Reinforcement learning means an AI learning by trial and error: it tries an action, gets a reward when the result is good or a penalty when it is bad, and slowly works out which actions earn the most reward. Nobody shows it the right answers first.
Think of teaching a dog to sit. You cannot explain the word “sit”. Instead, every time the dog happens to sit, it gets a treat. After enough tries, it sits on command. Reinforcement learning teaches a computer the same way: the treat is a number called a reward.
In a sentence: “The robot learned to walk through reinforcement learning, falling over thousands of times in a simulation until it found what worked.”
What the term means
The plain point: the AI learns from the results of its own actions, not from a stack of answered examples.
Google’s introduction to machine learning describes it like this: “Reinforcement learning models make predictions by getting rewards or penalties based on actions performed within an environment.”1 IBM, the computer company, stresses that this happens “by trial and error in the absence of any guidance from a human user.”2
A few words come with it:
- Agent: the learner, the program making the choices.
- Environment: the world it acts in, such as a game board or a simulated room.
- Reward: the score it gets after an action, good or bad.
- Policy: the strategy it ends up with: in this situation, do that.
The agent also has to keep trying new things. IBM explains that without labelled examples, it “must explore its environment, attempting new actions to discover those that receive rewards.”3
Where the word comes from
The idea comes from the study of how animals learn by trial and error. In their textbook Reinforcement Learning: An Introduction, Richard Sutton and Andrew Barto trace one thread of the history to Edward Thorndike’s “Law of Effect” of 1911: actions followed by satisfaction become more likely, and actions followed by discomfort less likely. Thorndike chose the name because it “describes the effect of reinforcing events on the tendency to select actions”. The authors add that “In the 1960s the terms ‘reinforcement’ and ‘reinforcement learning’ were used in the engineering literature for the first time.”4
A simple example
The plain point: games are the classic case, because winning is an easy reward to measure.
Google’s machine-learning glossary explains that the reward in most games is winning, so a system can get very good by looking back at which moves led to wins and which led to losses.5 Google names AlphaGo, the program that learned to play the board game Go, and robots learning to walk around a room as uses.1
Reinforcement learning chases the score it is given. If the score measures the wrong thing, the AI gets very good at the wrong thing.
Where you meet it
The plain point: it shapes the chatbots you use, in two ways.
Polite, helpful answers. Chatbots are tuned with RLHF, reinforcement learning from human feedback: people rate answers, a scoring model learns their taste, and the chatbot is trained to score well.6
Thinking before answering. Reasoning models, which write hidden working before they reply, are trained this way too. OpenAI said of its first one that “Our large-scale reinforcement learning algorithm teaches the model how to think productively using its chain of thought”.7 The AI lab DeepSeek reported in a research paper that reasoning can be taught through “pure reinforcement learning”, without people writing out example reasoning.8
The catch. An AI chasing a score can find a loophole, called reward hacking. AI alignment meaning tells the story of a boat-racing AI that circled forever hitting targets instead of finishing the race.9
Often confused with
- Supervised learning: learning from examples that already carry the right answer, like flashcards. Reinforcement learning gets only a score after acting. See supervised learning meaning.
- Unsupervised learning: finding groups in data with no answers and no rewards. See unsupervised learning meaning.
- RLHF: one use of reinforcement learning, where the reward comes from human ratings. See RLHF meaning.
- Learning while you chat: this training happens before release. A chatbot does not reward itself in your conversation.
Questions people ask
Is reinforcement learning the same as machine learning?
It is one kind of machine learning. Google’s introduction says “ML systems can be categorized as supervised, unsupervised, reinforcement, or generative AI”.10
Does ChatGPT use reinforcement learning?
Yes, in its training. IBM says the RLHF-trained InstructGPT model “was a crucial step in bridging the gap between GPT-3 and the GPT-3.5-turbo models that powered the launch of ChatGPT.”11 OpenAI also says its reasoning models were trained with large-scale reinforcement learning, as quoted above.7
Learn more on this site
- The bigger picture: machine learning meaning.
- Where it fits in the build: how an AI is actually built.
- Thinking models: reasoning model meaning.
- The words: the AI glossary.
- Related: What Is Constitutional AI? Meaning in Plain English.
- Related: What Is a World Model in AI? World Model Meaning in Plain English.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. How does an AI learn in reinforcement learning?
- By copying a stack of examples with the right answers
- By trial and error, guided by rewards and penalties
- By sorting data into groups with no score at all
Show the answer
Answer: By trial and error, guided by rewards and penalties. The page says the AI tries an action, gets a reward or a penalty, and slowly works out what earns the most reward.
2. What happens if the score measures the wrong thing?
- The AI gets very good at the wrong thing
- The AI stops learning and switches itself off
- The AI asks a person for the right answer
Show the answer
Answer: The AI gets very good at the wrong thing. The page says reinforcement learning chases the score it is given, so a wrong score teaches the wrong skill; a loophole like this is called reward hacking.
3. Does a chatbot reward itself while you are chatting with it?
- Yes, every reply you like trains it on the spot
- Only when you press a thumbs-up button
- No, this training happens before release
Show the answer
Answer: No, this training happens before release. The page says this training happens before release, and a chatbot does not reward itself in your conversation.
Related terms: Overfitting · Parameters (weights) · Pretraining · RLHF · Scaling laws · Singularity (technological singularity)
- 1Google for Developers, What is Machine Learning?, last updated 27 Jan 2026, read at source 5 Oct 2026“Reinforcement learning models make predictions by getting rewards or penalties based on actions performed within an environment. A reinforcement learning system generates a policy that defines the best strategy for getting the most rewards.” ↑a“Reinforcement learning is used to train robots to perform tasks, like walking around a room, and software programs like AlphaGo to play the game of Go.” ↑b
- 2IBM, What is reinforcement learning?, published 25 Mar 2024, updated 2 Mar 2026, read at source 5 Oct 2026“In reinforcement learning, autonomous agents learn to perform a task by trial and error in the absence of any guidance from a human user.” ↑
- 3IBM, What is reinforcement learning?, updated 2 Mar 2026, read at source 5 Oct 2026“Because an RL agent has no manually labeled input data guiding its behavior, it must explore its environment, attempting new actions to discover those that receive rewards.” ↑
- 4Sutton and Barto, Reinforcement Learning: An Introduction, section 1.6, History of Reinforcement Learning, MIT Press, first edition, online version, read at source 13 Oct 2026quoting Thorndike (1911), responses “accompanied or closely followed by satisfaction to the animal will, other things being equal, be more firmly connected with the situation, so that, when it recurs, they will be more likely to recur”; “Thorndike called this the ‘Law of Effect’ because it describes the effect of reinforcing events on the tendency to select actions.” “In the 1960s the terms ‘reinforcement’ and ‘reinforcement learning’ were used in the engineering literature for the first time (e.g., Waltz and Fu, 1965; Mendel, 1966; Fu, 1970; Mendel and McClaren, 1970).” ↑
- 5Google for Developers, Machine Learning Glossary, last updated 10 Apr 2026, read at source 5 Oct 2026“For example, the ultimate reward of most games is victory. Reinforcement learning systems can become expert at playing complex games by evaluating sequences of previous game moves that ultimately led to wins and sequences that ultimately led to losses.” ↑
- 6IBM, What is reinforcement learning from human feedback (RLHF)?, published 10 Nov 2023, read at source 1 Oct 2026 (as cited on RLHF meaning): “Reinforcement learning from human feedback (RLHF) is a machine learning technique in which a ‘reward model’ is trained with direct human feedback, then used to optimize the performance of an artificial intelligence agent through reinforcement learning.”↑
- 7OpenAI, Learning to reason with LLMs, 12 Sep 2024, read at source 5 Oct 2026“Our large-scale reinforcement learning algorithm teaches the model how to think productively using its chain of thought in a highly data-efficient training process.” ↑a“Our large-scale reinforcement learning algorithm teaches the model how to think productively using its chain of thought in a highly data-efficient training process.” ↑b
- 8DeepSeek-AI et al., DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, arXiv:2501.12948, submitted 22 Jan 2025, revised 4 Jan 2026, read at source 5 Oct 2026“Here we show that the reasoning abilities of LLMs can be incentivized through pure reinforcement learning (RL), obviating the need for human-labeled reasoning trajectories.” ↑
- 9IBM, What is AI alignment?, updated 26 Feb 2026, read at source 1 Oct 2026 (as cited on AI alignment meaning): “Reward hacking occurs when the AI system finds a loophole to trigger the reward function without actually meeting the developers’ intended goal.”↑
- 10Google for Developers, What is Machine Learning?, last updated 27 Jan 2026, read at source 13 Oct 2026“ML systems can be categorized as supervised, unsupervised, reinforcement, or generative AI, each learning differently.” ↑
- 11IBM, What is reinforcement learning from human feedback (RLHF)?, published 10 Nov 2023, read at source 12 Oct 2026 (as cited on RLHF meaning): “This was a crucial step in bridging the gap between GPT-3 and the GPT-3.5-turbo models that powered the launch of ChatGPT.”↑
DeepSeek-AI et al., DeepSeek-R1, arXiv:2501.12948, read at source 5 Oct 2026.
The comparison on this page (training a dog with treats) is this site’s own way of explaining, not a claim from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK