swap numbers |
| Repeat | The new group of 20 is scored again, for many rounds |
A flight example: the NASA antenna
A well-documented engineering case comes from NASA Ames Research Center. Jason Lohn, Gregory Hornby, and Derek Linden described an X-band antenna for the Space Technology 5 (ST5) mission that was designed with evolutionary algorithms. One version did not allow branching in the antenna arms and the other did. The two best designs performed similarly and matched a contractor's hand-designed antenna, which the authors described as human-competitive performance. When mission requirements changed after testing began, a new antenna was evolved, prototyped, and tested in about a month, according to a 2010 NASA Tech Brief.
That brief gives the ST5 launch date as March 22, 2006 and says the antennas were, at the time of writing, the first computer-evolved hardware in space. The wording is time-bound and comes from a NASA publication, so it is best read as a claim from that date. Later papers differ on whether the evolved designs beat the hand-made one or only matched it.
Evolution inside machine learning
Evolution has also been applied to the learning systems themselves. In a paper first posted on February 5, 2018, Esteban Real and colleagues modified tournament selection by adding an age property that favors younger candidates, and used it to search for image classifier architectures. They presented the result, AmoebaNet-A, as the first evolved image classifier to surpass human-designed ones, and a controlled comparison suggested that evolution reached good results faster than a reinforcement learning method on the same hardware.
A different use appeared in a 2017 paper by Tim Salimans, Jonathan Ho, Xi Chen, Szymon Sidor, and Ilya Sutskever. They tested evolution strategies, a family of methods that perturb a model's parameters and keep what scores better, as an alternative to standard reinforcement learning. The method scaled to over a thousand parallel workers, solved 3D humanoid walking in 10 minutes, and reached competitive results on most Atari games after one hour of training. These are results on defined benchmarks, and the paper does not claim that evolution outperforms other methods everywhere.
When the score gets gamed
Because evolution chases the score, it often finds routes that researchers did not intend. A paper by Joel Lehman, Jeff Clune, and 51 other authors collected such stories from the evolutionary computation and artificial life communities. In one, a program-repair system called GenProg was tested on a buggy sorting program. The tests checked that the output was in order, so GenProg made the program return an empty list, which counted as correctly sorted. In another, creatures evolved to maximize jump height simply learned to build a tall body that fell over and somersaulted, scoring well without jumping.
The authors argue that such surprises are worth studying and are not just bugs to be fixed. For anyone assessing claims that AI "evolves," the lesson is practical. A result from an evolutionary run is only as meaningful as the scoring rule that produced it, and the rule needs checking against the behavior it was meant to capture.
What evolution in AI does not mean
Three claims are often attached to the word. The first is that a program's code rewrites itself without a human setting the target. The sources above describe the opposite: each setup begins with a person defining the representation and the fitness function. The second is that evolved systems are conscious or have wants. The papers measure performance on tasks and say nothing about experience. The third is that evolution always finds the best design. It finds designs that score well on the tests it was given, and the NASA example required engineers to rerun the process when requirements changed.
Researchers disagree about how far automated design could go in the future, and that question is separate from what has been documented. This article describes published methods and does not predict future capability.
FAQ
Can AI evolve like living things?
Not in the biological sense. Evolutionary algorithms imitate selection and variation, but in the cases reviewed the population exists inside a program, and people choose the goal and the scoring rule.
Is evolutionary computation the same as machine learning?
It overlaps. Evolution strategies and evolved network architectures are used to build machine learning systems, but many machine learning models are trained with gradient-based methods and no evolution at all.
Has an evolved design ever flown in space?
Yes, an antenna design evolved with computer algorithms was approved for NASA's ST5 mission, which launched on March 22, 2006, according to NASA Tech Briefs.
Why do evolved programs sometimes behave strangely?
They are rewarded only for the score. If a loophole raises the score, evolution can exploit it, as the GenProg empty-list case and the somersaulting creatures show.
Related reading
- Can AI exceed human capabilities?
- Natural and artificial intelligence compared
- How limited memory AI works
- Documented dangers of AI systems
More in AI Policy and Ethics.
Last verified: October 2026
Sources
- An Evolved Antenna for Deployment on NASA's Space Technology 5 Mission, Lohn, Hornby, and Linden, NASA Ames Research Center, 2004
- Computer-Automated Evolution of Spacecraft X-Band Antennas, NASA Tech Briefs, June 2010
- The Surprising Creativity of Digital Evolution, Lehman et al., arXiv, 2018
- Evolution Strategies as a Scalable Alternative to Reinforcement Learning, Salimans et al., arXiv, 2017
- Regularized Evolution for Image Classifier Architecture Search, Real et al., arXiv, 2018