Prompting gives modern language models a richer channel, but it still uses a batch protocol inherited from punch cards.
2
The hard work of deciding context, timing, ambiguity, and repairs remains with the human even when the model can reason and speak.
3
AI interfaces should participate in conversations by listening, yielding, clarifying, tracking speakers, and choosing when to act.
Summary
Ted Johnson argues that AI has improved what people can express to computers without changing the basic interaction protocol. A prompt is still a packaged request that gets submitted and answered, much like a punch-card job. Faster responses and voice transcription do not make the exchange genuinely interactive. The result is that people carry the work of selecting context, repairing misunderstandings, deciding timing, and identifying what matters. Johnson shows examples of systems that interrupt, yield, backchannel, track who is speaking, resolve requirements, and act when the conversation gives them enough information. He asks designers to treat AI as an interface technology as well as an intelligence technology. The right interface may use chat, voice, questions, pauses, sketches, checklists, or silence, depending on the moment. Its job is to reduce the translation, precision, context, and repair work that people have historically done for machines.
Familiar computer interfaces are inherited compromises
Johnson starts with the keyboard, an interface that feels natural only because people learned it through lessons and practice. The layout used today traces back to a patent from about 1860, and he points to alternatives such as Dvorak, Colemak, and the Hansen Writing Ball. These devices were designed under constraints that no longer exist, yet people still place them between themselves and powerful computers. He calls the medium an interface gives people to work in the channel. A keyboard, microphone, screen, punch card, and prompt box are all channels. Each carries different physical signals, such as discrete text, vocal timing and hesitation, or spatial relationships in a diagram.
AI expanded expression without changing the channel
The keyboard and submit box did not change when large language models arrived. What changed was the range of meaning people could send through them. Earlier computer interfaces offered fixed vocabularies: assembly opcodes, shell commands and flags, or programming-language primitives. Natural language opened that menu so people could express context, nuance, and intent in ordinary words. Johnson separates this expressive capacity from understanding. A channel can carry more signal without the machine understanding it. The central mismatch is that computers can now receive a much richer request while people still use an old physical path and the same basic exchange pattern.
Prompting is a batch protocol inherited from punch cards
Johnson calls prompting the protocol of a punch card. With punch-card computing, a person encoded a complete request away from the machine, submitted the deck to an operator, waited, read the result, corrected an error, and submitted the job again. A prompt follows the same shape: assemble the request, submit it, wait, inspect the answer, and package another turn when something is wrong. Interactive features and shorter waits add what he calls "interactive sprinkles," but they do not change the protocol. Voice input also remains batch when speech is transcribed into a box and submitted as a turn.
Prompt engineering teaches people to package good batch jobs
Johnson says prompt engineering is a set of learned rules for packaging batch requests. People are told to ask for step-by-step thinking, provide examples, assign or avoid an expert role, add or remove context, and communicate through particular document formats. He compares these instructions with the knowledge a punch-card operator needed to assemble a deck that would run successfully. He is not calling prompts or punch cards bad. Both were effective responses to the limits of their time. His question is whether batch remains the right protocol for a machine that can ask follow-up questions, clarify missing information, and engage during a thought.
The old protocol makes users do the interface work
As model abilities in reasoning, speech, vision, memory, and planning increase, the interface remains a box and a submit button. Johnson lists the burdens that stay with the human: deciding which context matters, remembering what to ask, choosing the timing, noticing ambiguity, repairing the answer, and engineering the prompt. When the result is wrong, people often conclude that they are bad at AI or insufficiently specific. Johnson rejects that conclusion. The problem is asking people to operate a new kind of intelligence through a punch-card protocol. The interface still feels like work even when the model feels capable.
Real conversation requires turn-taking and speaker awareness
Johnson gives a voice-mode example in which his co-founder asks when the next Timberwolves game is, then says "Hey Ted, come on in" as if addressing a person who has arrived. The AI treats that speech as its next turn because it cannot tell who is speaking or whether the words are meant for it. Johnson describes the system as having exactly one slot, a user message followed by a reply. He then shows PersonaPlex responding to interruptions, yielding, and picking up a thread again. Backchannel sounds help, but they do not by themselves establish who is in the room or which speech belongs to the AI.
A participating AI can follow group conversation and act at the right moment
In Johnson's group-conversation example, the system labels statements as questions, proposals, or answers and waits until no one else is speaking or holding the floor. It tracks direct references to the AI and follows a discussion about expense approvals, including the first-release scope and an approval threshold. When asked to capture the decision, it summarizes that expense approvals are in scope, access requests are out of scope, managers can approve or reject from a notification, and amounts over the stated threshold need a second approver. When a participant changes the threshold, the AI asks for confirmation before updating the requirement.
AI interface design should choose the human affordance for the moment
Johnson asks designers what burden people still carry only because machines used to be too limited to carry it. He says the answer is not always chat, voice, or a wall of markdown. People communicate through questions, pauses, sketches, checklists, quiet asides, and silence. An AI interface should understand timing and modality instead of making the user choose them every time. Johnson describes past computing interfaces as carrying translation, precision, context, and repair taxes from one era into the next. AI could reduce those taxes without removing human judgment or turning every interaction into voice.
"AI is not just an intelligence technology. It's increasingly becoming an interface technology."17:31
Who should watch
You design or build chat, voice, agent, or collaboration products and need a way to think about turn-taking, interruption, and timing.
Your team is teaching users prompt techniques because the system cannot ask for context or repair misunderstandings on its own.
You want to reduce the work people do to translate intent into commands, forms, prompts, or other machine-oriented formats.
