I left grad school (physics & phil at U Chicago) back in 1984, to move to Silicon Valley to work on the then AI boom (which I did at Lockheed and NASA), and to help Xanadu in its quest for hypertext publishing.
I love the motivation behind this line of thinking, but I wonder whether such a system can ever become practical because most "news" is storytelling rather than reporting. (Another theme I remember you discussing recently.) Although the details of who, what, when, where, and how may be verified — and laziness or sloppiness with these can definitely give us hints as to the credibility of the storyteller — the truthfulness of a story as narrative is not itself falsifiable. A story is just an invitation to interpret the facts in a certain way.
An alternative approach might be to utilize a narrow machine learning model trained only to recognize positive or negative sentiment to detect https://en.wikipedia.org/wiki/Emotive_conjugation and report on the +/- valence of particular authors on particular issues. One can imagine this becoming the input to a more sophisticated source of news that blends the perspectives from different sources.
I have some free Claude Code credits to use up so I may just take a hack at this!
I never got around to thanking you properly for your kindness and generosity in meeting with members of my Facebook group to discuss Prediction Markets. Thank you. Manus recently made available One Trillion Tokens to people who create their own Apps. I wish we could all Vibe-code our own Prediction Markets. Someday, creation of Prediction Markets might be democratized. This might or might not address the burgeoning need to determine what is "Live and what is Memorex" if you remember the old Ella Fitzgerald commercial. We will all need our Personal Truth Engines...and a personal Prediction Market populated by those whom we know and trust might be accessible via an Icon in the messages we send. Or not. But, why shouldn't this be one more service ChatGPT or Meta offers?
I would be very interested in seeing what you build, if you do. My interest is somewhat tangential to what you are suggesting, but related to issues surrounding methodologies and technologies to arrive at "True Probabilities." Early approaches included simple Averaging of Expert's Opinions. Some of the stepchildren of the Good Judgement Project developed formulae for partitioning Error Variance, and one source can be attributed to personal bias. They actually developed an app using "R" to control for such bias which they published, if memory serves me correctly.
"I'm shocked—shocked—to find that [inaccurate /false reporting] is going on [ ] here!". (Captain Renault, Casablanca-1942)
I’m not a cynic. News media is incapable of reporting information without taking sides. This shouldn’t be a surprise. Individuals are ideological—even reporters and journalists!
Not only news (including social media) but any source (Yes even AI) are corrupted by their human masters.
I’m a betting man (thanks to Caplan and you). So whenever someone wishes to take an absurd position, I’m always willing to put my money where my mouth is—But shockingly few reciprocate.
“It is so easy to be wrong-and to persist in being wrong-when the costs of being wrong are paid by others." (Attributed to Thomas Sowell)
News inaccuracy isn't just a matter of ideology or "sides". It's easier to be wrong than right, and the incentives aren't strong enough to overcome that.
wow, this quote of Sowell made my day. Thanks! “It is so easy to be wrong-and to persist in being wrong-when the costs of being wrong are paid by others." (Attributed to Thomas Sowell)
It seems to me the practical challenge with all such betting markets is the judging. In cases where the judging is trivial – as in the NYSE – we already have efficient markets. This suggests that the difficulty of judging is an impediment to adoption.
Judging is a heavyweight operation, in general, and there will always be more claims than there is capacity to fairly judge them. Which suggests two strategies: (1) narrowing the judging scope to one or two overall claims made by a news article, or (2) narrowing the scope to claims that can be more easily judged (numerical statements, etc.).
I like the idea and would love to see it happen. Although I observe that for most people "objective truth" is not what they seek from news, so it would be a niche product.
For most all the markets that exist on Polymarket and Kashi today, but didn't five years ago, it wasn't gains in judging that have made them possible now.
These markets are expanding into areas that are easy to judge: Election outcomes, sports betting, how many launches SpaceX will have this year, what will be the top AI model at the end of the month, etc. Looking at Polymarket and Kalshi I don't see anything that is hard to judge, which is certainly by design.
This is qualitatively different from the kinds of "propositions" in a typical news article. To apply these techniques to news a central problem then becomes how to make judging scalable and (at least somewhat) objective. The two mechanisms that have been tried – having experts judge and letting the masses vote – are vulnerable to claims of bias.
But fact checkers we know well in fact are mostly ideologizing themselves.
The question is not *whether* judges could judge them, but how easily and how well and how authoritatively in the eyes of the consumer - and then the law, as the inevitable lawsuits that would follow from judgements that are not in fact obvious.
The problem of oracle is certainly non trivial, but if you are interested, it is not a sort of hard problem as you are proposing it to be.
Abstract incentive based systems like UMA optimistic Oracle can resolve many different types of markets without needing unbiased pundits in each domain. You should expect to also see the rise of mutually agreed upon AI oracle systems.
Practically because I know people who are working on them. If your point is more theoretical of the sort aren't they biased, there are a number of replies the first off the top off my head being:
1. Yes, but potentially not catastrophically so, especially if both people on each side of the trade agree to using a specific model/paradigm (i.e. prompt with access to web search whatever)
2. The power of incentives trumps all. Could implement a system like UMA optimistic oracle where being right is rewarded and wrong means you lose money, so human Creators of these oracles are incentivised to make them less biased/mitigate.
Etc.
As for your point who is doing the agreeing, I mean the parties buying yes and no. Unlike many other systems you may associate with prediction markets like gambling, its actually closer to a market with no house. Many systems like polymarket that operate using a central limit order book are just matching two people like you or me who have different takes or rather different credence about a take.
Dude, you fail to address who is doing the buying and selling (who is on which side of what trade) in this case of news accuracy. Which was the topic.
I’m a big believer in the power of incentives. You haven’t described how incentives are aligned or who is going to buy what - and why they would - just because you inserted the idea of an AI oracle.
Offhand, the concept seems overly complicated and convoluted. Kind of the proverbial Rube Goldberg machine, not that either is without merit -- I've periodically had recourse to the latter at least. Provides some "explanatory depth" as the philosophical jargon goes:
"Rerum cognoscere causas; Mechanisms in Science: things learned at my mother's knee and other low joints"
But your idea seems something of a case of who watches the watchers. Best bet seems to be the readers and their subscriptions -- and a question of who pays for them.
We tend to be a gullible bunch. I remember reading Douglas Hofstadter's "Gödel, Escher, & Bach" -- more than a few decades ago -- and I remember he had a "thought experiment" of sorts on the topic there:
DH: “How gullible are you? Is your gullibility located in some "gullibility center" in your brain? Could a neurosurgeon reach in and perform some delicate operation to lower your gullibility, otherwise leaving you alone? If you believe this, you are pretty gullible, and should perhaps consider such an operation.”
Though I'm somewhat chagrined to remember that, on reading the first couple of those sentences, I had thought that possible. And then somewhat "nonplussed" on reading the last one. 🙂
“A big problem is that news customers don’t seem very interested in news accuracy. For example, though it has been feasible for centuries for news sources to post bonds payable to those who can prove their stories false, few have ever done so.”
Your second sentence does NOT demonstrate that news customers are not interested in news accuracy.
It shows only that news incumbents and entrepreneurs/investors have to date not believed they could profit from the posting of such bonds.
Those incumbents and never-were entrepreneurs may or may not have been wrong. But suggesting that their absence demonstrates that customers are not interested in news accuracy is like suggesting in 1998 that customers were not interested in smartphones with internet access and touchscreen UIs because they hadn’t purchased any.
I would really like this to happen. I think that one thing we will need to be able to measure is "how well-substantiated" is this argument. Well-substantiated but still wrong is a category, of course, but when deciding whether to even bother reading an article it helps to know the track record of the journalists involved. The ones who are mostly writing historical fiction with the historical-period set to "right now" can be skipped, but only if we can judge this ahead of time.
One suggestion for your proposed user interface: if we lean into "true is blue" (instead of green), we will have a result that is usable by the 8% or so of the population which has red or green colour deficiency.
The MSM has willingly given away most of theirs as they both seek to promote the ideological agenda of most of their members, and switch to a business model of selling to the most activist, extreme leftists. Which reduces their reputation still further amongst those not on the hard left.
The most common way to give a false impression is by selectively picking facts that aren't wrong and ignoring other relevant facts that also aren;t wrong.
Note also that it is always necessary to select facts. So false stories and accurate stories often both look fine on the surface.
One example: "You use some of the same words that terrorists use". That is undeniably correct - terrorists use words like "the", "and", "sometimes", etc. - but it can convey an impression of you being a terrorist. String a buunch of statements along those lines together, and you get a completely stupid story made out of completeley undeniably true factual statements. it's how the media makes stories to attract attention, and social media selects such things for promotion.
I don't see how any market mechanism or fact checking helps correct those types of spreading of false understanding.
I love the motivation behind this line of thinking, but I wonder whether such a system can ever become practical because most "news" is storytelling rather than reporting. (Another theme I remember you discussing recently.) Although the details of who, what, when, where, and how may be verified — and laziness or sloppiness with these can definitely give us hints as to the credibility of the storyteller — the truthfulness of a story as narrative is not itself falsifiable. A story is just an invitation to interpret the facts in a certain way.
An alternative approach might be to utilize a narrow machine learning model trained only to recognize positive or negative sentiment to detect https://en.wikipedia.org/wiki/Emotive_conjugation and report on the +/- valence of particular authors on particular issues. One can imagine this becoming the input to a more sophisticated source of news that blends the perspectives from different sources.
I have some free Claude Code credits to use up so I may just take a hack at this!
The challenge is to get people to trust the output of your specially trained model.
I never got around to thanking you properly for your kindness and generosity in meeting with members of my Facebook group to discuss Prediction Markets. Thank you. Manus recently made available One Trillion Tokens to people who create their own Apps. I wish we could all Vibe-code our own Prediction Markets. Someday, creation of Prediction Markets might be democratized. This might or might not address the burgeoning need to determine what is "Live and what is Memorex" if you remember the old Ella Fitzgerald commercial. We will all need our Personal Truth Engines...and a personal Prediction Market populated by those whom we know and trust might be accessible via an Icon in the messages we send. Or not. But, why shouldn't this be one more service ChatGPT or Meta offers?
Today is Sunday, so it will be open source. ;-)
I would be very interested in seeing what you build, if you do. My interest is somewhat tangential to what you are suggesting, but related to issues surrounding methodologies and technologies to arrive at "True Probabilities." Early approaches included simple Averaging of Expert's Opinions. Some of the stepchildren of the Good Judgement Project developed formulae for partitioning Error Variance, and one source can be attributed to personal bias. They actually developed an app using "R" to control for such bias which they published, if memory serves me correctly.
Rough start https://github.com/riemannzeta/rationalizer
It's relatively easy to get up and running on a mac following the instructions in the readme
"I'm shocked—shocked—to find that [inaccurate /false reporting] is going on [ ] here!". (Captain Renault, Casablanca-1942)
I’m not a cynic. News media is incapable of reporting information without taking sides. This shouldn’t be a surprise. Individuals are ideological—even reporters and journalists!
Not only news (including social media) but any source (Yes even AI) are corrupted by their human masters.
I’m a betting man (thanks to Caplan and you). So whenever someone wishes to take an absurd position, I’m always willing to put my money where my mouth is—But shockingly few reciprocate.
“It is so easy to be wrong-and to persist in being wrong-when the costs of being wrong are paid by others." (Attributed to Thomas Sowell)
News inaccuracy isn't just a matter of ideology or "sides". It's easier to be wrong than right, and the incentives aren't strong enough to overcome that.
wow, this quote of Sowell made my day. Thanks! “It is so easy to be wrong-and to persist in being wrong-when the costs of being wrong are paid by others." (Attributed to Thomas Sowell)
It seems to me the practical challenge with all such betting markets is the judging. In cases where the judging is trivial – as in the NYSE – we already have efficient markets. This suggests that the difficulty of judging is an impediment to adoption.
Judging is a heavyweight operation, in general, and there will always be more claims than there is capacity to fairly judge them. Which suggests two strategies: (1) narrowing the judging scope to one or two overall claims made by a news article, or (2) narrowing the scope to claims that can be more easily judged (numerical statements, etc.).
I like the idea and would love to see it happen. Although I observe that for most people "objective truth" is not what they seek from news, so it would be a niche product.
For most all the markets that exist on Polymarket and Kashi today, but didn't five years ago, it wasn't gains in judging that have made them possible now.
These markets are expanding into areas that are easy to judge: Election outcomes, sports betting, how many launches SpaceX will have this year, what will be the top AI model at the end of the month, etc. Looking at Polymarket and Kalshi I don't see anything that is hard to judge, which is certainly by design.
This is qualitatively different from the kinds of "propositions" in a typical news article. To apply these techniques to news a central problem then becomes how to make judging scalable and (at least somewhat) objective. The two mechanisms that have been tried – having experts judge and letting the masses vote – are vulnerable to claims of bias.
If fact checkers can check article claims, judges could judge them.
But fact checkers we know well in fact are mostly ideologizing themselves.
The question is not *whether* judges could judge them, but how easily and how well and how authoritatively in the eyes of the consumer - and then the law, as the inevitable lawsuits that would follow from judgements that are not in fact obvious.
The problem of oracle is certainly non trivial, but if you are interested, it is not a sort of hard problem as you are proposing it to be.
Abstract incentive based systems like UMA optimistic Oracle can resolve many different types of markets without needing unbiased pundits in each domain. You should expect to also see the rise of mutually agreed upon AI oracle systems.
Why should we expect these “mutually agreed upon” AI oracle systems?
The systems are biased by the text they are trained on, their fine tuning and their so-called safety guidelines.
But the bigger point is, who is it that would be doing the “mutual” agreeing, and how might that even work?
Practically because I know people who are working on them. If your point is more theoretical of the sort aren't they biased, there are a number of replies the first off the top off my head being:
1. Yes, but potentially not catastrophically so, especially if both people on each side of the trade agree to using a specific model/paradigm (i.e. prompt with access to web search whatever)
2. The power of incentives trumps all. Could implement a system like UMA optimistic oracle where being right is rewarded and wrong means you lose money, so human Creators of these oracles are incentivised to make them less biased/mitigate.
Etc.
As for your point who is doing the agreeing, I mean the parties buying yes and no. Unlike many other systems you may associate with prediction markets like gambling, its actually closer to a market with no house. Many systems like polymarket that operate using a central limit order book are just matching two people like you or me who have different takes or rather different credence about a take.
Dude, you fail to address who is doing the buying and selling (who is on which side of what trade) in this case of news accuracy. Which was the topic.
I’m a big believer in the power of incentives. You haven’t described how incentives are aligned or who is going to buy what - and why they would - just because you inserted the idea of an AI oracle.
Offhand, the concept seems overly complicated and convoluted. Kind of the proverbial Rube Goldberg machine, not that either is without merit -- I've periodically had recourse to the latter at least. Provides some "explanatory depth" as the philosophical jargon goes:
"Rerum cognoscere causas; Mechanisms in Science: things learned at my mother's knee and other low joints"
https://humanuseofhumanbeings.substack.com/p/rerum-cognoscere-causas
But your idea seems something of a case of who watches the watchers. Best bet seems to be the readers and their subscriptions -- and a question of who pays for them.
The problem is that readers don't care much about accuracy https://www.overcomingbias.com/p/the-real-problemhtml
We tend to be a gullible bunch. I remember reading Douglas Hofstadter's "Gödel, Escher, & Bach" -- more than a few decades ago -- and I remember he had a "thought experiment" of sorts on the topic there:
DH: “How gullible are you? Is your gullibility located in some "gullibility center" in your brain? Could a neurosurgeon reach in and perform some delicate operation to lower your gullibility, otherwise leaving you alone? If you believe this, you are pretty gullible, and should perhaps consider such an operation.”
https://www.goodreads.com/quotes/52221-how-gullible-are-you-is-your-gullibility-located-in-some
Though I'm somewhat chagrined to remember that, on reading the first couple of those sentences, I had thought that possible. And then somewhat "nonplussed" on reading the last one. 🙂
“A big problem is that news customers don’t seem very interested in news accuracy. For example, though it has been feasible for centuries for news sources to post bonds payable to those who can prove their stories false, few have ever done so.”
Your second sentence does NOT demonstrate that news customers are not interested in news accuracy.
It shows only that news incumbents and entrepreneurs/investors have to date not believed they could profit from the posting of such bonds.
Those incumbents and never-were entrepreneurs may or may not have been wrong. But suggesting that their absence demonstrates that customers are not interested in news accuracy is like suggesting in 1998 that customers were not interested in smartphones with internet access and touchscreen UIs because they hadn’t purchased any.
I would really like this to happen. I think that one thing we will need to be able to measure is "how well-substantiated" is this argument. Well-substantiated but still wrong is a category, of course, but when deciding whether to even bother reading an article it helps to know the track record of the journalists involved. The ones who are mostly writing historical fiction with the historical-period set to "right now" can be skipped, but only if we can judge this ahead of time.
One suggestion for your proposed user interface: if we lean into "true is blue" (instead of green), we will have a result that is usable by the 8% or so of the population which has red or green colour deficiency.
Agreed in particular with your first paragraph.
Reputation matters.
The MSM has willingly given away most of theirs as they both seek to promote the ideological agenda of most of their members, and switch to a business model of selling to the most activist, extreme leftists. Which reduces their reputation still further amongst those not on the hard left.
And this death spiral will likely continue.
The most common way to give a false impression is by selectively picking facts that aren't wrong and ignoring other relevant facts that also aren;t wrong.
Note also that it is always necessary to select facts. So false stories and accurate stories often both look fine on the surface.
One example: "You use some of the same words that terrorists use". That is undeniably correct - terrorists use words like "the", "and", "sometimes", etc. - but it can convey an impression of you being a terrorist. String a buunch of statements along those lines together, and you get a completely stupid story made out of completeley undeniably true factual statements. it's how the media makes stories to attract attention, and social media selects such things for promotion.
I don't see how any market mechanism or fact checking helps correct those types of spreading of false understanding.