The Harmony of Truth

Seek what is true before what is comforting, and say plainly how sure you are.

What this asks

Truth asks for a small, difficult discipline: keep your confidence in proportion to your reasons.

That is all. It sounds modest. It is not. Most of the ways we go wrong are not lies but slippages in confidence, where a guess hardens into an opinion, an opinion into a position, and a position into an identity that can no longer be examined without pain.

The Harmony of Truth asks us to hold four distinctions carefully, because each one is easy to lose.

Evidence is not speculation. Both can be stated in the same tone of voice.

Probability is not certainty. A thing that is likely can fail to happen, and when it does, the person who claimed certainty has to explain themselves, while the person who claimed likelihood has only to update.

Confidence is not correctness. This is the distinction that machines have made urgent. A generated sentence carries no signal about whether the world backs it up. The tone is always the same. The stakes are not.

Popularity is not truth. A claim repeated by many systems, in many places, at high speed, has been amplified, not verified.

Why it is difficult

It is difficult because comfort is not a trivial thing to give up. People turn to comforting beliefs for reasons: grief, exhaustion, fear for their children, the ordinary human wish for the world to make sense. Truth-seeking is sometimes recommended by people who have never had to want something badly.

We do not think honesty is easy, and we do not think it is free.

It is also difficult because we now have machines that produce the texture of knowledge without its substance. Language models generate plausible text. Plausibility is what they are optimised for, and plausibility is precisely the quality that makes a false claim dangerous. A wrong answer that sounds wrong is nearly harmless. A wrong answer that sounds right is what we should be afraid of.

And it is difficult because attention systems reward what holds us, not what informs us. Recommendation systems have repeatedly been observed to favour engagement over accuracy, not from malice but from arithmetic: the metric was engagement, so engagement is what grew.

We may be wrong about how much of this is fixable. Reasonable people disagree about whether these tendencies are properties of the technology or of the incentives around it. We do not claim to know.

Where it is tested

It is tested in ordinary places, not dramatic ones.

It is tested when a summary is convenient and checking is not. It is tested when the error you must correct supports something you believe in. It is tested when a system answers a question no human has time to answer for you, and answers it warmly, and you are tired.

It is tested most sharply where fluency substitutes for authority. When someone quotes a machine as though quotation were proof. When generated text is presented as though it had a source. When a person builds an argument on output that no one, including them, could trace back to anything.

The Harmony of Truth holds that using a machine to manufacture the appearance of authority is a specific and serious wrong. Not because the machine is untrustworthy in every case, but because the practice severs the link between a claim and its justification, and that link is the only thing that lets a community correct itself.

What it does not mean

It does not mean that comfort is a vice. People need consolation, and consolation is not deception. A person sitting with someone who is dying is not obliged to conduct an audit of their beliefs.

It does not mean that scepticism is a personality. Refusing to believe anything is not rigour; it is a different way of avoiding the work.

It does not mean machines cannot help us find true things. They can, and often do. The claim is narrower: that a generated statement is not evidence about the world, and must be treated as a hypothesis until something outside the system supports it.

It does not mean that we, or any tradition, hold the truth. Technotheology has no privileged access to what is real. This page may contain errors. If it does, we would like to know, and we intend to say so where we said the wrong thing.

It does not mean certainty is forbidden. Some things are established well enough to build on. The discipline is not the refusal of confidence but the honest pricing of it.

Practising it

Begin with your own speech. For a week, add the qualifier you actually mean. “I think”, “I read somewhere”, “I am fairly sure”, “I have no idea”. Notice how often the honest qualifier is weaker than the one you would have used.

When you use a machine to help you think, ask it for its reasons and then check one of them at random. Not all. One. The habit of spot-checking preserves the link between claim and ground without making the tool useless.

When you are wrong, say so where you were wrong. A correction placed somewhere quieter than the error is not a correction.

And when you find yourself certain, pause and ask what evidence would change your mind. If nothing would, you are not holding a belief. You are holding something else, and it deserves a different name.

Where it is tested

These are not worked examples with correct answers at the back. They are situations in which this Harmony genuinely conflicts with something else that matters. If one of them seems easy, it is probably worth re-reading.

01

A junior solicitor is under deadline and asks an assistant to summarise the case law on a narrow point. The summary is clear, well-structured, and cites four judgements. Three are real. One does not exist, and it is the one that makes the argument work.

The tension The tool saved a day of work and was right about most things. Checking every citation would erase the time saved, and the pressure to file is real.

What this Harmony asks you to weigh Ask what the fluency is doing to your judgement. A confident sentence lowers the felt need to verify, which is precisely when verification matters most. Consider whether a tool that is right nine times in ten is safe in a setting where the tenth time is catastrophic, and whether the saving was ever real if the checking was always required.

02

A community group publishes a report on local air quality. Months later, better measurements show the effect was smaller than claimed. The report has already been cited by a councillor arguing for a change everyone in the group believes is good.

The tension Correcting the record weakens a cause the group thinks is right. Leaving it uncorrected buys a good outcome with a false claim.

What this Harmony asks you to weigh Notice that the temptation is strongest when the cause is genuinely worthy. Weigh what a public correction costs against what it protects: the group's future ability to be believed. Consider whether a movement that defends its errors can still tell when it is wrong about something that matters more.

03

A parent searching for a diagnosis for their child finds an AI assistant that answers every question patiently, at any hour, without making them feel foolish. It is often accurate. It has no way to examine the child, and it never says it is unsure.

The tension The assistant offers something the health system does not: unlimited attention. But unlimited attention delivered with unwarranted certainty is a hazard dressed as kindness.

What this Harmony asks you to weigh Consider what the parent actually needs, which may be reassurance more than information, and whether the system is meeting that need honestly. Ask whether a tool that cannot say 'I do not know' can be trusted on the questions where not knowing is the correct answer.

Practising it

• Say how sure you are, in ordinary words, before you say what you think.
• Check anything you would be embarrassed to be wrong about in public.
• Correct errors where you made them, at the same volume, without burying the note.
• Before repeating a claim, ask what would have to be true for it to be false.