Sour Milk

I first heard it in a room over a year ago. A partner was pitching, and she described the pressure on the client the way everyone used to describe it, as FOMO. Someone else moving faster, buying earlier, getting there first.

My boss corrected her without looking up.

“FOBO now.”

Fear of becoming obsolete. The conversation carried on and nobody remarked on it. I have heard it constantly since, and I have come to think the substitution is the most honest thing said in most of those rooms.

FOMO measured you against an evening you missed. It assumed you would still be there afterwards, merely less well informed, and it always came with a purchase attached: buy the platform, learn the capability, attend the thing, catch up. FOBO has the same shape and a different object. It does not ask what you are missing. It asks whether the thing being left behind might be you.

The word is not slang, either. Gallup put it into a poll release in September 2023 as a play on FOMO, reporting that twenty-two per cent of American workers feared technology would make their job obsolete, up from fifteen two years before. The headline number is not the story. Among workers without a degree the figure was flat at twenty-four per cent. Among graduates it went from eight to twenty. The people who had newly become afraid were the ones whose credentials were supposed to be the protection.

Milk has a date on the bottle, and even then the date settles almost nothing. We open it, we smell it, and we decide. The carton is usually still in date when it goes down the sink, and the thing that finished it was not the milk turning. It was somebody saying it had.

Nothing in this business has a date on it at all. No platform stops working on the morning its date comes up. No business model expires on a Thursday, and nobody sends a certificate confirming that your expertise remains current for another twelve months. Obsolescence is not a property things acquire. It is a verdict somebody passes, and it can be passed on something that still works perfectly.

We are not the first to invent a verdict and then forget that we invented it. Everybody knows what a thumb turned down in an arena means. In 1872 Jean-Léon Gérôme exhibited Pollice Verso, a canvas of a victorious gladiator standing over a fallen one while the crowd, Vestal Virgins among them, aim their thumbs into the sand. The painting travelled, it was reproduced everywhere, and it eventually reached cinema, and much of what we now picture a Roman crowd doing with its hands came to us through it.

The Roman writers who mention the gesture never describe it. The most careful modern reading belongs to the classicist Anthony Corbeill, who worked through the Roman vocabulary of the hand and concluded that Gérôme very probably had it backwards. An extended thumb was hostile, a jabbing thing, standing in for the blade going in. Mercy was the thumb folded away inside a closed fist, the sword sheathed. If he is right, the sign we all make to say you are finished is closer to the sign for let him live.

Nobody checks. The gesture is instantly legible and possibly inverted, and it has never needed to be accurate in order to work. It only needed everyone to agree that they recognised it. That is the condition our own verdicts are delivered in, and it is why they sound like findings.

Standing underneath all of it, it is tempting to look for the emperor. The board. The market. The client. AI itself. Some authority above the arena deciding what still has value and what has reached the end of its useful life.

Look properly and mostly there is nobody there. Nobody in your organisation has been handed the thumb. The gesture is horizontal. It is made across tables, on calls, over coffee, almost always about a platform or a team or a person who is not in the room. Everyone making it is standing in the same sand.

Mostly is doing some work in that sentence. Plenty of organisations have now made fluency with these tools a condition of getting on, and in some of them a condition of staying, which is an emperor by any reasonable definition. It rarely announces itself as a verdict. It arrives as a utilisation figure, as staffing that quietly stops coming your way, as a client asking for somebody closer to the work, as a line in a review about keeping pace. But an emperor does not replace the horizontal gesture. He is why it multiplies. Once a verdict can come down from above, showing you are on the current side of the line stops being vanity and becomes maintenance, and the cheapest way to show it is to point at somebody who is not.

Behind is a relative position. You cannot place anybody there without quietly locating yourself somewhere ahead of them. Calling a company a Blockbuster sounds like market analysis right up until you notice that the speaker has assigned himself Netflix.

Nobody making that comparison experiences it as self-defence. It arrives as observation, which is the only form in which it would work, and it is the only available answer to a question nobody else will answer for you.

None of which makes the underlying change imaginary. Some platforms really are past it, some skills really are depreciating, and the people saying so are frequently right. The classification does not run instead of the economics; it sits on top of them. But it is not the same event, and only one of the two is carried out hourly, in rooms, by people with something of their own riding on the answer. How you tell one from the other in the moment, I cannot say. Neither can anybody else, and that is not a gap in the argument. It is the condition the argument is about.

Am I still useful.

There used to be crude proxies. Experience. Tenure. Expertise that accumulated slowly enough to still be worth something by the time you had it. The perceived half-life of all that has shortened to the point where a platform can work perfectly and be legacy, and a person can be excellent at a job while suspecting that excellence at that job is what is losing value. The employers surveyed for the World Economic Forum's 2025 Future of Jobs report expect nearly two-fifths of current skill sets to be outdated by 2030, a precise-sounding number attached to nobody in particular.

Nobody will tell you either way. You find out, if you find out at all, because somebody used the past tense about you in a room you were not in. Most of the time there is no verdict at all, only its absence, going on indefinitely, and an absence is harder to stand in than a sentence would be.

So we look sideways. At the colleague who still does it manually. At the leader who has not understood the tool. At the platform everyone used two years ago. At the company that missed the turn. We find the carton that has gone off, and for as long as it takes to say so, the one in our own hand smells fine.

I have made the comparison myself, in a room, about someone.

There is a second defence, and it looks more respectable than the first. If nobody will certify that you are current, you can demonstrate it. Use the tools visibly, bring the new vocabulary into the meeting, be the person who has already tried the thing everyone else is circling. It is the sensible response, it is what most organisations are now explicitly asking for, and it is where the trap closes.

Because demonstrating it costs you as well. Researchers at Duke ran four preregistered experiments with more than four thousand participants and found a social penalty for exactly this. People who use AI at work expect to be judged lazier and less competent, and they turn out to be right. Colleagues rate the same work lower when told a machine helped produce it, and the effect follows candidates into hiring decisions. Two verdicts pointing in opposite directions, both passed sideways by people at the next desk.

There is one detail in that study that I have not been able to put down. The penalty largely disappears when the person doing the evaluating uses AI themselves. Which means the judgement was never really a reading of the work. It was a readout of where the judge was standing when they made it. The same act is lazy or it is capable depending entirely on the position of the person watching, and nobody involved experiences their own reaction as anything other than an assessment of quality.

So the safest posture is a performance in two directions at once: fluent enough to be current, careful never to look dependent. Talk about the tools, do not appear to need them. A good many people I know in this industry are performing some version of it, myself included.

What it costs is not obvious, because the cost does not fall on the performers. It falls on the organisation's ability to know anything true about itself. Where saying I have not used that yet is heard as an admission of expiry, it tends not to get said. Where saying I used it heavily for this is heard as an admission of dependence, that tends not to get said either. Both of the sentences a company would need in order to find out what it can actually do have been made expensive. Adoption gets reported. Competence is inferred from the reporting. Nobody is lying, and none of it is reliable.

There is a simple way to see the gap. Take the training completion figure, whatever it is, and set it beside the number of people in the same organisation who could tell you what one of these tools got wrong last week. The first number is easy to produce and easy to report. The second requires somebody to have looked closely enough to notice, and then to say so, and saying so is one of the two sentences that has been made expensive. Almost nobody collects the second number. It is the only one of the pair that would have told you anything.

And the exposure is not spread evenly. Expertise that took fifteen years to build is among the hardest to demonstrate in a forty-minute meeting, because it shows up as judgement, as knowing which of four plausible approaches will fail in month three, and judgement can look like hesitation to a room that reads hesitation as lag. The people best protected are often the ones whose competence is most performable. That is not the same population as the ones who are good at the work.

If an organisation wanted to know what was actually happening inside it, it would have to make different things safe to say. None of what follows is a training programme.

Stop reading adoption as capability. Licence counts, completion rates and self-reported fluency measure willingness to be seen using the tool, and in an environment that rewards being seen, those two numbers separate on contact. Ask instead what somebody leaned on it for, and treat a detailed answer as a good sign rather than a confession.

Make the first sentence free, and prove it once in public. Not a policy line about psychological safety. A senior person saying I have not used that yet, in a room, with nothing happening to them afterwards. Until that has visibly occurred, nobody believes the policy, and they are right not to.

Make sure the people assessing AI-supported work have used the tools themselves. The Duke finding suggests why: the penalty largely disappears once the evaluator has crossed the bridge too. That does not make an evaluation clean, and it will not fix a performance process on its own. But an assessment of AI capability run entirely by people who have not tried it is measuring something other than the work.

And watch who goes quiet. The exposure falls on the people whose contribution is judgement, and judgement is slow to display. If your most experienced operators have stopped speaking first in meetings about this, that is data.

Gérôme gave the world an emperor because an emperor makes a verdict comprehensible. Authority has a seat. The condemned man knows where the decision came from, even if the gesture was wrong. The rooms I sit in offer nothing so clean. Everyone is in the arena and everyone is in the stands, looking up for a thumb that never comes while extending their own across the table.

On Monday somebody will mention a platform that is already legacy, a competitor who did not see it coming, a colleague whose skills belong to the old model. I will know exactly what they mean. I may well agree. And knowing all of this, I expect I will join in, because sitting still under a judgement nobody is ever going to pass is harder than passing one along.

Sources

Gallup (2023). Saad, L., “More U.S. Workers Fear Technology Making Their Jobs Obsolete.” Work and Education poll, fielded 1–23 August 2023, published 11 September 2023. 22% of US workers worried technology would make their job obsolete, up seven points from 15% in 2021; the rise came almost entirely from college-educated workers (8% to 20%), while concern among non-graduates was unchanged at 24%. Gallup uses the term FOBO in the release as a play on FOMO.

Reif, J. A., Larrick, R. P., & Soll, J. B. (2025). “Evidence of a social evaluation penalty for using AI.” Proceedings of the National Academy of Sciences, 122(19), e2426766122 (Duke University, Fuqua School of Business). Four preregistered experiments, N = 4,439. Participants anticipated and received negative evaluations of competence and motivation for AI use; the effect carried into assessments of job candidates, and was substantially reduced when the evaluator used AI themselves.

Corbeill, A. (1997). “Thumbs in Ancient Rome: Pollex as Index.” Memoirs of the American Academy in Rome, 42, 61–81. Expanded in Corbeill, A. (2004), Nature Embodied: Gesture in Ancient Rome, Princeton University Press. Argues that the extended or hostile thumb (infesto pollice) signalled death and the thumb enclosed within a closed fist signalled mercy, reversing the popular understanding.

Gérôme, J.-L. (1872). Pollice Verso. Oil on canvas, Phoenix Art Museum. The accuracy of the thumbs-down gesture was disputed almost immediately; an 1879 pamphlet collected evidence on both sides, including a letter from Gérôme dated 8 December 1878.

Juvenal, Satires (c. 110–127 AD) and Prudentius, Contra Symmachum (late 4th century AD). Both refer to a turned thumb in the context of the arena; neither specifies the direction or form of the gesture.

World Economic Forum (2025). Future of Jobs Report 2025. Employers expect 39% of workers' existing skill sets to be transformed or become outdated by 2030.

Note on sources not used: several widely circulated figures on “skill obsolescence anxiety” originate in vendor-commissioned surveys rather than peer-reviewed research or primary polling, and have been left out rather than hedged.