The Pile Was Protecting You

There is a file on your company. It has been growing for nearly twenty years, and it is full of the things people say about work when nobody from the company is in the room. The sharpest entries tend to come from people on their way out the door.

Recruiting has always run on one quiet asymmetry. The company has been inside. You have seen only what it chose to show you. Everyone knows the career page is fiction and everyone reads it anyway, because the alternative was a pile: thousands of anonymous reviews, contradictory and badly spelled and endless. A diligent candidate would read six, notice that the angriest person always writes the longest one, and apply the usual discount for disgruntled ex-employees.

I have come to think that discount was doing an enormous amount of hidden work. The pile's accidental defense was effort. It protected employers from scrutiny, and it protected candidates from a distorted verdict. Nobody reads a thousand reviews, so a thousand reviews were, in practical terms, private.

Now a summary costs nothing to read. New Indeed data, reported by Fortune this week, says some 70% of job seekers use generative AI tools like Claude, ChatGPT and Perplexity to research employers, and that when they ask about company culture, roughly 40% of the sources cited come from Glassdoor and Indeed rather than company career pages. (The figures come from Indeed, which has an obvious stake in the answer. A weather vane, not a survey.)

Taylor Meadows, a principal advisor at Glassdoor, is admirably calm about it. "Your employer brand exists whether you are participating in it or not," he says, and he is right that AI "isn't uncovering anything that isn't already online."

That second observation is exactly where the interesting part starts. Nothing new has been revealed. What changed is the cost of reading.

A summary does not add information. It removes the effort that kept information inert.

A pile of documents is a promise that answers exist somewhere in there. A paragraph is the answer.

There is a second change hiding inside the price, and it may matter more. An archive waits to be consulted. The pile was something you might look at on a Sunday afternoon if you were unusually conscientious. Suppose the summary turns up at eleven at night instead, the evening before you have to tell the recruiter yes or no. That is the one hour of the whole process when information can actually change a life. A verdict delivered at the moment of decision is not really an archive at all.

Which brings me to the half of this I like. Meadows warns about offboarding, because departing employees leave reviews that models later surface, so their final impression of the company counts twice. I would go further. The way a company ends employment is now a recruiting document. The last day has become a first impression.

Qantas is a live example. This week the airline told staff that 850 Australian corporate roles are going to Accenture, most of them moving to teams in India, with the changes beginning in mid-2027 and affected staff to be redeployed "where possible". The BBC's account framed the deal as the airline fast-tracking AI investment. Chief executive Vanessa Hudson's internal email is a good one by the standards of the genre. She wrote that "this is hard news for some and I'm sorry for the uncertainty it brings". She said the "scale and speed of change in technology and AI is something none of us has seen in our lifetime". She described more than 1,000 legacy applications that were never designed to work together, and she put a date on the thing.

And that email is corpus now. Whatever a model tells a nervous graduate about Qantas in 2030, some of it will be built out of that message, the reporting around it, and whatever the people affected write in the years between. Here the incentive gets interesting: the honesty of a layoff email is now a decade-long recruiting asset. Polished corporate mush will be summarised as polished corporate mush. A real apology with a real date on it comes through the summary intact.

Careful, though, because the verdict has a defect, and it is not the one people are complaining about. A summary can be faithful to its documents and still be built on a rigged sample. Compression is mostly honest. Sampling is where the trouble lives, and this sample was never fair.

Almost nobody writes a workplace review after a merely fine Tuesday. The furious write, the devoted write, and the whole middle is missing. What kept the pile roughly honest was that everyone could see what it was: "that's just a disgruntled ex-employee" has been a cliché for twenty years because it is often true and always available.

A synthesis keeps the skew and drops the discount.

It takes a biased sample and hands it over in a calm, reasonable register, and calm registers read like findings. The funhouse mirror now has a court record's posture. The confidence comes free with the summary; the evidence underneath never earned it.

Meadows' prescription is to "feed the beast": ask the models "What is it like to work at X company?", "reverse engineer" whatever surfaces, publish Q&As including "the good, the bad, and the ugly", and encourage employees to share authentic experiences. It is sensible advice. It is also a description of landscaping. Employers are about to start gardening the very corpus the model reads, so the verdict of 2028 will be part genuine grievance and part cultivated humility, delivered in one even tone, with no seam showing where one stops and the other starts. "None of this is about perfection and none of this is about control," Meadows says. "It's just about informing and influencing." Lovely sentence. Also: gardening.

So here is what I would do with a verdict like that. Treat it as a lead, not a judgment. Make it show its sources, then go read the people who left, because that is where the sample runs hottest and where you have to put the discount back in by hand. And if you are the one hiring: the exit interview is public property now. The email you send on the worst day of somebody's year is a document your future hires will meet in summary for the next decade.

The file was never sealed. It was just heavy.