Screen Science, but for Whom?

There are plenty of reasons to read Thomas Petzinger Jr.’s Hard Landing: fare wars, improbable survival, and executives whose meetings we might prefer to experience from a considerable distance.

One of its most contemporary moments, however, begins with a travel agent sitting down at a Sabre terminal.

The customer wants to fly from LaGuardia to Los Angeles the next morning. The agent enters the cities, date, and time. American Airlines’ reservation system searches its schedules and produces a ranked list.

So far, quite sensible.

Petzinger reports that more than half of reservations came from the first line, and about 92 percent from the first screen. American understood what that meant. Its people called the work of exploiting the display “screen science.”

The sales advice was wonderfully economical: “Trust the machine.”

Bob Crandall might have enjoyed knowing how well that expression would travel.

The answer already depended on the question. Change the departure time or destination, and Sabre’s results changed. The system was responding to a particular request, with some interpretation of what would make a flight convenient.

But the traveller entered the computer’s world through a narrow opening. Cities, dates, hours, and the service criteria the system had been programmed to weigh. There might be an entire conversation behind the booking; the terminal received a few selected parts.

Even inside an enormously complicated airline system, that query represented only a small part of the traveller’s circumstances.

The opening has become considerably wider.

On the Web, a query could be accompanied by a much longer trail. What we clicked, watched, searched for, or returned to could become material for personalisation. Google’s account of its YouTube recommendation system, for example, describes using viewing and search histories among its inputs. The question immediately in front of the system was only part of what it had to work with.

We could type what we wanted. Our behaviour could provide another account, occasionally without asking whether we agreed with it.

Spending an evening investigating something does not necessarily mean we would like to spend the next fortnight being shown more of it. Curiosity can have consequences for the furniture of our screens.

A better account of our preferences can improve a recommendation. It can also improve a platform’s ability to sell us something or hold our attention. Both may happen in the same apparently helpful result.

With AI, we can supply a much more elaborate account on purpose.

We can give an assistant a book, point it towards earlier articles, and ask for a draft. When the draft sounds generic, we can say so. We can explain which part of the argument matters and ask it to stop producing sentences that could belong to anybody.

The complaint becomes another input.

In a coding session, the possibilities widen further. The assistant might have a repository, conversation history, an AGENTS.md file, and the results of commands it has run. A skill can supply a workflow for a particular task. An MCP connection can bring external information and tools into the session.

These inputs affect both what the assistant can discover and how it approaches the work.

They also interact. An instruction can change what it looks for. A tool can return something that changes its next question. A correction can redirect the argument. The conversation keeps adding to the circumstances from which the next response emerges.

The combinations multiply.

The difference from that airline query is considerable. We can now supply examples, constraints, working methods, unfinished thoughts, and disagreements with an earlier answer. Much of the particular situation can become part of the request.

And the system can compose the thing we encounter.

Sabre selected from flights already in its database. A feed selects from available posts or videos. A generative system can produce an explanation, an argument, or a proposed change to a project in response to the materials assembled around it.

We have entered something like a just-in-time world of answers.

There is an extraordinary freedom in being able to bring so much of a problem into the room. There is also the question of what has actually arrived.

A repository might be current while its instructions describe an earlier arrangement. A document might preserve a decision that was later abandoned. A leading question might invite an agreeable explanation whose agreement feels like independent support.

An answer can be particular to us and still depend on a rather poor account of us.

Here, Petzinger supplies a stubborn detail worth returning to.

American wanted its Dallas–Honolulu service to appear ahead of Braniff’s. Braniff flew nonstop; American stopped in Los Angeles and arrived two hours later. Petzinger describes internal demands to alter the display in American’s favour when ordinary adjustments to the scoring formula could not overcome that disadvantage.

The flight still took longer. Moving it up the screen did nothing for the passenger sitting through the connection.

There was a journey outside the display against which the display could be judged.

An AI answer has things outside it too: the passage in the book, the code that actually runs, the evidence behind a claim. Yet a response that incorporates our vocabulary, priorities, and previous corrections can feel unusually close to our understanding.

That closeness can make difficult material accessible. It can also make the answer comfortable to accept.

We recognise what we asked for. We recognise the qualifications we supplied. We may even recognise our own objections, now absorbed into a more polished argument.

How much of that recognition should count as a reason to trust it?

Our ability to shape the session sits alongside the provider’s decisions about training and operating rules. The sources have their own authors. The instructions may have accumulated through several people’s work. A personal answer can carry quite a crowd behind it.

American’s salespeople could point towards the terminal and say, “Trust the machine.”

Now we can help furnish the machine with the circumstances through which it answers us.

The screen has become much better at taking us into account. When its conclusions begin to sound like our own, whose science are we looking at?