You have almost certainly made this inference, probably this week. A proposal arrives that is thorough, well-structured, and carefully written. Before evaluating a single argument in it, you formed an impression: someone took this seriously. The polish told you something about the person, not just the document. Their investment was evidence of their conviction.

That inference was reliable for a long time, and it has quietly stopped being reliable. Not because people stopped caring, but because the thing you were reading as evidence of care now costs almost nothing to produce. The polished document and the perfunctory one look identical, because they can be generated by the same two-minute action. And this matters far beyond documents, because organizations run on inferences of exactly this kind, dozens of times a day, mostly without noticing they are making them.

Why effort carried information

The mechanism here is well understood and won a Nobel Prize. In 1973, the economist Michael Spence published Job Market Signaling, and in 2001 shared the Nobel Prize in Economics for work on markets where one party knows more than the other. His insight was that when quality cannot be observed directly, people communicate it through actions that are costly, and specifically through actions that cost less for the high-quality type than the low-quality type.

Education was his example. Even if a degree taught nothing useful, it could still convey real information, provided that capable people found it easier to obtain than less capable people did. The signal works because of the cost differential, not the content. Remove the differential and the signal stops separating anyone from anyone, regardless of how impressive it looks.

That last sentence is the whole issue. A signal is not informative because it is expensive. It is informative because it is differentially expensive. And generative AI reduced the cost of producing polished, thorough, articulate output to roughly the same near-zero for everyone, whatever their underlying quality, care, or competence.

What this actually costs an organization

The hiring case is the most visible, because the collapse happened fast enough to measure. LinkedIn reported roughly a 45% surge in applications submitted through its platform in a single year, reaching an average of about 11,000 applications per minute, while job postings declined. The Financial Times reported employers and recruiters estimating that as many as half of applicants were using generative AI, with one recruitment platform chief executive describing more than double the number of candidates per role and, in her words, higher volume and lower quality. Recruiters told the New York Times that applications tailored by AI to stand out have instead become suspiciously similar to each other.

So the cover letter has stopped working as a signal. It used to cost a candidate an hour and some anxiety, and that cost was higher for someone who did not much want the job. Now it costs a prompt. The document still exists, still gets read, and no longer separates the committed candidate from the indifferent one.

The response has been to add verification, which is where the real expense shows up. Employers deploy AI screening against AI applications, candidates escalate in response, and both sides add cost without adding information. Robert Half found in early 2026 that 67% of US HR leaders said reviewing AI-generated applications had slowed their hiring, and 65% of hiring managers said it had become harder to verify candidates' actual skills. Time-to-hire and cost-per-hire have moved the wrong direction. This is the pattern to watch for generally: when a signal collapses, organizations replace cheap inference with expensive verification, and the verification is a permanent new operating cost that appears nowhere in the AI business case.

The internal version is bigger and less visible

Hiring is where this is being discussed, but it is not where most of the damage sits. Inside an organization, effort-as-evidence is load-bearing in ways nobody has written down.

A detailed analysis signaled that someone thought the question mattered. A long, careful email signaled engagement with the problem. A comprehensive proposal signaled conviction, since nobody writes forty pages about something they do not believe in. A quick, thorough response signaled that a colleague prioritized your request. Each of these was a genuine information channel, and each has degraded at the same time, without any announcement.

The consequence is not that leaders now receive worse work. It is that they have lost a channel they were using to allocate attention and trust, and most have not noticed the loss, so they keep inferring from a signal that no longer carries the information. That is worse than knowing the channel is gone, because a broken instrument you still trust is more dangerous than no instrument at all. The manager who is impressed by a thorough proposal is having a real reaction to a document that may represent five minutes of interest.

There is a related effect worth flagging, which is that volume itself becomes uninformative. When producing more costs nothing, more gets produced, and length stops indicating depth. Organizations are now generating documents faster than anyone can read them, and the length that once signaled thoroughness now signals only that someone had access to a tool.

The honest part: some of these signals deserved to die

It would be a mistake to treat this purely as loss, because effort-based signals were always crude proxies and some were straightforwardly unjust.

Polish favored people with time, which meant it favored people without second jobs or caregiving responsibilities. It favored native speakers of the dominant language, and penalized brilliant people who communicated awkwardly in a second tongue. It favored those educated in institutions that taught the conventions, and it rewarded a particular register of professional writing that has approximately nothing to do with competence. A candidate whose application is now clear because AI helped them express themselves is not gaming a signal, they are being freed from a barrier that was measuring the wrong thing.

More fundamentally, effort was never what anyone actually wanted. Nobody hires a person because they suffered over a cover letter. Effort was a proxy for commitment, competence, and judgment, and it was a mediocre proxy. Losing a mediocre proxy is only a problem if you have nothing better, which raises the real question: what were you actually trying to learn, and is there a more direct way to learn it?

That reframing turns this from a loss into a forcing function. The signals that collapsed were the ones we relied on because they were convenient, not because they were good.

Signals that still separate

A signal survives if producing it remains differentially costly, which means the question is what AI has not made cheap. A few things have held.

• Presence in real time: Unscripted conversation, live problem-solving, and responding to an unexpected follow-up cannot be pre-generated. This is why interviews have gained value relative to written materials, and why the most informative interview question is the one that goes somewhere the candidate did not prepare for.

• Specificity that requires actual access: Details that could only be known by someone who was genuinely there, did the work, or understands the particular situation remain expensive to fake, because the cost is in having had the experience rather than in the writing.

• Consequences borne: A track record where someone was accountable for outcomes carries information that no document does, because the cost was paid in the world rather than on a page.

• Stakes attached: A prediction someone will be held to, a commitment with a date, a recommendation the recommender's reputation rides on. AI can produce the words. It cannot assume the risk.

The practical move is to shift weight from artifact-based signals to interaction-based and stake-based ones. In hiring, that means less inference from written materials and more from live work with real problems. Internally, it means less impression from documents and more from conversation, follow-through, and whether someone will attach their name to a specific claim with a specific date.

None of this is free, and it would be dishonest to pretend the replacements cost nothing. Live assessment takes senior people's time. Judging track records requires actually tracking them. The organizations that adapt well are the ones that recognize they are now paying explicitly for information they used to get for free, and budget for it, rather than continuing to read a signal that stopped meaning anything and wondering why their judgment has gotten worse.

This connects to a pattern running through everything AI changes about organizations: the mechanisms companies depended on were often invisible, unnamed, and load-bearing. The middle management layer was doing translation nobody had documented, as covered in what happens to the middle of your org chart. Junior work was building expertise nobody had priced, as covered in the career ladder AI breaks. And effort was carrying information nobody had noticed it was carrying, until it stopped.

FAQs

Q1. What is a costly signal?
It is an action that conveys otherwise unobservable quality because producing it is expensive, and crucially, more expensive for the low-quality type than the high-quality one. Michael Spence formalized this in 1973 with education as the example, showing that a degree could convey real information about ability even if it taught nothing, purely because capable people found it easier to obtain. The cost differential is what makes the signal work.

Q2. Why does AI break these signals specifically?
Because it collapses production cost to roughly the same near-zero for everyone, eliminating the differential the signal depended on. A polished document produced in two minutes by a deeply committed person and by an indifferent one are indistinguishable. The signal still looks the same, which is what makes the failure hard to notice, but it no longer separates anyone from anyone.

Q3. What is the evidence this is happening in hiring?
LinkedIn reported roughly a 45% surge in applications in a year, reaching around 11,000 per minute, while postings fell. The Financial Times reported recruiter estimates that up to half of applicants were using generative AI, with one platform seeing more than double the candidates per role at lower quality. Robert Half found 67% of US HR leaders saying AI-generated applications had slowed hiring and 65% of hiring managers finding skills harder to verify.

Q4. Isn't the internal effect just that we get better documents?
The documents may well be better, but the inference channel is gone. Leaders were using thoroughness and polish to gauge how much someone cared and how seriously they had thought about a problem, allocating attention and trust accordingly. That information channel has degraded while the appearance of it remains, so people keep making the inference from a signal that no longer supports it.

Q5. Is losing these signals entirely bad?
No, and that is worth taking seriously. Effort-based signals systematically favored people with more time, native fluency in the dominant language, and education in the relevant conventions, none of which measure competence. Someone whose application is now clear because AI helped them express themselves is not cheating a signal, they are escaping a barrier that was measuring the wrong thing.

Q6. What is the verification tax?
It is the cost organizations incur replacing cheap inference with active checking once a signal stops working. In hiring it appears as AI screening deployed against AI applications, longer processes, and rising time-to-hire and cost-per-hire. It is a permanent new operating expense that rarely appears in any AI business case, because the signal it replaced was free and therefore never budgeted.

Q7. Which signals still carry information?
Ones where the cost has not collapsed: unscripted real-time interaction, specificity that requires genuine access or experience, track records where someone bore consequences, and claims with stakes attached such as a prediction the person will be held to. AI can produce words but cannot assume risk or manufacture having been present.

Q8. What should a leader change first?
Notice which judgments currently rest on effort as evidence, then decide what you were actually trying to learn and find a more direct route to it. In practice this usually means weighting live interaction and demonstrated track record over produced artifacts, and accepting that this costs senior time you were not previously spending. The alternative is continuing to trust a broken instrument, which is worse than having no instrument at all.