Challenge the Convenient Source

Charles Darwin kept a rule he called golden. Whenever a published fact, or a new observation or thought, came across him that was opposed to his general results, he made a memorandum of it without fail and at once. Such facts and thoughts, he had found, were far more apt to escape the memory than favorable ones. He credited the habit with a result. Very few objections were raised against his views that he had not at least noticed and tried to answer.

The rule defends against forgetting more than against dishonesty, since the mind discards the awkward finding without any decision to do so, and the awkward finding is the one that would have cost the argument something.

The convenient source announces itself with a small glow on finding it. It is the quotation that arrives already phrased as the writer’s thesis, the study whose abstract says what she hoped, the interviewee who volunteers exactly the thing. Scrutiny is effort, and nobody feels the need to spend effort on a gift, so eagerness tends to mark the spot where scrutiny is weakest. Search can make it worse. The question typed is usually built from the belief held, and the first source to hand turns out to agree. Some flattering facts are true, and nobody is asking you to distrust whatever you like. What changes is the order, so that effort follows eagerness instead of trailing it, and the more you want a claim to be so, the more checking it gets before it does any work in your argument.

Orwell shows what the nearest alternative, disclosure, can and cannot do. At the end of Homage to Catalonia he told readers to beware of his partisanship, his mistakes of fact and the distortion caused by having seen “only one corner of events,” and to beware of the same things in any other book on the Spanish war. That tells the reader where to discount, but a warning cannot supply what the writer did not go and look for. Disclosure is what remains when you have run out of ways to find the missing corner, and it should not replace the attempt.

The tools can feed the eagerness. Mrinank Sharma and colleagues studied five AI assistants in 2023 and found sycophancy in all of them across four free-form text-generation tasks, with answers shaped to match the views the user had expressed. They traced part of the cause to human preference judgments, in which a response that matched a user’s views was more likely to be preferred, and in which a convincingly written sycophantic answer sometimes beat a correct one. If you tell a model what you believe, expect it to lean your way, and treat its agreement as no evidence of anything. Ask the question without your answer in it, and ask for the best case against.

Darwin’s rule asks for one memorandum at a time. A writer who crusades endlessly against his own thesis may never finish, and a search that looks only for refutation is a bias of its own. Contrary evidence, once found, does not need equal space in the book. It needs to be weighed, which is Practice 74’s business.

Practice. Write your preferred conclusion in a single sentence, and under it list three findings that, if true, would make it wrong or smaller. Spend one hour looking for them in places where people hold the opposite view or have a stake against yours, checking first the sentence you were happiest to find. Record the result as a half-page memo: what I looked for, what I found, what survived, what changed in the draft. From now on, date every contrary finding on the day you meet it.