Say it back: readback, hearback, and trusting what machines understood
Trust is only as good as the channel it travels through, and a cheap way to check it is to repeat things back.
TOWER: Tiger six six, descend two four zero zero, cleared for NDB approach runway three three.
…
CAPTAIN: Okay, four zero zero.Flying Tiger Line flight 66, Kuala Lumpur, 19 February 1989. Lines lightly simplified from the cockpit voice recorder transcript. Tellingly, the published transcript itself writes the controller's words as "descend (two/to) four zero zero". See the official accident report, which includes the transcript.
…
The other half of trust
In the previous post, I described trust as permission to stop checking. We focused on the thing you trust: the floor, a colleague, a friend, an AI. How you come to trust it, and how you lose that trust.
But there is a second component to trust: the channel of communication. Even when you trust someone completely, what you say still has to reach them, and the other way round. Words go through air, radio, postal systems. Those are channels, and you shouldn't trust them too blindly.The picture of a sender, a channel and a receiver comes from Claude Shannon's A Mathematical Theory of Communication (1948). On why making the channel more reliable can never fully replace a check at both ends, see Saltzer, Reed and Clark, End-to-End Arguments in System Design (1984). And on a more abstract level, meaning itself is carried by words, or sometimes by cute stick figures, and those are channels too, which deserve a closer look.
Two blocks then: trust in the entity, and trust in the channel. If the channel is misleading, the most reliable person in the world can lead you straight into a hill.
What happened over Kuala Lumpur?
In February 1989, Flying Tiger flight 66, a Boeing 747 freighter, is on approach to Kuala Lumpur, over hilly ground.
The controller wants the aircraft at 2,400 feet. He says "descend two four zero zero". Said quickly, over a radio, that sounds a lot like "descend to four zero zero". Four hundred feet above sea level.
That is what the captain hears. He replies "okay, four zero zero", and starts heading down. But in the region, hills were higher than 400 feet above sea level...
Say it back: what is a readback?
What should have happened is for the tower to notice, at the slightest ambiguity, that the answer was different from the instruction. Aviation has a procedure for exactly this: the readback. When someone gives an instruction or a piece of information, the person receiving it must repeat it, word for word, and the original sender must listen very carefully.
A lot of work goes into building genuinely reliable channels. But a readback is what we may call a runtime check. You don't need to trust the channel: every readback checks that it worked this time. It tests, every single time, that the message went through faithfully.Readback is everywhere once you look. In hospitals, where it is part of closed-loop communication, orders given over the phone and critical lab results must be written down, read back, and confirmed by the person who gave them: the Joint Commission has required it for phone orders since 2003, and for critical test results since 2004. On ships, the helmsman repeats each order before turning the wheel. And when you try to delete an important resource on some websites, they ask you to type its name first.
Simple, but subtle
What I call readback actually has two equally important parts: the readback itself, and the hearback. Going back to the transcript, the readback was there: "okay, four zero zero". The error was said out loud, on the right frequency, in plain words.
But a readback is only half of the job. The original sender has to compare it with what they meant. That second half is the hearback, and without it, the readback is just politeness.
What the readback verifies is narrow, and very valuable: that what was understood is what was meant. Not that the instruction was a good one. Not that the pilot will fly well. Just that the two of you are now talking about the same thing. And it only verifies it if both ends play their part.
However, especially where the channel is mostly good and the people are trustworthy, it is tempting to get lazy about the readback and the hearback. When FAA researchers went through 49 hours of tower tapes, fewer than one readback in a hundred was wrong. But of those few errors, controllers missed more than a third.Kim Cardosi, An Analysis of Tower (Local) Controller-Pilot Voice Communications, FAA, 1994: 49 hours of tapes from ten control towers, more than 11,000 controller messages. 19 readback errors, of which controllers missed 7. When something almost always works, nobody is really listening the day it doesn't.NASA's Aviation Safety Reporting System warned about this as early as 1991: controller silence is not confirmation. A later FAA study of incident reports found that missing readbacks altogether contributed to a quarter of the communication errors analysed.
Okay then. Don't we have to just remind people to do the hearback?
Machines don't always read back
AI is writing code. AI is doing maths. AI is helping you decide. AI is ordering your pizza… Plenty of people are working on making AI trustworthy. We think another aspect is just as important to keep humans in the loop: how reliable our channel of communication with AI-enabled tools can be.
More and more often, the entity on the other end of the channel is a machine. You tell an agent what you want, it goes off and does something, and it tells you it is done. Where is the readback before the action? Sometimes it asks for your permission, with a long computer command you barely read. Or worse: a confident summary two screens long, which you skim because it looks good, without ever checking that it is also faithful to what you needed.
In formal methods, the field Clinch comes from, this problem has been known for a while, but it is (in our opinion) surprisingly overlooked. A proof assistant like Lean checks that a proof proves a statement. It cannot check, and does not pretend to check, that the statement is the one you meant. That step, from the sentence in your head to the formal statement, is a channel. And the only readback we have is a human reading the formal statement and comparing it with the original.
Lean, like most of these tools, was built by computer scientists, for computer scientists. Mathematicians are increasingly learning to use them, but the translation from what you understand to the formal language (formalisation, or autoformalisation when an AI does it) is a noisy channel. When people actually do the hearback, the results are sobering. One review of a popular benchmark of formalised competition problems found mismatches between the formal and informal statements (misformalisations) in more than half of them.miniF2F-Lean Revisited, NeurIPS 2025. See also the corrections to ProofNet (2024), and the taxonomy of misformalisations in the Formal Conjectures benchmark (2026). As AI starts formalising mathematics at scale, I expect we will find many more of these, and those issues may well extend far beyond mathematics…
A readback you can actually read
Keeping that loop as easy to close as possible is one of the things we care about most in all the tools we build. Whatever your input goes through, numerical solvers or AI assistants, you should always be able to read back what the machine understood, in your own notation and words, and check for yourself that everything is as you meant.
That sounds obvious, but it has a precise technical name: Pollack-consistency.Named by Freek Wiedijk, Pollack-inconsistency (2012), after Randy Pollack's How to believe a machine-checked proof (1998). If it fails, the readback is broken and your hearback is worthless.
A closer look. A system is Pollack-consistent if it never shows things in a way it wouldn't understand back. What it means in practice is that if you copy what it shows you, and paste it back into it, it will understand the exact same thing it meant. For instance, if it is 19:30, 7:30pm, and your computer displays 7:30 only, when you copy that and paste it into a date field, it may understand 7:30am, breaking Pollack-consistency.
And we do not stop there: if the readback is faithful but tedious to read, you'll skip it, and the loop won't be closed either. Building that is as much UX and cognitive science as it is computer science, but more on that in a later post!
Spot it in the wild
Next time you spell out your phone number to someone, think of us when they read it back to you after writing it down!
Once you start spotting it, you'll see it everywhere. Feel free to send us your best examples 🙂
If you like what you read and want to be kept in the loop too, feel free to drop your email!