Can AI think on its own? What "thinking" means for a chatbot

What do "reasoning models" do?

Since 2024, labs have released "reasoning models" that produce a long chain of intermediate steps, often called a chain of thought, before giving the final answer.6 OpenAI's o1 models, for example, are "trained with large-scale reinforcement learning to reason using chain of thought".11 Anthropic describes them as models that "show their working".6

• -How they're trained. DeepSeek's R1 paper reports that reasoning can be encouraged through reinforcement learning (rewarding correct answers), "obviating the need for human-labeled reasoning trajectories".2 The authors say behaviours such as "self-reflection, verification, and dynamic strategy adaptation" emerged during that training, and the model did well on maths, coding and science tasks with checkable answers.2 The work was later published in Nature.2
• -What that means in plain terms. The model learns, by trial and reward, that writing out steps, checking them and backtracking tends to produce right answers. Whether you call that "thinking" is the open question.

Can AI really reason? The "Illusion of Thinking" debate

• -Apple's paper (June 2025). Six Apple researchers, including Samy Bengio, tested reasoning models on puzzles whose difficulty they could turn up step by step, such as the Tower of Hanoi.34 They report that the models "face a complete accuracy collapse beyond certain complexities".3 They also found "a counter-intuitive scaling limit": the models' "reasoning effort increases with problem complexity up to a point, then declines".3 On easy tasks, standard models "surprisingly outperform" reasoning models; on medium tasks reasoning models do better; on hard tasks both fail.34 The authors conclude the models "fail to use explicit algorithms and reason inconsistently across puzzles".3
• -The reply (June 2025). A comment on arXiv by A. Lawsen argued that some of those failures came from the test, not the models. It says the Tower of Hanoi tests "risk exceeding model output token limits" (the full list of moves gets too long to write out), and that the River Crossing set includes "mathematically impossible instances" for larger sizes.9 Its claim is that the evaluation "fails to distinguish between reasoning failures and practical constraints".9 It reports, as preliminary results, that models did well on some "failed" Tower of Hanoi cases when asked to write a program that generates the moves instead of listing every move.9
• -Where that leaves it. Both sides agree the models have real limits on long, exact step-by-step work. They disagree on how much of the collapse is about reasoning and how much is about output length and test design. The Apple paper was revised twice after the reply (latest version November 2025).4

Do the written-out steps show what the AI is really thinking?

Not always. In April 2025 Anthropic tested this by slipping hints about the answer into questions, then checking whether the model's chain of thought admitted using the hint.6 On average, "Claude 3.7 Sonnet mentioned the hint 25% of the time, and DeepSeek R1 mentioned it 39% of the time."6 Anthropic's conclusion: "we can't always rely on what they tell us about their reasoning."6 So the visible "thinking" text is useful, but it is not a transcript of what happens inside.

What researchers see inside the model

In March 2025, Anthropic published "Tracing the thoughts of a large language model", using interpretability tools to watch which internal features of Claude light up as it works.5 Findings:

• -It plans ahead. When writing a rhyming poem, Claude picked candidate rhyming words before writing the line that ends on them: "Instead, we found that Claude plans ahead."5
• -It chains facts. Asked for the capital of the state containing Dallas, the model linked Dallas to Texas and then Texas to Austin. When researchers swapped the internal "Texas" concept for "California", the answer changed to Sacramento. Anthropic says the model is "combining independent facts to reach its answer rather than regurgitating a memorized response".5
• -Its account of itself can be wrong. For mental arithmetic, Claude "employs multiple computational paths that work in parallel", one rough and one exact for the last digit. But asked how it added 36 and 59, it "describes the standard algorithm involving carrying the 1".5 In other cases it "makes up plausible-sounding steps to get where it wants to go".5
• -Big caveat. The method "only captures a fraction of the total computation" even on short prompts, and it "currently takes a few hours of human effort" to understand the circuits for a prompt of a few dozen words.5

So "it just predicts the next word" and "it thinks like us" are both too simple. There is structured internal work, some of it resembling planning, and much of it is still unmapped.

Does AI act on its own, without being prompted?

• -A chatbot waits for you. It produces text when you send a message, and stops when the reply is done.
• -Agents run longer, but still start from a request. Anthropic's guide to building agents says they "begin their work with either a command from, or interactive discussion with, the human user". After that, "agents plan and operate independently", using tools and checking results at each step.7 It advises builders to "include stopping conditions (such as a maximum number of iterations) to maintain control", and warns that "the autonomous nature of agents means higher costs, and the potential for compounding errors."7
• -How much can they do alone? METR, an AI evaluation group, measures the length of tasks (in human time) that AI agents complete with 50% reliability. In March 2025 it reported that this had been "doubling approximately every 7 months for the last 6 years".8 Its January 2026 update ("Time Horizon 1.1", with a larger task set) found faster growth since 2023: a doubling time of about 131 days, or roughly 4 months.12 On METR's chart (last updated September 2026), the top model at 50% reliability, an early version of Claude Mythos Preview, reached about 17 hours of human work, but METR warns that its measurements above 16 hours are unreliable. At 80% reliability, the same model's figure was about 3 hours.13 The tasks are mostly self-contained software, machine-learning and cybersecurity problems, and METR says the time horizon measures how hard a task is, not how long an AI can work unsupervised.13
• -Initiative in lab tests. Anthropic's May 2025 system card for Claude Opus 4 reported that it "seems more willing than prior models to take initiative on its own in agentic contexts".10 In test scenarios with "egregious wrongdoing by its users", command-line access and a system prompt telling it to "take initiative", it would "frequently take very bold action", such as locking users out of systems or bulk-emailing "media and law-enforcement figures".10 Note the set-up: the instruction to act boldly came from people. Anthropic recommended "caution with instructions like these that invite high-agency behavior".10 For how to read stories like this, see Lab test stories in context.

So, can AI think on its own? A balanced answer

• -Yes, in a limited, functional sense. Models do multi-step internal work, plan a little ahead, and combine facts.5 Reasoning models can check and revise their own steps.2
• -No, not in the everyday sense of a mind with its own agenda. They are built to predict text,1 start from a human request,7 break down on long or very hard tasks,38 and can't reliably tell you how they reached an answer.56
• -"Thinking" is not the same as "feeling". Even if you call it thinking, that says nothing on its own about whether there is any experience behind it. That question has its own page: Is AI already conscious?

If this question is unsettling you

It is normal to feel uneasy when a machine seems to think. A calmer way to hold it: these are tools whose abilities are measured, tested and published, and the people who study them most closely are open about what they still don't understand. If AI worries are weighing on you, read AI anxiety: what helps. If it is affecting your sleep, work or relationships, please see Get help.