The Old Signals

If you have hired anyone in knowledge work in the last year, you have probably had this moment.

You are reading a cover letter. It is good. Maybe too good. The phrasing is a little too clean. The bullet points line up a little too well. You catch yourself wondering whether the candidate wrote it or whether a model did. Either way, you have to make a decision about a person , and the signal you used to use, can this person communicate clearly in writing, just stopped working.

This is happening everywhere now. Quietly. Across hiring loops, performance reviews, slide decks, sales emails, GitHub commits, PR descriptions, customer complaints , manuscript submissions, and Tinder profiles. The signals professionals used to read each other by are noisy now. Not destroyed. Noisy. Which is worse, in some ways, because we still use them, just less reliably.

This chapter is about what happens to a workplace when the cues stop being trustworthy.

***

Every workplace ran on a set of small signals.

You could tell who was good at their job by how they communicated , by how they handled ambiguity , by how clean their code was, by how their drafts read on a first pass, by how they showed up in meetings. You could tell who was junior by how they wrote and what mistakes they made. You could tell who was a craftsperson by what they refused to do . You could tell who was a pretender by what they over-explained.

These signals were never perfect. They were biased. They were often class - coded . They missed people . But they were legible. The people who hired and managed and mentored knew how to read them. The signals were the social technology of professional work .

AI scrambled them all in about eighteen months.

***

The ladder is missing some rungs now .

Start with hiring.

There is a story you will find , in some variation , told by senior engineers all over the internet in the last year. A junior candidate sends in the coding exercise her interviewers asked her to complete at home. The code is clean . The architecture is right . The README is thoughtful. Three years ago, this candidate would have been a strong hire on this evidence alone. She gets the offer.

Then she shows up on Monday and struggles through tasks the exercise should have predicted she could handle easily. The submission had been mostly generated. The struggle had not.

This is why many companies have started quietly retiring take-home assignments and going back to live coding instead. Power days . Whiteboarding. Watching the candidate write code in front of you , in real time, with no tab open in the background. An entire common hiring practice is being abandoned in under two years, not because anyone announced anything, but because the signal stopped working and somebody had to do something.

The live-coding pivot doesn't solve the underlying problem. It just moves it. Candidates are already preparing for live interviews with AI in ways that don't survive contact with the actual job. Some are running models on a second screen during the call. The signal got noisy in one place, so the industry built a new one somewhere else, and that one is already going noisy too. This is the texture of the moment : not the death of any specific signal , but the constant work of replacing signals as fast as AI scrambles them.

This story is small but it is everywhere. Take-home assignments don't filter the way they used to. Cover letters don't. GitHub repos don't. The traditional signal for can this person actually do junior work got drowned by a tool that does junior work very well.

So companies have started hiring fewer juniors . Why pay for an entry-level engineer who will spend three months ramping up when a senior with AI can ship the same output today? The short-term math works. The long-term math is grim . Juniors are how you get seniors . Without an entry pipeline, the senior layer eventually runs out, and there is no plan for what happens then.

The ladder still exists. The bottom rungs are rotting. Nobody is being asked to fix it because the symptom is invisible until five years out , and most companies do not plan that far ahead . We are running an experiment to see what happens to a profession when its entry-level practitioners do less actual work in their first three years than any previous generation . The result , I suspect , is going to be expensive , and most of us will be retired by the time the bill comes due.

The juniors who do get hired walk into a strange situation. The work they would have done is now done by their seniors plus AI . So they get fewer reps . Fewer chances to fail. Fewer chances to build the intuition that makes them senior eventually. We are growing a

generation of professionals who skipped the awkward middle , and we have no idea yet what that will produce .

***

Oracle or toy.

On every team there is a person who has decided AI is an oracle and another person who has decided AI is a toy.

The oracle person prompts the model and accepts whatever comes out. They paste in the result with mild adjustments. They are productive on the surface. They are also routinely shipping things that are wrong, because the model is wrong sometimes , and sycophancy actively rewards the oracle reading . The model is built to sound confident, especially when it shouldn't be. The oracle person looks the same as a calibrated user from a distance. Same volume. Same speed. Different correctness rate.

The toy person prompts the model once, gets something mediocre , decides the model is useless , and goes back to working the old way . They miss real value . They also avoid real risk . Their output is slower than the oracle person's , but more reliable . They are starting to look slow next to the oracle person , even though the oracle person is shipping bugs.

The calibrated user, the one who treats AI as a useful but unreliable colleague , is rarer than either of the other two . Calibration takes work . It requires noticing when the model is wrong, building intuition for when to trust and when not to , pushing back on outputs,

re-prompting, integrating skeptically. This work is invisible to a manager looking at output volume. So the manager often promotes the oracle. The toy person looks lazy . The calibrated user looks slow. The oracle person looks fast . It is the worst possible alignment of incentive and accuracy.

The old signal was simple. Did the person produce good work? Now you have to ask a different question . Did the person produce good work because they are good , or because they prompted well, or because they accepted whatever came out and got lucky ? You can't answer this from a distance. You have to look at the work . Most managers, most of the time, don't.

***

Vinyl and slow food.

There is a third group , and they are doing something stranger.

In some circles , not using AI has become the status signal.

The artisanal refuser shows up in a code review with a hand-rolled implementation. No Copilot suggestions accepted. No Claude-assisted refactors. They mention it. They mention it casually, but they mention it. There is a small show of having written this themselves . In some communities they get praise. In some communities they get hired specifically for it.

This is interesting because the same gesture used to mean the opposite thing. Five years ago , refusing the modern toolchain meant I don't know how or I am behind . Today it can mean I have the seniority and the taste to refuse. Same refusal. Opposite read.

There is something real in the artisanal flex . Knowing how to do the underlying thing matters. The pianist who can play scales unaccompanied is still a better pianist than the one who has only ever used backing tracks . Resistance has value.

But it scrambles another signal. Refusal to use the tool used to be evidence of incompetence . Now it can be evidence of mastery. And often it is just performance. The performance and the mastery look identical from outside. You cannot tell which one you are watching, which means the artisanal flex is doing some of the same work the AI is doing , just from the opposite direction . Both make people look more capable than they are. Both rely on the observer not being able to tell .

So the team has the oracle , the toy , the calibrated user , and the artisanal refuser. Four ways to relate to the tool . Four ways the work gets done. Four signals that look similar enough from a distance that you have to do real work to tell them apart. The performance review process did not get updated . Most managers are still using the old signal book , on a workforce that no longer matches it .

***

The deeper move.

The problem isn't any one signal being broken. It is that the system of signals is broken. The professional world ran on a small social technology, the ability to read each other at distance, and that technology has temporarily stopped working .

Some of it will rebuild. New signals will emerge. We will figure out , eventually , how to tell who is calibrated from who is an oracle, how to tell a real junior from a generated cover letter , how to read the artisanal refuser correctly. Five years from now there will be tools and conventions and probably entire job categories built around interpreting AI-mediated work. Someone is going to make a fortune selling AI-resistant assessment frameworks to HR departments. Bet on it.

But right now , today , this year , we are working in a workplace where the old signals are noisy and the new signals don't exist yet. Decisions are being made on partial information. People are being hired who shouldn't be, and people are being passed over who shouldn't be. Promotions are being given based on output that may not survive scrutiny. The performance review you wrote last quarter , the one that was supposed to summarize someone's competence , was probably partly summarizing their prompting skill. You did not know which was which. Neither did they.

This is not fixable in a chapter. It is barely diagnosable. But naming it matters. Most of the friction on your team right now, the friction you cannot quite explain, is partly downstream of this. The team isn't broken. The way you read the team is.

***

Back to the cover letter.

When you finish reading it, you are no closer to knowing whether you should hire the person. You make a decision anyway , because hiring decisions get made . Half the time you are right. Half the time you are not. The miss rate used to be lower. You don't yet know what to do about it .

The old signals are noisy . The new ones haven't arrived . We are working without a map , on a project we don't fully understand, evaluating each other with tools that don't quite work anymore.

That's the world now. The book is about it.

***