VISTA·RESEARCH™
How it worksThe LibraryResearchHeritageAboutTalk to the Desk
Essay

The Prophet Who Hoped to Be Wrong

By Russ W. Rosenzweig · Vista Research · 14 November 2026 · Why I write with AI

Within two years, Hananiah told the people in the Temple, the king would be home.

It was the fifth month of the fourth year of Zedekiah, king of Judah. Nebuchadnezzar, king of Babylon, had already carried off King Jeconiah, the nobles of Judah and Jerusalem, and vessels from the Temple itself. The prophet Jeremiah stood in the crowd with a wooden yoke bar on his neck. He had been told to wear it. "Make bonds and bars, and put them on your neck." (Jeremiah 27:2, WEB). The bar was his forecast, worn where everyone could see it. Babylon's rule would last, and the way to survive it was to bend.

Hananiah, the son of Azzur, from Gibeon, gave the priests and the people a different one.

“Yahweh of Armies, the God of Israel, says, ‘I have broken the yoke of the king of Babylon. Within two full years I will bring again into this place all the vessels of Yahweh’s house that Nebuchadnezzar king of Babylon took away from this place and carried to Babylon. I will bring again to this place Jeconiah the son of Jehoiakim, king of Judah, with all the captives of Judah, who went to Babylon,’ says Yahweh; ‘for I will break the yoke of the king of Babylon.’ ” (Jeremiah 28:2 to 4, WEB)

Hananiah's was the welcome forecast, the one that would have won a vote. I like to think I would have known whom to believe. In the Temple that day I would have wanted Hananiah to be right. So, it turns out, did Jeremiah.

He said, "Amen! May Yahweh do so." (Jeremiah 28:6, WEB).

Hananiah answered with a piece of theater. He took the bar from Jeremiah's neck and broke it. He repeated his forecast and its deadline, the yoke of Babylon broken "within two full years." (Jeremiah 28:11, WEB). The text says only, "Then the prophet Jeremiah went his way."

I thought of that crowd when I read a headline. I have been in the expert business since 1994, when I founded Round Table Group, arguably the first expert network. Vista Research, which brings experts and expertise to bear on investment decisions, is its younger cousin. Put 570 experts in a headline and I am going to read the article.

The headline was "What 570 Experts Predict the Future of Work Will Look Like," by Nicky Dries, Joost Luyckx and Philip Rogiers, in Harvard Business Review, September 16, 2024. The idea for this essay originated there, in late October 2026.

The study behind the article surveyed five groups. Four were professional groups, from technology experts to policy makers. The fifth was called "Engaged Citizens." I gave the other four rather more attention. After thirty-two years of introducing people to experts, I have developed a few habits. The joke was on me. The engaged citizens put the same two futures at the top of their list as every group of experts did. In one, workers must keep retraining. In the other, people and machines work together.

So, at the top of the list, all five groups agreed. What is that agreement worth before anyone knows the outcome, when a forecast is actually needed? Investors face this question every day. A forecast rarely arrives alone. It arrives as the consensus of the analysts, or the view of the experts, or what everyone in the room already knows. The short answer is that a crowd is worth only as many people as saw the evidence for themselves. That can be far fewer than the number who agree. The long answer took me through a 1950 Weather Bureau rule for grading forecasts and a 1785 theorem about juries. Then it took me back to the Temple, where I had miscounted the prophets.

Predicting the predictors

What I admire in the study is where it starts. Before asking anyone for a prediction, the three scholars went to find out where predictions come from. In 485 articles that ran in the Belgian press from 2015 to 2021, they found seven stories about the future of work told over and over. Then the 570 rated each story on a five-point scale that ran from never to sooner than we think.

The groups did differ, by degree, and that is the article's finding: what people predicted depended a good deal on the group they belonged to. In the article's telling, the economics professors call the bestselling authors doomers and the tech entrepreneurs hypers. The scholars sum up their finding in a line I wish I had written: "we can't predict the future of work, but we can predict your prediction."

They do not call expertise worthless. Their conclusion is that no one can yet tell which group is right, and that the future will be what people make it. Their advice is to look at who is making a prediction and what that person has riding on it. What follows is my attempt at the arithmetic of that advice, and of the agreement that surprised me.

The study set out to map beliefs, not to grade them. Its scale mixed whether a thing would happen with how soon. A mark on that scale names no event and no day on which to look. The 570 also gave a year and a degree of certainty for each prediction. But most of the predictions are too loosely worded for any outcome to settle. What would count, on the appointed day, as enough people learning new skills or working alongside machines?

Hananiah gave us vessels to look for. If the vessels were still in Babylon at the end of those years, nobody would have to ask what he had meant. I enjoy reading market commentary as much as legal complaints, as these are my two professions. I would be grateful for a few more forecasts on those terms.

Jeremiah, before the bar was broken, had supplied the rule for grading him: "As for the prophet who prophesies of peace, when the word of the prophet happens, then the prophet will be known, that Yahweh has truly sent him." (Jeremiah 28:9, WEB).

Hananiah's forecast could fail. A rating on a five-point scale cannot, and so it cannot succeed either. What, then, does grading take?

Fifty points apart

Take two forecasters and one yes-or-no event. The optimist says 80 percent. The skeptic says 30. These numbers are invented, like those in the models that follow. They measure no real forecaster.

Suppose the event happens. The optimist missed by 20 points and the skeptic by 70. Suppose it does not. Now the optimist missed by 80 and the skeptic by 30. Either way their misses are 50 points apart, the same 50 points that separate their forecasts. Listening to the two forecasters argue would have given us that distance for free.

To find out which of them was closer, we need the outcome and a rule. In 1950 Glenn Brier of the U.S. Weather Bureau gave weather forecasters one. Write 80 percent as 0.8. Write the outcome as 1 if the event happened and 0 if it did not. In its usual form today:

Brier score = the average of (forecast − outcome)2

Lower is better, and zero is perfect. Each miss is squared, so a large miss costs far more than a small one. If our event happens, the optimist's miss of 0.2 costs 0.04. The skeptic's miss of 0.7 costs 0.49.

One event proves little, so give them ten, with the same two forecasts each time. If 7 of the 10 happen, the optimist scores 0.22 and the skeptic 0.37. If 3 happen, it is 0.46 against 0.21. The optimist wins the scoring only if more than 55 percent of the events happen. That break-even is the midpoint of their two forecasts.

0.00.10.20.30.40.50.60.70%20%40%60%80%100%Share of the events that actually happenedAverage squared error (lower is better)Their disagreement, 80% minus 30%,is 50 points whatever happensBreak-even at 55%The optimist, who always says 80%The skeptic, who always says 30%
Two forecasters on the same ten yes-or-no events, one always saying 80%, the other 30% (an illustration). How far apart they are, 50 percentage points, never moves: it is the same whatever happens. The two lines are their average squared errors, the Brier score, which depends on what happens. If 7 of the 10 events happen, the optimist scores 0.22 and the skeptic 0.37; if 3 happen, 0.46 and 0.21. The optimist wins only if more than 55% of the events happen, the midpoint of their two forecasts. From model.py.

So a record can tell us who forecast well, once the outcomes are in. Nothing the 570 said has been graded. What they have is their agreement, and it is tempting to count that as evidence.

How many is 570?

Every one of the five groups, engaged citizens included, put the same two futures on top. Two models in the Vista Decision Library say what such agreement is worth, and they give very different answers.

The first is the Condorcet jury theorem. In 1785 Condorcet asked when the majority of a jury can be trusted. Let each judge decide independently, and let each be right more often than wrong. Then the larger the jury, the more often its majority is right. Take judges who are each right 60 percent of the time. A majority of 3 is right 64.8 percent of the time. A majority of 9 is right 73.3 percent of the time. A majority of 569 is right essentially always. I use 569 so that the vote cannot tie. The 570th expert may abstain.

By that arithmetic, if the 570 were such a jury, their agreement would be close to proof.

The second model is the one the Library calls "Herding, cascades, and the wisdom of crowds." It fastens on a word in the theorem that is easy to read past: independently. Its instruction is to count independent sources of evidence, not voices. Judges who each saw the evidence for themselves are a jury. Judges who all heard it from the same place are closer to one observation, repeated.

Which were the 570? I do not know, and the published study cannot say. But recall where the seven stories came from: 485 articles in the press of one country. The authors saw the link themselves: "the four most prevalent narratives in the print media were also most strongly subscribed to by the actors in our survey." If the 570 drew on the same stories, their judgments move together, and statisticians have a formula for what such a crowd is worth. Let n be the number of judges. Let rho be the correlation between any two of their judgments. It is 0 when they are independent and 1 when they are copies. The formula gives the number of fully independent judges whose average would be just as precise as the crowd's.

n_eff = n / (1 + (n − 1) × rho)

At a correlation of 0, the 570 count as 570.

If every pair of the 570 judgments had a correlation of only 0.05, their average would be as precise as the average of about 19 independent judgments.

01002003004005006000.000.050.100.150.200.250.30How much the judgments move together (correlation)Independent judgments they are worthFully independent: 570about 19 at 0.05about 10 at 0.10about 3 at 0.30
How many is 570? If the judgments of 570 experts move together, because they read the same stories or one another, their average is only as precise as the average of far fewer independent judgments: about 19 at a correlation of 0.05, about 10 at 0.1, about 3 at 0.3 (n / (1 + (n - 1) x correlation)). An illustration: the 570's real correlation is not published. From model.py.

At 0.1 the figure is about 10. At 0.3 it is about 3, and the crowd in the headline would fit in a taxi.

This is an illustration. The correlation among the 570 is not published, and no one can estimate it without their individual answers. It may be smaller than 0.05, or larger than 0.3. What the formula shows is how little shared reading it takes to shrink a crowd.

Nor is the formula about Belgium. The same arithmetic applies to any consensus an investor leans on. Twenty analysts working from one model, or a room of experts who have read the same reports, may be fewer opinions than they look. It is an uncomfortable formula for a man in the expert business.

The prophets, plural

I went back to the Temple with that formula and found I had miscounted. I had pictured two men, one against one. But in the chapter before Hananiah speaks, Jeremiah has been warning the priests and the people about a forecast they keep hearing: "Don't listen to the words of your prophets who prophesy to you, saying, 'Behold, the vessels of Yahweh's house will now shortly be brought again from Babylon;' for they prophesy a lie to you." (Jeremiah 27:16, WEB).

Prophets, in the plural. The text does not count them. Hananiah put a date on what the others were already saying. Jeremiah stood nearly alone, and he was the one who hoped he was wrong. Condorcet's arithmetic, applied in the Temple that day, would have sided with the many.

I do not tell this to hang Hananiah's name on anyone, least of all on the 570, who answered a survey in good faith. A consensus can be right. Theirs may be.

Before ChatGPT

The study was conducted, its authors write, in Belgium from 2015 to 2021. Nicky Dries reported early results from its survey of 570 in June 2021. So the 570 answered before ChatGPT was released to the public on November 30, 2022. Since then the future of work has become a question companies and investors are betting capital on.

I can offer one witness, though hardly an independent one. I have just turned 56. I would never have imagined that I would spend most of my day coding in and engaging with the latest AI models. They are now part of the fabric of my work.

I am also in the robotics business, as the founder of Steel Collar Robotics and Aeronautics, an independent dealer of robots and drones. The scholars advise asking what a forecaster has riding on a forecast, so there it is. I stand to gain if my next sentence is believed, so mark it down accordingly. In my view, the future of work will likely arrive through hands as well as words. If the 570 are ever asked again, I would put to them the question you should now be putting to me: which outcome do you hope for?

The stake in sounding bold

I have just declared a stake. Here is the arithmetic of what one can do to a number.

Brier's rule is what statisticians call proper. A forecaster who cares about nothing except her average score does best by reporting exactly what she believes. All the trouble is in the word "nothing."

Suppose she also gains a little each time she sounds bolder: a larger audience, the next invitation, a product to sell. Call that gain her stake.

Best report = belief + stake / (2 × how much the score counts)

Let each extra point on her forecast be worth to her three tenths of a point on her score. A forecaster who believes 40 percent will say 55 percent. Shading her number that far from her belief costs her 0.0225 of score on average. The stake pays her 0.045, so she takes the trade.

The remedy is in the same line. Make the score count double and the shading is cut in half. The same 40 percent comes out as 47.5.

30%40%50%60%70%80%0.00.10.20.30.40.50.6What the forecaster gains from each point of boldness (the stake)The probability reportedWhat the forecaster believes: 40%Says 55%47.5% when the score counts doubleScored at the usual weightScore counted double
A forecaster scored honestly (the Brier rule) but who also gains from sounding bold: a vendor, a promoter, an expert whose next invitation depends on the headline. Believing 40%, the forecaster does best by reporting 55% when each point of boldness is worth 0.3 against the score; counting the score double halves the shading, to 47.5%. An illustration from model.py.

I have seen that stake at close range. In my memoir I described a professor of the dot-com years whom I called Net Prof. I wrote of him: "Unlike other faculty who stuck with tools, techniques, and methodologies, Net Prof got into the business of prophesying, which made him even more mystical and sought after." (The Scholars' Treasure, ch. 7, p. 92).

The sentence describes a reward, not a shaded forecast. Our stake of three tenths is that reward with a number on it.

The stake in the outcome

Now give the forecaster a stake in the event itself: a prize if it happens. She is still graded by Brier's rule. And make her cautious about money in the ordinary way.

I would have guessed she reports higher. In our example a belief of 40 percent is reported as about 33 percent. The prize already gives her a good day if the event happens. So she shades her report down, which makes her score pay more if it does not. A homeowner does the same when she buys fire insurance.

Here are the terms. She starts with one unit of wealth. A perfect score pays her one more unit, the worst score nothing, and a score in between pays in proportion. The prize is one more unit again. Her caution is log utility, the textbook form: a dollar means half as much to her once she has twice as many. Take the prize away and she reports about 42 percent, because caution alone pushes a report toward 50 percent. Other terms would give other figures.

This follows Olivier Armantier and Nicolas Treich (2013). They tested it with dice, where the true odds are known. Subjects who were promised a bonus if an event happened reported significantly lower probabilities for it.

So a stake can push a number either way, and you cannot correct for a stake you have not identified.

The walk-down

Does any of this happen outside models and dice? Earnings are the case investors meet every quarter. Scott Richardson, Siew Hong Teoh and Peter Wysocki studied analysts' earnings forecasts from 1984 to 2001. Their subject was the walk-down, in which forecasts begin optimistic and are walked down to a level the company can beat on the day it reports.

Take an average company and the last forecast before its announcement. When its insiders, its own officers and directors, were net buyers of the stock in the twenty trading days after the announcement, the company had met or beaten that forecast 54 percent of the time. When they were net sellers, it was 66 percent.

The authors read this as managers guiding analysts toward a beatable number when stock is about to be sold. Their evidence is an association and proves no one's intent, and they urge caution in reading it.

When a forecast arrives

So when a forecast arrives, I now ask four things.

Can it be graded? That takes a defined event, a deadline, and a rule set in advance. Hananiah, to his credit, gave the first two.

How many of the voices behind it saw the evidence for themselves?

What does the forecaster gain from the report, and separately from the outcome? A stake does not make a forecast false. It makes the number an unreliable guide to what the forecaster believes.

And which outcome was I hoping for? I try to be hardest on the forecast I want to hear. Fleetwood Mac's "Don't Stop" (1977) is a fine song about tomorrow. I try to keep it off the research desk.

Why we freeze our tests

All of this applies to us. I like hearing that something I wrote is bold. That is a second stake, and the same discount applies. What we can do is the remedy the model gives: make the score count for more.

So every Vista inquiry sets down its tests, with dates, on the day it is published. A public scorecard, at vistaresearch.ai/notes/scorecard.html, records how each one comes out when its date arrives, right or wrong. It grades no probability, though two of our inquiries have put odds on an outcome.

Our new short form, the Hypothesis, will do both. Each Hypothesis will state our probability that its test, if it reaches a verdict, confirms the hypothesis, and Brier's rule will grade that number.

The next question

None of this is a complaint about the study. It is the best map I know of what experts believed about work before ChatGPT arrived. Its authors suggest following such stories over time, and at the very least they offer their study to "future historians" for "retrospective scenario reviews." I am taking them up on it a little early.

How independent were the 570? Part of the evidence already exists, in the individual answers, which only the authors hold.

More of it would take a second survey, the follow-up the authors themselves suggest. They guessed that their survey, run again after ChatGPT, would read differently. A second wave could put the same seven stories to the experts again. It could turn each story into a dated forecast that public data can grade. It could ask each expert which sources shaped the view, and what stake the expert has in it. And it could put my question to every story. Their survey asked it of four extreme scenarios, and Jeremiah answered it without being asked: which outcome do you hope for? Each hope could then be set beside its forecast. Vista would be glad to help field such a panel and to share what it finds.

Amen

I stopped the story in the Temple with the bar broken and Jeremiah walking away. "Then Hananiah the prophet took the bar from off the prophet Jeremiah's neck, and broke it." (Jeremiah 28:10, WEB).

It was a bar of wood, and Jeremiah was sent back to say what breaking it had bought. "You have broken the bars of wood, but you have made in their place bars of iron." (Jeremiah 28:13, WEB).

By the time he finished, there were two deadlines and three things to watch.

The first deadline was Jeremiah's own, and it was the shorter. "This year you will die" (Jeremiah 28:16, WEB). Hananiah had spoken in the fifth month. "So Hananiah the prophet died the same year in the seventh month." (Jeremiah 28:17, WEB). That forecast was settled in about two months. Hananiah did not live to see his own deadline.

The second deadline was Hananiah's. Within two full years, he had said, the vessels would be back. Seven years after he spoke, in Zedekiah's eleventh year, the Temple was burned and the vessels that were left went to Babylon too (Jeremiah 52:5 to 19).

The third thing to watch was the king. Hananiah had promised his return in the same oracle as the vessels, and his story ran the longest. The book's last chapter calls him Jehoiachin. He is Hananiah's Jeconiah, the son of Jehoiakim whom Nebuchadnezzar carried to Babylon (2 Kings 24:6 to 15). "In the thirty-seventh year of the captivity of Jehoiachin king of Judah, in the twelfth month, in the twenty-fifth day of the month, Evilmerodach king of Babylon, in the first year of his reign, lifted up the head of Jehoiachin king of Judah, and released him from prison." (Jeremiah 52:31, WEB). Hananiah had said two years. The book counts to the thirty-seventh year of the captivity, and even then the king was released, not sent home. It leaves him in Babylon, eating the king's bread all the days of his life.

The crowd could not have known any of this when the bar broke. What it had heard earlier was Jeremiah's wish for the king and the captives: "Amen! May Yahweh do so." (Jeremiah 28:6, WEB). I keep coming back to that part. His forecast ran against the other prophets and against his own hopes, and he said so in front of the priests and all the people. That is as close as a person can come to showing that the number and the wish have been kept apart.

When a forecaster tells me he hopes he is wrong, and gives me the forecast and the date anyway, I listen.

I can put a date on a forecast. The Amen takes more practice.

Sources

  • Nicky Dries, Joost Luyckx and Philip Rogiers, "What 570 Experts Predict the Future of Work Will Look Like," Harvard Business Review, September 16, 2024, https://hbr.org/2024/09/what-570-experts-predict-the-future-of-work-will-look-like.
  • Nicky Dries, Joost Luyckx and Philip Rogiers, "Imagining the (Distant) Future of Work," Academy of Management Discoveries 10, no. 3 (2024): 319 to 350, DOI 10.5465/amd.2022.0130. The study's details here, including the five groups' ratings in its Table 4, are from the authors' manuscript (revision R2) in the KU Leuven repository, https://lirias.kuleuven.be/retrieve/718388; its wording may differ from the printed article.
  • Nicky Dries, "So you think you can predict the future of work?" BI Business Review, June 15, 2021, https://www.bi.no/en/research/business-review/articles/2021/06/so-you-think-you-can-predict-the-future-of-work/. An early report of the same survey of 570, framed as four broad scenarios.
  • OpenAI, "Introducing ChatGPT," November 30, 2022, https://openai.com/index/chatgpt/.
  • Glenn W. Brier, "Verification of Forecasts Expressed in Terms of Probability," Monthly Weather Review 78, no. 1 (1950): 1 to 3, DOI 10.1175/1520-0493(1950)078<0001:vofeit>2.0.co;2, and the Vista Decision Library entry "Was the seventy percent right seventy percent of the time?"
  • The Vista Decision Library, the entry "The Condorcet jury theorem" and the structure it runs under, "Herding, cascades, and the wisdom of crowds," on the Library's page How the Models Work. The formula in "How many is 570?" is the standard one for the number of independent judgments whose average is as precise as the average of n equally correlated ones.
  • Olivier Armantier and Nicolas Treich, "Eliciting beliefs: Proper scoring rules, incentives, stakes and hedging," European Economic Review 62 (2013): 17 to 40, DOI 10.1016/j.euroecorev.2013.03.008. The theory and the authors' account of their experiment are from the working paper, Toulouse School of Economics Working Paper 10-156 (April 2010); the journal version may differ.
  • Scott Richardson, Siew Hong Teoh and Peter D. Wysocki, "The Walk-down to Beatable Analyst Forecasts: The Role of Equity Issuance and Insider Trading Incentives," Contemporary Accounting Research 21, no. 4 (2004): 885 to 924, DOI 10.1506/khnw-pjyl-adub-0rp6.
  • World English Bible (public domain), 2 Kings 24 and Jeremiah 27, 28 and 52, from ebible.org/eng-web.
  • Russ W. Rosenzweig, The Scholars' Treasure, ch. 7, pp. 91 to 92.
  • The model figures in "Fifty points apart," "How many is 570?," "The stake in sounding bold" and "The stake in the outcome" are illustrations computed in this essay's model file; none is a measurement of any real forecaster or of the 570. The analysts' figures in "The walk-down" are Richardson, Teoh and Wysocki's.

Disclosures

This essay is analysis and commentary, not investment advice, and nothing in it is a recommendation about any security. Its numerical examples are illustrations. This essay names no publicly traded company. Russ W. Rosenzweig, Vista's founder, owns shares of publicly traded companies, none of them named here, and holds no options on any company named here. Holdings through mutual funds and exchange-traded funds are not counted. Russ W. Rosenzweig is the founder of Steel Collar Robotics and Aeronautics, an independent dealer of robots and drones and a robotics research think tank. How this essay was made: written with the help of AI tools. The models were built and every model figure computed in code, each public source was checked against the copy named in Sources, and the draft was fact-checked before I read it. I read it and listened to it before it was published. Why I write with AI.