The Freedom to Think Requires the Ability to Disagree with AI

The Freethinker has spent 145 years defending freedom of thought against forces that try to do our thinking for us. Artificial intelligence presents a stranger version of the same problem. Nobody has to censor us. Nobody has to compel belief. We can voluntarily hand over enough of the work of thinking that our capacity to disagree begins to weaken.

That risk sits underneath the Freethinker’s recent examination of AI-generated political propaganda. The obvious concern is that synthetic images and text can flood the information environment with cheap persuasion. A less visible concern comes earlier in the chain: what happens when people increasingly use the same technology to interpret the world, formulate arguments, summarise evidence, and decide what deserves attention?

Freedom of thought has always required more than access to information. It requires a mind capable of examining a claim, comparing it with an independent model of reality, and rejecting it when the two do not fit. AI can support those capacities. It can also let us bypass the practice through which we acquire them.

The problem is delegation without retained judgement

There is no sensible case for preserving every cognitive chore. I do not want to return to long division for accounting, paper maps for every journey, or card catalogues for literature searches. Technological progress often works by moving human effort upward. Once a tool reliably handles a lower-level task, people can spend more time on interpretation, synthesis, creativity, and judgement.

I have argued elsewhere that losing some skills to AI can be a feature rather than a bug. The hard question is which skills we can safely surrender. The answer depends less on whether a task feels intellectually respectable than on whether losing it leaves us unable to evaluate the system doing it for us.

Call this metacompetence: knowing what to delegate, what you still need to understand, how to check the output, and when to descend below the tool’s abstraction layer. A person with metacompetence can use AI aggressively without treating fluency as evidence. A person without it can become highly productive while also becoming more dependent on a system whose mistakes are difficult to recognise.

That distinction matters because AI performance is uneven. In Fabrizio Dell’Acqua and colleagues’ field experiment on the ‘jagged technological frontier’, consultants using GPT-4 performed substantially better on tasks that fell within the model’s capabilities. On a task outside that frontier, however, participants using AI were less likely to reach the correct answer. The danger was not that the system was useless. It was useful often enough to make misplaced trust tempting.

Learning creates a second problem. A 2025 PNAS field experiment by Hamsa Bastani and colleagues studied nearly a thousand secondary-school students using GPT-based tutors in mathematics. Students with access to a conventional GPT tutor performed better during practice, yet some of that advantage disappeared when the tool was removed for the exam. A more carefully designed tutor that gave hints rather than answers largely avoided the learning penalty. The lesson is narrower than ‘AI harms learning’. The way we delegate matters. Assistance that preserves struggle can build capability; assistance that replaces the relevant mental work can conceal its erosion.

Free thought needs a competence reserve

The Freethinker’s interviews on freedom of thought have already pointed toward this deeper issue. Susie Alegre warned that pervasive digital systems can weaken the ability to think independently, while Simon McCarthy-Jones described free thought in terms that include attention, reasoning, reflection, and the courage to think publicly. AI changes the environment in which all four capacities operate.

The useful response is not abstinence. It is a competence reserve: a deliberately maintained set of abilities we can still exercise without the tool when the stakes justify it. Pilots train for failures they rarely encounter. Cybersecurity teams practise incident response before the breach. Organisations test backup systems they hope never to use. People who rely on AI for intellectual work need an equivalent discipline.

That does not mean periodically writing an essay by hand for moral purification. It means identifying the abilities that make supervision real. Can you reconstruct the central reasoning behind an answer? Can you identify what evidence would change your mind? Can you spot a contradiction between the output and the underlying source? Can you explain why one interpretation is stronger than another? Can you recover when the model is confidently wrong?

Those are not nostalgic tests of authenticity. They are operational tests of independence.

Four kinds of friction are worth keeping

The easiest mistake is to divide work into ‘AI allowed’ and ‘AI banned’. That treats the tool as the unit of analysis. The better unit is the cognitive function.

First, keep enough pre-AI effort to form an initial model. Before asking a system to solve an unfamiliar problem, make a prediction, outline the competing explanations, or state what you think matters. The point is not to beat the machine. It is to give yourself something independent against which to compare it.

Second, preserve source contact. If an AI system summarises a study, contract, policy, or historical source that materially affects a decision, inspect the controlling document yourself. Otherwise, the abstraction layer becomes the reality layer.

Third, practise error diagnosis. When AI makes a mistake, do more than correct the final answer. Ask what kind of failure occurred: missing context, bad source selection, false premise, brittle reasoning, or simple fabrication. An error you merely patch teaches less than an error you understand.

Fourth, retain occasional tool-off performance where failure would be costly. A student may need to solve selected problems unaided. A manager may need to explain the assumptions behind an AI-assisted forecast. A researcher may need to reconstruct the evidence chain. A writer may need to defend the argument without leaning on generated prose. These checks should be selective because the goal is not ritual hardship. It is calibration.

This also changes how we should think about convenience. Some friction is waste. Some friction is training. Good tools remove the first while preserving enough of the second to keep users capable of judgement. The difference cannot be determined by asking whether AI makes the task easier. We have to ask what ability the difficulty was building.

Disagreement is the test

The most important question about AI and free thought is therefore not whether a machine can produce persuasive language. It obviously can. The question is whether the person reading that language retains enough intellectual independence to say: this is plausible, polished, and wrong.

That standard should appeal to both enthusiasts and sceptics. AI is too useful to reject wholesale, and too fallible to obey. The mature relationship is neither dependence nor abstinence. It is delegation under conditions that preserve the capacity to take back control.

The history of freethought is full of fights against external authorities that demanded assent. AI introduces a subtler possibility: we may increasingly outsource the work needed to know when assent is unwarranted. A society can enjoy unprecedented access to information while becoming less capable of independent judgement.

Freedom of thought survives only when the freedom to use powerful cognitive tools is paired with the ability to disagree with them.

Related reading

Ethical future? Science fiction and the tech billionaires, by Rahman Toone

Artificial intelligence and algorithmic bias on Islam, by Kunwar Khuldune Shahid

Capitalism Is a Nightmare. AI and Amazon Make It Worse. By Jonathan Church

The end of the world as we know it? Review of Susie Alegre’s ‘Human Rights, Robot Wrongs: Being Human in the Age of AI’, by Ralph Leonard