12 min read

The Signal Was Never the Thing Itself, and We All Agreed Not to Mention It

The Signal Was Never the Thing Itself, and We All Agreed Not to Mention It
Photo by Joshua Hoehne / Unsplash

My colleague Todd spent eleven months studying for a project management certification. I know this because he mentioned it approximately once a week for most of those eleven months, and I don't mean that as criticism (there is something about working toward a large credential that requires a certain amount of external narration to stay sustainable, the way people training for long-distance runs announce their mileage not to brag but because telling someone makes it more real).[1] He passed. The certification arrived as a digital badge, a small rectangle of color that could be embedded in email signatures and LinkedIn profiles, and which Todd displayed with a pride that seemed slightly out of proportion to the badge's visual impact, which was modest.

I thought about that badge recently, in a way I hadn't anticipated, when I was told about three people at a friend's company laid off in the same week. The stated reason, issued through official channels in the dry language of strategic adjustment, was, in the language of official HR communication, that their work had been automated. All three had comparable certifications. All three had the kind of resume that would have, five years ago, produced multiple interviews before the week was out. One of them had the same certification Todd had spent those eleven months earning.

The certifications did not help. Which, I understand, is not exactly news. But the specific way they did not help is what I keep turning over.

There is a clean economic theory about why credentials work, which is to say why they ever worked, and the theory is almost brutally simple once you see it. In 1973, Michael Spence published a paper that would eventually help earn him a Nobel Prize, about what he called job market signaling. The problem he was addressing is that employers cannot directly observe a job applicant's actual productivity. They're buying something they cannot fully inspect before purchase. So they need signals, things about the applicant that correlate with productivity even if they don't cause it. Education became the dominant signal not necessarily because it made workers better at their jobs, but because acquiring it was costly enough that only the motivated and capable would bother. The signal worked because of what it filtered out.

Spence's framework was, as these things often are, more complicated than the summary makes it sound, but the core insight is this: a signal has value proportional to how hard it is to fake.[2] A degree from a rigorous program takes years of effort. A certification from a serious professional body requires sustained preparation. The signal carries information precisely because it costs something to acquire. Remove the cost, or introduce a cheaper way to appear to have paid it, and the signal degrades. The correlation with actual ability weakens. The information the signal carries drains away.

This is not a glitch in the system. This is the system working exactly as designed. Today, this system is encountering a condition it was never designed to handle: the cost of appearing qualified collapsing faster than the infrastructure built around that cost can adapt.[3]

What makes this stranger, though, is something that the economic model of signaling tends to underweight: credentials were never just signals. They were also identities.

Randall Collins, the sociologist who wrote what is probably the most unsparing examination of credential inflation, observed in The Credential Society that the relentless expansion of educational requirements for jobs had very little to do with the actual skills those jobs required. The evidence was not subtle. Janitors were being asked for high school diplomas. Administrative assistants began needing bachelor's degrees. The credentials kept climbing even when the work itself stayed roughly constant. What Collins argued, somewhat controversially at the time, was that credentials function primarily as social currency, tokens of group membership, markers of cultural respectability.[4] The point was not that you learned anything. The point was that you had been inside the system that conferred the badge, and the badge said something about who you were, where you belonged, what group could reasonably claim you.

This is what he purely economic account misses. The credential was not merely a signal to employers. It was a story you told yourself about the arc of your professional life. The MBA meant something. The PMP meant something. Not just as evidence of competency but as events in a life's narrative, things you had done, ways you had organized effort across time, demonstrations to yourself that you were the kind of person who completed difficult things. Todd was not spending eleven months on that certification purely because he thought it would make him more hireable. He was spending eleven months on it because of what finishing it would feel like, what it would mean that he had done it.[5]

When a signal degrades, the information it carries to the labor market diminishes. That's the economic story. But when a credential degrades, something more personal happens. The totem loses its charge. The story it was supposed to tell about you becomes a story about a system that no longer works. You did the thing and the thing turns out to have been not quite the thing you thought you were doing. That is a specific and unpleasant feeling that has no good name, which is possibly why nobody talks about it directly.[6]

Goodhart's Law describes, in terms so concise they feel almost smug, what happens when a measure becomes a target: it ceases to be a good measure. The original formulation came from British economist Charles Goodhart, writing about monetary policy, but the principle turns out to be nearly universal. The moment you announce some metric will be used to evaluate performance, people optimize for the metric. They may simultaneously degrade the underlying thing the metric was supposed to reflect. Sometimes they don't even notice they're doing it. The optimization is often unconscious, and the thing being measured is often abstract enough that the person gaming it genuinely believes they're producing the real thing.[7]

The credential economy ran on a slow-motion version of this, playing out across decades. Once hiring managers began using degrees as filters, students optimized for degrees. Once PMP certifications became standard job requirements in certain industries, the preparation for the exam became an industry of its own. This is not a story of cynicism. Most of the people doing the optimizing were doing exactly what rational actors in a given information environment are supposed to do. They were responding to the incentives correctly. But the correct response to a proxy system, pursued by enough people with enough sophistication, eventually hollows the proxy out. The measure becomes the target. The target ceases to be a good measure. Then someone invents a tool that makes the optimization so cheap and easy that the whole superstructure collapses at once.

What AI did to the credential economy was not introduce a new problem. It accelerated an existing instability to the point where it became impossible to pretend the underlying dynamic wasn't there. The resume was always, at some level, a polite fiction. A curated presentation of selected truths organized to imply competence that might or might not be present. Hiring managers knew this. Candidates knew this. The whole system was sustained by a shared agreement not to examine it too closely, because examining it closely would require replacing it with something harder, and there was never quite enough pressure to do that work.[8]

Now there is.

Here is the part I keep returning to, and is, for obvious reasons, most uncomfortable to sit with.

The people who feel the credential collapse most acutely are not, primarily, the cynics who gamed the system. They are the people who took it seriously. Todd spent eleven months because he genuinely wanted to learn the material, and because he believed the certification would mean something durable, and because he organized his effort in good faith around a signal he had every reason to expect would hold its value. He was not deceived. He was just operating in an information environment that changed faster than his investment could mature.

This is the specific shape of the loss. It is not the loss of something you had. It is the loss of something you were in the process of building toward, and the discovery, too late to course-correct, that the destination had quietly moved. Psychologists have a term for the grief that attaches to futures that never happened: ambiguous loss, described by Pauline Boss, and it has a quality of incompleteness that ordinary grief does not. With ordinary loss you can locate the thing that is gone. With ambiguous loss you are mourning an absence that is hard to point at, a trajectory interrupted, a story whose ending has been removed without announcement.[9]

Nobody told Todd the certification would matter less. The system just   shifted. And the strange thing about shifts of this kind is that they are invisible at the individual level until they aren't, and by then the investment has already been made.[10]

The question underneath all of this is not "what is the credential worth now" but "what does proof of competence even look like when the old proof system breaks."

Because the problem Spence identified hasn't gone away. Employers still cannot directly observe a job applicant's actual productivity before hire. The information asymmetry hasn't resolved. If anything it has deepened, because now the things that used to help bridge the gap, the resume, the portfolio, the credential, have all been flooded with AI-generated material that makes the signal-to-noise problem considerably worse. Filtering inputs are failing at the same moment as filtering needs are intensifying.[11]

What this points toward, though nobody has quite put together a clean institutional version of it yet, is a shift in what kind of evidence is legible at all. The credential said: I completed this process, and the process correlates with competence. The AI-generated resume says: here is a document formatted to imply I completed such processes, and good luck determining the correlation. What breaks through that noise is not more sophisticated credentials. It is evidence of actual work. Specific, observable, explainable work. Not a claim that you can do a thing, but a demonstration that you did a thing and knew what you were doing while you did it.[12]

This is not a new idea. It is basically how reputation worked before credentials existed, which is to say before information asymmetry at scale made reputation systems too slow and too local to bear the weight. The difference now is that the infrastructure for making demonstrated work portable and discoverable is finally good enough that the old reputation model could scale in ways it couldn't before.[13]

The apprentice who fixed machines in a specific shop and was known by the fifteen people who'd seen him work is now the person whose documented solutions to specific problems can be seen by anyone who needs to evaluate whether they should trust his work. Same basic information. Different reach.

Todd is fine, for what it's worth. He is good at his job in ways the certification never fully captured and never needed to, because the people he works with know him and have observed his judgment over time. The credential did not make him competent. It did not unmake him either.

But he has not mentioned the certification in a while. I noticed, actually, sometime in the fall. The email signature changed. The thing that had been central enough to narrate for eleven months just quietly disappeared.

I don't think he made a conscious decision. I think he just stopped feeling like it was pointing at something real, and once you stop feeling that, you stop showing it, and then you stop thinking about it, and eventually it is just a file somewhere in a folder you rarely open. The evidence that you did a hard thing, preserved in a format that no longer carries the weight you intended it to carry.

The credential didn't die. It just stopped being a story worth telling. And what takes its place, if anything does, is the work itself, documented clearly enough that someone who has never met you can look at what you actually did and form a judgment about whether you understood it. Which is harder than a badge. More exposed. Less controllable. And, probably, more honest about what competence actually is.[14]

  1. The narration is also, I think, a form of accountability. If you tell people you're doing a thing, the thing becomes harder to quietly abandon. This is not a bug in how Todd organized his eleven months. It's a feature, and a reasonably intelligent one. The problem is that the same mechanism that made the commitment durable also made the discovery of its diminished payoff more visible. He had been publicly building toward something. The something had changed. Neither thing was his fault. ↩︎
  2. Spence's model technically requires that the signal be more costly for low-ability workers than high-ability workers, which is the mechanism that makes the equilibrium stable. If a degree takes the same effort regardless of underlying ability, it fails as a filter and the signaling equilibrium collapses. The current situation is approximately: AI has dramatically reduced the cost of appearing to have acquired certain credentials, which is a special case of Spence's collapse condition. This also happens to be why the response "just make credentials harder to earn" will not, on its own, solve the problem. The optimization target will follow. It always has. ↩︎
  3. There is a version of this argument where the whole credential system was always a collective delusion and good riddance. I find this argument too clean. The delusion was functional. It was doing real work, routing talent through a noisy labor market with imperfect but non-trivial accuracy. The question of what replaces a functional delusion when it stops working is considerably harder than the question of whether the delusion was true. ↩︎
  4. Collins was not entirely dismissive of what credentialed programs sometimes taught. His argument was about the primary social function, which is cultural rather than technical. The badge is the point. The learning, if it happens, is adjacent to the point. This seems harsh until you look at the data on how much of what is studied in most credentialing programs actually shows up in the jobs those programs are ostensibly preparing people for. ↩︎
  5. I want to be careful not to reduce this to "Todd did it for the feeling," which is too dismissive. He also genuinely wanted to know the material. He read the books. He did the practice problems. The credential was partly a container for a real learning project. What collapsed was the container's external value, not the learning inside it. Which is, honestly, its own kind of loss: you worked hard to build something real, and the box it came in turned out to matter more to the market than the contents. ↩︎
  6. There is probably a word for this in German. There usually is. The closest available in English is something like "credential grief," describing the specific disappointment of discovering the system you invested in was less durable than you had been led to believe. It is different from standard career disappointment, which is about outcomes. This is about the rules of the game changing after you have already played several hands. ↩︎
  7. There is a subtler version of this that applies specifically to managers rather than individual contributors. When a manager uses an AI tool to generate performance reviews, or an AI to summarize the output of direct reports, or an AI to draft strategic recommendations, they are not gaming anything in the conventional sense. They are using available tools efficiently. But the management credential, the MBA, the years of ascending seniority, was partly a proxy for the judgment developed through the actual work of doing those things without AI assistance. If the AI does the work, the credential no longer tracks the judgment. The manager may not notice. That is the part that keeps me up. ↩︎
  8. This is, I think, why the conversation about AI and credentials tends to generate such intensity from people who have recently completed expensive credentialing programs. The problem is not really about AI. It is about a polite collective fiction becoming unsustainable at a specific historical moment. AI is just the mechanism by which the fiction became unsustainable. Shooting the mechanism feels emotionally satisfying and does not address the underlying condition. ↩︎
  9. Boss developed the concept primarily to describe the experience of families of people with dementia or those who had gone missing: someone who is physically present but psychologically absent, or physically absent but psychologically present. The application to professional credentials is less dramatic but structurally similar. The credential is still there. The meaning it was supposed to carry is what has gone. ↩︎
  10. The invisibility until it isn't is worth dwelling on. Signal degradation in a credential system doesn't happen at a single moment. It happens the way Hemingway described bankruptcy: gradually, and then all at once. The credential was losing value for years before Todd completed his. The people who understood this were not distributing that information in any place where Todd would have encountered it during the eleven months he was studying. The market for credential devaluation reporting is, for obvious reasons, thin. ↩︎
  11. The filtering problem is particularly acute at the entry level, where there are no observable work histories to reference and credentials were doing the most work. Junior candidates are now presenting AI-generated portfolios, AI-assisted cover letters, AI-optimized resumes, and no amount of credential-checking alone will distinguish the capable from the convincing. The hiring manager problem is genuinely hard. I don't have a clean solution to it. I am just noting that the hardness is real and symmetrical: it is equally hard for capable junior candidates to distinguish themselves and for hiring managers to find them. ↩︎
  12. The "knew what you were doing while you did it" part is doing a lot of work in that sentence and deserves expansion. The specific failure mode of AI-assisted work is not that it produces bad outputs. Often the outputs are excellent. The failure mode is that you can have an excellent output while having no model of why it is excellent, what would make it worse, or how you would change it if the requirements shifted. The demonstration that matters is not the output but the explanation of the output. Which is harder to generate with AI assistance than the output itself, at least for now. ↩︎
  13. The pre-credential reputation system was also terrible in specific ways. It was accessible only to people already inside the right social networks, it had geographic limits, it was susceptible to bias and favoritism, and it primarily reproduced existing hierarchies. Whatever replaces the credential system should preserve the democratizing function that credentials, for all their fictions, actually delivered. The replacement cannot simply be "know the right people again." That would be a regression dressed up as progress. ↩︎
  14. Though not inevitably more fair. Demonstrated work is easier to evaluate for roles that produce visible artifacts. It is harder for roles that primarily involve judgment, facilitation, or relationship management, none of which leave clean traces. The people most disadvantaged by a shift toward demonstrated-work systems are not, in general, the people who were gaming the old credential system. They are people who do genuinely valuable work that simply does not photograph well. Which suggests that whatever replaces the credential also cannot be purely artifact-based, and the hard design problem is still ahead of us. ↩︎

Richard Mangahas