信 - 麻省理工学院人工智能研究专业

信

Communicated by Terrence Sejnowski

TruthSift: A Platform for Collective Rationality

Eric B. Baum
eric@truthsift.com
TruthSift, 普林斯顿大学, 新泽西州 08540, 美国.

TruthSift is a cloud-based platform that logically combines members’
contributions into a collective intelligence. Members add statements and
directed connectors to diagrams. TruthSift monitors which statements
have been logically established by demonstrations for which every chal-
lenge raised has been refuted by an established refutation. When mem-
bers run out of rational objections, the result is a converged diagram
succinctly representing the state of knowledge about a topic, 包括
plausible challenges and how they were refuted. Previous computer sys-
tems for collaborative intelligence did not have a qualitatively better
solution for combining contributions than voting and are subject to
groupthink, interest group capture, and inability to follow a multistep
logical argument. They did not settle issues automatically point by point
and logically propagate the consequences. I review indications that many
practically important statements most people believe to be firmly es-
tablished will be revealed to be firmly refuted upon computer-assisted
审查. TruthSift also supports construction of powerful probabilis-
tic models over networks of causes, implications, 测试, and necessary
因素.

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d

我

F
/

3
5
3
5
3
6
2
0
7
1
8
3
2
n
e
C
哦
_
A
_
0
1
5
6
2
p
d

1 介绍

What is a proof? According to the first definition at Dictionary.com, a proof
is “evidence sufficient to establish a thing as true, or to produce belief in
its truth.” In mathematics, a proof is equivalent to a proof tree that starts
at axioms, which the participants agree to stipulate, and proceeds by a se-
ries of steps that are individually unchallengeable. Each such step logically
combines several conclusions previously established or axioms, 或两者.
The proof tree proceeds in this way until it establishes the stated proved
结论. Mathematicians often raise objections to steps of the proof, 但

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

Readers who are interested in exploring applications of the technology explored
in this letter to their own problems or organization, please contact me. I also suggest
going to https://tssciencecollaboration.com/graph/You%20should%20use%20TruthSift
%20for%20your%20or%20your%20organizations%20plans%2C%20projects%2C%20doc
uments%2C%20and%20decisions/289/0/-1/-1/0/0#lnkNameGraph

神经计算 35, 536–553 (2023)
https://doi.org/10.1162/neco_a_01562

TruthSift

537

if it is subsequently established that all such objections are invalid or if a
workaround is found around the problem, the proof is accepted.

The scientific literature works in a similar way. Each paper adds some
novel argument or evidence that previous work is true or is not true or
extends it to establish new results. When people run out of valid, 小说
reasons why something is proved or not proved, what remains is an estab-
lished theory or a refutation of it or of all its offered proofs.

TruthSift is a platform for diagramming this process and applying it to
any statements members care to propose to establish or refute. 一个可能
state a topic and add a proof tree for it, which is drawn as a diagram with
every step and connection explicit. Statements are simply added as replies
to previously added statements, either a supporting reply or a refuting re-
层数. If somebody thinks they have found a hole in a proof at any step or
thinks one of the original assumptions needs further proof, they can chal-
lenge it, explaining the problem they see. Then the writer of the proof (或者
others if it is in collaboration mode) may edit the proof to fix the problem
or make clearer the explanation if they feel the challenger was simply mis-
taken and then may counterchallenge the challenge explaining that it had
been resolved or mistaken. This can go on recursively, with someone point-
ing out a hole in the proof used by the counter-challenger that the challenge
was invalid. With TruthSift, the whole argument is laid out graphically and
essentially block-chained, which should prevent the kind of edit wars that
happen for controversial topics on Wikipedia (Attkisson, 2015; Wilson &
Likens, 2015). Each challenge or post should state a novel reason, 什么时候
the rational arguments are exhausted, as in mathematics, what remains is
either a proof of the conclusion or a refutation of it or all of its proofs.

As statements are added to a diagram (see Figures 1, 2, 和 3), TruthSift
keeps track of what is established and what is refuted, writing “TE” on ten-
tatively established statements and “TR” on tentatively refuted statements
(“tentative” meaning based on all the statements that have been added to
the graph so far though more replies could be added and change the result).
You can instantly tell what is established and what refuted. TruthSift com-
putes this by a simple algorithm that starts at statements with no incoming
边缘 (no incoming challenges or proofs), which are thus unchallenged as
assertions that prove themselves, are self-evident, or appeal to an authority
everybody trusts. These are considered established. (If you don’t trust the
权威, that is a valid reason to challenge.) Then it walks up the diagram
rating each statement in turn after all its parents have been rated. A state-
ment will be established if none of its challenges are, and if it has proofs,
at least one is established. A challenge may request that a proof be added
if a statement does not have one already or adequately prove itself, or if it
states a reason that the existing established proofs do not, 放在一起,
establish the statement. If a statement has an established challenge or if all
of its proofs are refuted, it is refuted.

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d

我

F
/

3
5
3
5
3
6
2
0
7
1
8
3
2
n
e
C
哦
_
A
_
0
1
5
6
2
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

538

乙. Baum

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d

我

F
/

3
5
3
5
3
6
2
0
7
1
8
3
2
n
e
C
哦
_
A
_
0
1
5
6
2
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

数字 1: Screenshot of a portion of a graph. The conclusion (a.k.a. the “topic
statement”) of the graph is at the top. Replies are rectangles showing the title
of the reply, whether it is TE or TR, and with a green bar at the bottom if the
reply supports the topic statement (PRO) and a red bar if it opposes it (CON).
A supporting reply of the topic statement or of a PRO statement is PRO, as is a
rebuttal to a CON statement. A supporting reply of a CON statement is CON,
as is a rebuttal to a PRO statement.

TruthSift

539

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d

我

F
/

3
5
3
5
3
6
2
0
7
1
8
3
2
n
e
C
哦
_
A
_
0
1
5
6
2
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

数字 2: Detailed screenshot of the upper right corner of Figure 1. Whichever
node of a graph is selected, that statement is shown on the right side. Selecting
“REPLY” pops a window in which you may choose if the reply is SUPPORT or
REFUTE and enter content. 因此, the graph can be built up. After each state-
ment is added, TruthSift updates the ratings. The body of the statement defines
它.

If it is not refuted, then it follows that all of its challenges are refuted,
and it has an established proof or is accepted as providing one, and so it
is established. Work your way backward on a diagram, centering on each
statement in turn, and examine the reasons why it is established or refuted.
It is important to understand that TruthSift directly maps mathematical
实践. 如果, based on published proofs and refutations and proof fixes and
counter-refutations, a rational human would consider a proof to be estab-
lished by the TruthSift diagram, then so will TruthSift’s simple rating algo-
rithm. In mathematical practice, what it means to be established is whether
there is a step-by-step proof offered with all the proposed refutations
反驳.

TruthSift implements a mapping of actual human mathematical practice
into a machine-human collaboration that enables it to extend to all fields

540

乙. Baum

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d

我

F
/

3
5
3
5
3
6
2
0
7
1
8
3
2
n
e
C
哦
_
A
_
0
1
5
6
2
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

数字 3: TruthSift also supports an alternative tree view of graphs. In the tree
看法, the structure of connections is shown by indents, so that just as in making
an outline, the user indents one tab each time he or she goes one deeper in the
hierarchy and shifts left one tab each time he or she goes up a level in the hier-
archy. I find it easier to navigate large graphs in the Tree view and far better for
viewing graphs on cell phones. Only the top of this graph is shown.

of human endeavor, not just mathematics. The algorithm decides the same
thing a rational human would about what has been established given the
contributions to date, and thus serves as an unemotional arbiter, 使能
human collaboration, even if it is in some cases adversarial, to achieve su-
perhuman feats of reasoning.

It also forces the discussion to be more precise. If you believe it has been
demonstrated that “vaccines are safe,” how much residual damage was al-
lowed in the determination? What precisely is the definition of safe? 和
TruthSift, there is a spelled-out topic statement. Careful phrasing of state-
ments is essential (just as in mathematical practice) to being able to actu-
ally establish them and enables more to be established than one might have
预测的.

TruthSift

541

Without TruthSift, people are often unable to agree on what the truth is.
With TruthSift, they have a block-chained, semantically organized diagram
of the course of their argument, and when they’ve reached a conclusion, 全部
objections raised have been addressed. TruthSift doesn’t always succeed in
establishing truth, but it does show when something is provable.

The process divides a problem into natural subtopics that get settled
point by point, although not sequentially in time: an old conclusion about
a subtopic can always be challenged with a new rational argument or evi-
登塞. These different subtopics may involve different collaborators or op-
ponents. The process guides people in the essence of critical thinking and
fruitful collaboration.

The process leaves a careful record of how decisions were made and how
all the arguments against every subpoint were rebutted. This is like the sci-
entific literature diagrammed and automatically updated as new conclu-
sions are added. This documentation is useful for businesses quite aside
from the advantages of getting the right answer and facilitating collabora-
的. The transparency exposes if some employee has an agenda other than
the bottom line or truth. The process is also useful for teaching the essence
of critical thinking and would make for far better refereeing and author re-
sponse process than currently in use in the sciences (TruthSift, n.d.c). 这
documented decision processes provide data to refine decision making over
时间.

For TruthSift to work properly, posters should respect the guidelines and
make a good-faith effort to post only proof or challenge statements that they
believe rationally support or rebut their target. Posts do not have to be cor-
直角 (that’s what challenges are for), but they have to be honest attempts,
not spam or ad hominem attacks.

Frequently a user wants to assemble arguments for a proposition stating
something like, “The preponderance of the evidence indicates X,” and these
arguments are not individually proofs of X. It is safe to simply add them as
proofs of the proposition. If not enough of them are established, the target
may be challenged on that basis. The goal is a diagram that transparently ex-
plains a proof and what is wrong with all the objections people have found
plausible. Edits that move in that direction are useful and desired.

TruthSift ran a public site where members could debate issues of their
choosing. There were a few dozen dedicated users, but we believe that the
concepts of establishing proofs are difficult for a consumer audience. 一个-
other problem is that many things people believe are not in fact true (Le Bon,
1895; Bernays, 1928; TruthSift, n.d.a), and they are disturbed to see this and
turn away from a technology they do not fully understand, thinking it must
be crackpot. 还, I designed the user interface, and it was too sophisticated
and not user-friendly enough. I thought it was great for a power user, 但
opinion is unanimous that it was hard to train novices in.

We have now pivoted and are trying to sell decision-making services to
公司. People are more motivated to learn to use TruthSift for their

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d

我

F
/

3
5
3
5
3
6
2
0
7
1
8
3
2
n
e
C
哦
_
A
_
0
1
5
6
2
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

542

乙. Baum

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d

我

F
/

数字 4: Screenshot of a workspace containing seven graphs, each making a
case for a different vendor.

job than as a social network, we can provide guidance, and they are less
likely to engage in ad hominem attacks and the like. Corporations need
documentation on their decisions, and they need to make better decisions,
so there is a lot more motivation. 基本上, we are bringing the scientific
method to business decision making, so we expect there will be a lot of
room for profit in it.

The user interface as presented here has been greatly simplified. 对于前-
充足, we cut back from multiple reply types to just two: support and re-
fute. (We are currently supporting a site where users may freely debate sci-
entific topics at https://tssciencecollaboration.com. Users can sign up for a
free account.)

For business users, we have added features such as a dashboard with
multiple work spaces. One business decision for which TruthSift has been
used repeatedly is deciding among alternatives such as applicants for a po-
位置, vendors for a technology, or marketing plans (见图 4). We sup-
port multiple shared workspaces, each of which may be a graph debating,
说, whether a particular applicant is qualified for the job. The author of a
workspace that contains graphs for each applicant also may designate sev-
eral features on which to weigh the different candidates. The contributors
to the individual graphs provide numerical assessments of the candidate

3
5
3
5
3
6
2
0
7
1
8
3
2
n
e
C
哦
_
A
_
0
1
5
6
2
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

TruthSift

543

according to these criteria, in addition to providing pro and con statements
to the graphs. TruthSift keeps track of a weighted score for each of the can-
didates, as well as whether his or her qualifications are tentatively estab-
列出. This aids the comparison of the qualified candidates, and one may
also create a graph establishing that the selected candidate is in fact the
right one. Templates are available for various types of business decisions.
As a business refines its decision process with experience, these will likely
be improved.

2 This Technology Is Important Because We Have Suffered

from Its Lack

Peer-reviewed surveys agree: a landslide majority of medical practice is not
supported by science (Ezzo et al., 2001; Garrow, 2007; Greenberg, 2009; 的-
fice of Technology Assessment, 1978). Within fields like climate science and
vaccines that badly desire consensus, no true consensus can be reached be-
cause skeptics raise issues that they feel are not adequately addressed by
大多数 (exactly what Le Bon, 1895, warned of more than 100 年
前). Widely consulted sources like Wikipedia are reported to be largely
paid propaganda on many important subjects (Attkisson, 2015) 或者, 最好,
the most popular answer rather than an established one (Wilson & Likens,
2015). Quora shows the most popular answer, not necessarily the correct
一, and the answers are individual contributions, with little or no collabo-
配给, and often there is little documentation of why anyone should believe
他们. Existing systems for crowd- wisdom thus largely compound group
think rather than addressing it.

Corporate or government planning is no better. Within large organiza-
系统蒸发散, where there is inevitably systemic motivation to not pass bad news
向上, leadership needs active measures to avoid being kept in the dark as to
the real problems (Siitari et al., 2014). Corporate or government plans are
subject to groupthink or takeover by employee or other interests compet-
ing with the mission. Individuals who perceive mistakes have no recourse
capable of rationally persuading the majority and may anyway be discour-
aged from speaking up by various consequences (LeBon, 1895).

Feynman (1974) famously described the phenomenon of cargo cult sci-
恩斯. In the South Seas, the natives saw planes landing on airstrips and
delivering cargo, so in an effort to get cargo, they built detailed replicas of
landing strips, including wood radios. Feynman observed that the science
that results when practitioners do not rigorously discuss and address all the
opposing arguments, but instead ignore them or replace them with straw-
男人, is cargo cult science. Like the islanders’ radios, it is missing the key
ingredient—the integrity to examine publicly every objection raised and
rationally refute it—and so will not function.

We see such examples when the medical literature is assessed. Appar-
ently many researchers have been using the wrong cell lines and know it,

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d

我

F
/

3
5
3
5
3
6
2
0
7
1
8
3
2
n
e
C
哦
_
A
_
0
1
5
6
2
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

544

乙. Baum

but haven’t done anything about it. When a doctor prescribes medicine, 它
may have been mistakenly tested on, 说, a different cancer (Fung, 2014).

In the vaccine literature, apparent systematic biases corrupting almost
all their measurements are never discussed (Baum, 2016). 同样的道理
of areas of proven danger such as contaminants, aluminum, timing during
发展, or multiple vaccine interaction. (Stratton et al., 2002, 2012,
provide Institute of Medicine safety surveys). For a graph proving the ma-
jority is often wrong, even when they are credentialed and widely admired,
see TruthSift (n.d.a.).

3 Private Planning Anecdotes

Users also report that private diagrams have unique potential for intu-
itive yet rigorous personal and business planning. TruthSift breaks debates
down and settles them point by point, propagating the rational conse-
quences of one conclusion on to other decisions. To decide A, a user can
create a statement, “Evidence for A,” with all the arguments he or she can
think of as proof statements and “evidence for Not A” with proof state-
ments of that. Then this person shold consider the counterarguments on the
specific arguments, followed by what actual proofs and refutations can be
offered to the overall conclusions. Everybody can contribute their point of
看法. Experience indicates that users find the process intuitive and helpful.

4 Previous Related Work

Mathematicians usually write proofs to appeal to other mathematicians,
leaving many of the steps somewhat intuitive and incompletely described.
然而, 原则, every correct mathematical proof should be re-
ducible to a series of syntactic checks. Languages like HOL Light, Coq,
PVS, Isabelle, and Mizar have enabled mathematicians to write computer
programs checking many major results (Hales, 2008), and there is an ongo-
ing project to verify all of mathematics (Formalizing 100 Theorems, 2016).
TruthSift, 相比之下, aims to provide formal verification for informal ra-
tional discourse such as occurs in the social and physical sciences.

There has been research going back to the Middle Ages on formal sys-
tems for persuasion dialog, reviewed by Prakken (2006). These to some
extent mirror (but formalize) the intuitive discussion of the mathematical
过程, as does TruthSift. One way I differ philosophically from some of
this work is in hypothesizing that there is an objective reality and maybe
also a Platonic reality that can help guide members to discovery and agree-
蒙特. Another example of modeling mathematical argument was Lakatos
游戏 (Pease et al., 2014). TruthSift appears to be the first to implement a
representation of such a system on a platform for collaboration and point-
by-point proof and refutation, with machine verification of what is estab-
列出, and to test it on general topics.

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d

我

F
/

3
5
3
5
3
6
2
0
7
1
8
3
2
n
e
C
哦
_
A
_
0
1
5
6
2
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

TruthSift

545

Feynman (1968) 写了, “That is what science is: the result of the discov-
ery that it is worthwhile rechecking by new direct experience, and not nec-
essarily trusting the [人类] 种族[’s] experience from the past. I see it that
方式. That is my best definition. . . . Science is the belief in the ignorance of
experts.” TruthSift directly implements this vision of science, requiring that
all objections raised be rationally refuted before a conclusion can be estab-
列出. What is constant in cargo cult science and other delusions is that the
objections raised are simply ignored and usually censored.

There have been a number of previous efforts to achieve crowd intelli-
gence or to map arguments, 或两者. (For recent reviews, see Michelucci
& 狄金森, 2016; 克莱因, 2015; and Quinn & Bederson, 2011. For a sur-
vey of argument maps, see Dwyer et al., 2012.) In Klein’s (2015), classifi-
阳离子, TruthSift may be considered to have a novel means of aggregation
of contributions compared to any surveyed methods, combining contribu-
tions according to logic, and a novel means of quality control, using the
aggregation to compute what is logically established. Another characteris-
tic used to classify existing systems is the motivation of users to contribute
to public collaboration, for which they list the existing alternatives: 支付, 阿尔-
truism, enjoyment, reputation, and implicit work (克莱因, 2015). 此外
to invoking altruism, enjoyment, and reputation, TruthSift also offers the
motivation for individuals or organizations (例如, advertisers or advocates)
that want to publish a verified proof of some proposition,. It also supports
private collaborations within corporations to make better decisions.

In the classification of Michelucci and Dickinson (2016), ”The human
computation ecosystems of the future” are type C systems that “combine
the cognitive processing of many human contributors with machine based
computing to build faithful models of the complex, interdependent systems
that underlie the world’s most challenging problems.” However, the exist-
ing prototype C systems seem to be for specialized applications and also not
entirely automated, like the polymath project (Gowers & Nielsen, 2009).

Notable efforts to provide crowd-sourced question answering or infor-
mation include Wikipedia and Quora. Wikis suffer on controversial topics,
and Wikipedia has been reported to represent paid propaganda on contro-
versial scientific topics (Attkisson, 2015; Wilson & Likens, 2015). 维基百科
reports only the last edit, and this is often apparently controlled by influ-
enced parties. On TruthSift, such edit wars would be transparent, flagged
as contrary to the guidelines, and possibly available as specific alterna-
tive stipulations. Quora reports the most popular answer to a question
rather than any more powerful collaboration, together with a list of alterna-
tive answers. Other fact-checking or question-answering websites present
some “expert opinion,” justified if at all by argumentation that demonstra-
bly often comprises logical fallacies such as straw men and ad hominem
attacks.

Klein’s (2015) review of online deliberation technologies classifies sys-
tems into time-centric, question-centric, topic-centric, debate-centric, 和

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d

我

F
/

3
5
3
5
3
6
2
0
7
1
8
3
2
n
e
C
哦
_
A
_
0
1
5
6
2
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

546

乙. Baum

argument-centric. Time-centric sites like Twitter or comment threads orga-
nize content by when it is contributed. Question-centric sites like Quora
organize answers by questions. As Klein notes, both of these types of sys-
tems produce voluminous output, most of which is of low quality thought
there might be some concealed. There is little collaboration beyond voting
on one most popular individual contribution.

Topic-centric sites like wikis organize content by topic, producing a more
concise output, but they suffer on controversial topics, becoming battle-
grounds that are not won on the basis of logic. Debate-centric sites like
whysaurus.com, debatepedia.com, debatewise.org, and debate.org allow
users to contribute pros and cons. 然而, they have no method of break-
ing down issues into subpoints and establishing point by point what can be
已确立的. They do not allow linking of arguments to arguments, much
less automatic update of consequences, and they do not support reasoning
forward on open-ended problems, as is possible with TruthSift.

Argument-centric systems like Klein’s MIT Deliberatorium (克莱因, 2011)
(and TruthSift) have advantages. Members prepare a concise and informa-
tive diagram summarizing an argument. Solutions are genuinely collabo-
rative, and thus potentially powerful. The idea of determining statements
to establish and work forward to establish other statements is integral to
science but missing from previous argument map systems like the Deliber-
atorium. Without that, as with Quora and Wikipedia, there is no good way
of composing the right answer, as opposed to merely selecting the most
popular answer at each choice. Crowd-based systems mostly compound
groupthink, not correct it.

The best existing system for determining truth using natural language
argumentation and evidence by a large collaboration has been the scien-
tific literature. It too, 然而, has suffered under questionable refereeing,
the lack of a good mechanism to agree on what is actually established at
any given time, and the ignorance or disregard of contrary arguments by
authors and referees or editors. It would be much improved if it adopted
TruthSift for refereeing (TruthSift, n.d.c).

5 Probability Mode

TruthSift supports probabilistic ratings if switched to probability mode. (看
数字 5.) This allows users to easily collaborate on constructing a Bayes net
modeling any question on all supplied evidence and automatically com-
putes the probabilities predicted using a fast Monte Carlo algorithm. 你
may make a number of observations about the world that are pertinent to
whether a given hypothesis is true, 例如, “It was reported in the news-
paper” or ”The newspaper isn’t always accurate,” or “If that hypothesis
was true, some other thing I observed would be much more likely to hold
than if it were false.” TruthSift offers combining assessments and intuitions

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d

我

F
/

3
5
3
5
3
6
2
0
7
1
8
3
2
n
e
C
哦
_
A
_
0
1
5
6
2
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

TruthSift

547

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d

我

F
/

数字 5: A graph drawn in probability mode. A Bayesian graph may be com-
posed of proof statements whose truth makes their parent more likely, chal-
lenges whose truth makes the parent less likely, and test statements for which
the author supplies two parameters: the likelihoods given target true and false.
TruthSift runs a fast Monte Carlo estimation of the probability of each node. 这
selected test node is shown in the upper right. The likelihood that Jane actually
has breast cancer is reported as the probability of the top statement as 1% 后
rounding or more precisely if the top node is selected.

about different pieces of evidence into a rigorously calculated probability
estimate.

The author of a topic can set it to probability mode by checking the “make
this a probability graph” box right below the title on the “add new topic”
window. In probability mode, contributors are asked to assign a ”proposed
信仰” to the statements they add, between zero and one, to reflect how
much confidence they have that their statement either proves or refutes its
目标.

In probability mode, in addition to pro and con statements, users can
add test statements, which are like epidemiological tests that provide ev-
idence favoring the truth or falsehood of their target. Given some obser-
vation claimed in the body of the test statement, the author of the test
statement supplies a likelihood of the observation given that the target
statement is true and the likelihood of the observation given that the

3
5
3
5
3
6
2
0
7
1
8
3
2
n
e
C
哦
_
A
_
0
1
5
6
2
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

548

乙. Baum

target statement is false. According to Bayes’s law, this evidence multiplies
the likelihood of the target statement by the ratio of the first number to the
第二.

例如, consider the medical test for breast cancer (见图 5).
It has a false-positive probability of .12, meaning for a woman without the
状况. 有一个 .12 probability the test will show positive anyway, 和
it has a false-negative probability of .27, meaning that if she does have the
状况, 有一个 .27 the test says she does not and thus a .73 chance it
says she does. 因此, a woman who has breast cancer is .73/.12 = 6 次
as likely to get a positive test as one who doesn’t. But since only 1 在 700
women has breast cancer, the likelihood of a woman with a positive test
having breast cancer is 6/700 = .00857.

TruthSift probability mode allows the woman to represent this and cal-
culate the likelihood of having breast cancer, using a test node with a likeli-
hood given true target = .73 and a likelihood given false target of .12, and a
prior node giving proof with expected belief 1/700. (See for this probability
图形, TruthSift, n.d.b. This simple example is taken from the discussion in
Pearl & Mackenzie, 2018, PP. 104–105.)

TruthSift then calculates the probability that the topic node is true (这
woman actually has cancer) by drawing instances from the population con-
ditioned on having a positive test. It does this by sampling instances with a 1
在 700 chance of having cancer and then weighting the women who actually
have cancer by the factor .73 (the likelihood estimate given target true) 和
then weighting ones who do not by the factor .12 (likelihood estimate given
target false). Then the likelihood that the topic node is true is given as the
ratio of the total weight of positive instances divided by the total weight of
examples. This is a Monte Carlo calculation that gives as the expected value
the same result as the Bayesian calculation.

TruthSift thus allows building a big causal model with epidemiological
tests bringing to bear all of the evidence on a given subject and rapidly esti-
mates the probability of the statements. Those who are unskillful in adding
to the graph multiple times different nodes that are highly correlated, ba-
sically the reflection of a single measurement and not adding independent
信息, cannot expect the results to be accurate. 然而, my experi-
ence is that in the hands of a practiced analyst, it is invaluable for intelli-
gence modeling.

6 Discussion and Future Work

TruthSift is attempting to create a collective rationality in a way that has
not been tried before. We have had in the past a number of public users
on our consumer websites and now have corporations testing for private
decision making, but it remains to be seen whether it will recruit a huge base
of users or what would result from large-scale public or enterprise use. 这
expectation is that members will explore rational diagrams as is somewhat

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d

我

F
/

3
5
3
5
3
6
2
0
7
1
8
3
2
n
e
C
哦
_
A
_
0
1
5
6
2
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

TruthSift

549

achieved in the scientific literature, but it is possible that too many members
will make irrational posts or that it will be hard to agree on what is rational
or that people won’t get over the learning curve to use it.

TruthSift diagrams are currently demonstrating that there are huge logic
holes in things many people believe, so the potential impact of Truth Sift on
society is substantial. 而且, it may demonstrate that there are many
similar costly misconceptions and groupthink delusions throughout the
corporate, intellectual, and government worlds. A noncontroversial exam-
ple of a widely held confusion in the business world is given by Money-
ball (Lewis, 2004). Baseball managers were systematically misvaluing assets
worth hundreds of millions of dollars, although Bill James (1982) 曾经是
trying to point this out for decades. Without TruthSift, people mostly disre-
gard any evidence or arguments against their cherished positions. TruthSift
flags their conclusions until they have come up with a rational argument ad-
dressing the issues. After a TruthSift diagram is built, assessing the state of
the argument is quick.

A critical reason for the propagation of groupthink delusions is that most
people do not take the time to understand the issue but assume others are.
But with TruthSift, they do not have to invest much time to verify so can
come to rely on TruthSift as a better authority than the group. 因此, a rela-
tively small group of contributors can have a large impact on society.

TruthSift expects to add more features, including n-choice statements
(one and only one of the n statements may be established, such as negations
that are 2-choice); connectors between diagrams, allowing a big web of ver-
ified concepts to be built; and a system allowing rewards to be posted for
contributions to topics that are significant because they affect the status of
the topic and remain established. This should allow members to motivate
researchers to a question and also to make claims that they have demon-
strated something in spite of offering a reward for disproof.

TruthSift could be used by AIs rather than or in addition to humans,
provided they could sometimes suggest proofs or refutations or tests of
statements. TruthSift is currently experimenting with AI-human collabora-
tive posts. The criticism of TruthSift may enable a collection of AIs, 也许
with some humans, to produce a higher-level intelligence than any of them
individually could achieve. Even if they sometimes interject challenges or
proofs that are off-base, if they more often can identify the problem and
challenge it, they may be able to bootstrap to a more rational intelligence.
同时, they can help create a wider base of interesting and ex-
plored topics.

Readers who are interested in exploring applications of the technology
explored in this letter to their own problems or organization, please contact
我. I also suggest going to https://tssciencecollaboration.com/graph/You
%20should%20use%20TruthSift%20for%20your%20or%20your%20organi
zations%20plans%2C%20projects%2C%20documents%2C%20and%20deci
sions/289/0/-1/-1/0/0#lnkNameGraph.

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d

我

F
/

3
5
3
5
3
6
2
0
7
1
8
3
2
n
e
C
哦
_
A
_
0
1
5
6
2
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

550

Conflicts of Interest

乙. Baum

I founded/own TruthSift Inc and have patents on the technology.

Coda: Many Worlds Theory: An Empirical Observation and Memoir

When in graduate school at Princeton, 我认为 1978, I conceived the notion
that the many worlds theory of quantum mechanics predicted that I would
eventually be the oldest person in the world because there would always
be a branch in which I was still alive, and in that branch I would remember
making this prediction, and since a 1 在 8 billion event had successfully pre-
dicted, I would eventually know in finite time that the many worlds theory
was correct (although if it was false I might never gain such confidence).

On October 7, 2016 in the evening I was at a bonfire with my wife and
small daughters on Lake Tahoe Beach thrown by the Princeton Club of
Northern California. My daughters’ fingers got sticky on smores, so my
wife took them to the lake to wash them, leaving me talking to somebody
who had introduced himself as also a Baum. When they came back a few
minutes later, 我 (and he) were gone. Shortly after she came upon me where I
had been found by a passing doctor lying in soft sand with my neck broken
and my heart stopped. I remember none of this; my first recollection is from
waking up later in an excellent trauma center in Reno.

Since then I have been a vent-dependent quadriplegic. I have a good life
仍然. I have a wife and daughters who love me and I love them. In spite of
the fact my heart stopped for some indeterminate period in the dark be-
fore I was found, my mind is still sharp. I play far better chess than I ever
做过 (2350 on worldchess.net). I get Queen Bee level on the New York Times
website “spelling bee” in under 30 minutes almost every morning. I read
widely and have had about the best understanding of the bizarre evolu-
tion of the world over the last decade of anyone I know. I am fortunate for
a fracture of my type in being able to eat anything normal people do and
speak well. I have excellent technology to communicate with my computer
(Glassouse Bluetooth device I wear like glasses that moves the mouse with
my head, bite switch to click, speech recognition), have a mobility van, A
great 24/7 nursing staff and a power wheelchair I operate with a mouth
joystick.

A few years before my injury I invented TruthSift. It supports the con-
struction and verification of proof refutation trees by individuals or collab-
orators for any statements you care to debate. This settles issues point by
point and propagates the conclusions up. People always have a strong ten-
dency to ignore the arguments against their position and TruthSift prevents
这. It makes clear what you actually know, how you know it, and how all
arguments against have been refuted. Basically it implements the Platonic
ideal of the scientific literature as a truth verification and discovery system
and applies it to any situation in life. We are currently marketing it as a

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d

我

F
/

3
5
3
5
3
6
2
0
7
1
8
3
2
n
e
C
哦
_
A
_
0
1
5
6
2
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

TruthSift

551

scientific and far superior approach to business decision making, 包括
specializations for various subfields such as hiring.

Living as I do on mechanical ventilation, my life is precarious. 我开始于
lying for an indeterminate time with a broken neck and my heart stopped
before presumably somebody walking through the dark on the beach found
我, and he happened to be a doctor and revived me. On at least a dozen
occasions since then, my airway has become obstructed or disconnected to
the point that I passed out. This mostly happens due to the fact that my
body regards the trach as a foreign object and produces mucus in an effort
to get rid of it. This mucus has to be suctioned out periodically so that I can
breathe, it can get dried into plugs that can clog the airway. 例如,
as I write, this happened this morning. Then the oxygen in my blood drops
while a nurse comes running to try and suction the plug out of the way.
But it’s not always that easy to get the plug out, and a nurse doesn’t always
respond immediately, so on a dozen occasions, I have passed out before
they removed the plug, and on many more come close.

I haven’t quite become the oldest man in the world yet, but I have sur-
vived a series of threats that seem to be very long odds, so I am basically
ready to conclude that the many worlds theory is correct. It’s unfortunate
my loved ones are mourning me in all the other universes.

致谢

An earlier version of this letter was presented as “TruthSift:A Platform for
Collective Rationality” at the IEEE 2016 Future Technologies Conference.
Thanks to Chaitra Keshav-Baum, Amit Sengupta, Vineet Velso, 和别的
team members and contributors to TruthSift for contributions to website
设计, 软件, and diagrams, and to Stefi Baum for a critical reading of
the manuscript.

参考

Attkisson, S. (2015). Astroturf and manipulation of media messages. TEDx. 大学-

sity of Nevada, https://www.youtube.com/watch?v=-bYAQ-ZZtEU

Baum, 乙. (六月 28, 2016). The top ten reasons I believe vaccine safety is an epic mass

delusion. Atlantic (四月 23).

Bernays, 乙. (1928). Propaganda. Ig Publishing.
Dwyer, C. P。, Hogan, 中号. J. & 斯图尔特, 我. (2012). An evaluation of argument mapping
as a method of enhancing critical thinking performance in e-learning environ-
评论. Metacognition Learning, 7, 219–244. 10.1007/s11409-012-9092-1

Ezzo, J。, Bausell, B., Moerman, D. E., Berman, D ., & Hadhazy, V. (2001). 审查-
ing the reviews. International Journal of Technology Assessment in Health Care,
17, 457–466. http://journals.cambridge.org/action/displayAbstract?fromPage=
在线的&aid=101041. 10.1017/S0266462301107014

Feynman, 右. 磷. (1968). What is science? Physics Teacher, 7(6), 31–320. http://www-oc
.chemie.uni-regensburg.de/diaz/img_diaz/feynman_what-is-science_68.pdf

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d

我

F
/

3
5
3
5
3
6
2
0
7
1
8
3
2
n
e
C
哦
_
A
_
0
1
5
6
2
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

552

乙. Baum

Feynman, 右. 磷. (1974). Cargo cult science. Commencement address at CalTech.
https://sites.cs.ucsb.edu/∼ravenben/cargocult.html http://discovermagazine
.com/2014/nov/20-trial-and-error
https://lifeboat.com/blog/2016/06/the
-top-ten-reasons-i-believe-vaccine-safety-is-an-epic-mass-delusion

Formalizing 100 theorems. (2016). http://www.cs.ru.nl/∼freek/100/
Fung, 乙. (2014). “Oops, wrong cancer”: How contaminated cell lines produce bad

研究. Atlantic (四月 23).
(2007). 什么

Garrow,

J. S.

to do about CAM? How much of orthodox
medicine is evidence based? BMJ, 335(7627), 951. http://www.dcscience.net/
garrow-evidence-bmj.pdf. 10.1136/bmj.39388.393970.1F

Gowers, T。, & Nielsen, 中号. (2009). Massively collaborative mathematics. 自然,

461(879), 879–881. 10.1038/461879A

Greenberg, S. A. (2009). How citation distortions create unfounded authority: Anal-

ysis of a citation network. BMJ, 339, b2680. 10.1136/bmj.b2680

Hales, 时间. 磷. (2008). Formal proof. Notices of the A.M.S. 55 (11), 1355–1380.
James, 乙. (1982). The Bill James baseball abstract. Ballantine Books.
克莱因, 中号. (2011). The MIT Deliberatorium: Enabling large-scale deliberation about
complex systemic problems. International Conference on Agents and Artificial Intel-
智慧. 罗马, 意大利.

克莱因, 中号. (2015). A critical review of crowd-scale online deliberation technologies. SSRN.

http://ssrn.com/abstract=2652888

Le Bon, G. ([1895] 1995). The crowd. Transaction.
Lewis, 中号. (2004). Moneyball: The art of winning an unfair game. 诺顿.
Michelucci, J. L。, & 狄金森, J. L. (2016). Human computation: The power of

crowds. 科学 351 (6268), 32–33. 10.1126/science.aad6499

Office of Technology Assessment, 我们. 国会. (1978). Assessing the efficacy and

safety of medical technologies. http://www.fas.org/ota/reports/7805.pdf

Pearl, J。, &, Mackenzie, D. (2018). The book of why: The new science of cause and effect.

基础书籍.

Pease, A。, Budzynska, B., 劳伦斯, J。, & 芦苇, C. (2014). Lakatos games for math-
ematical argument. 在S. 帕森斯, 氮. Oren, C. 芦苇, & F. Cerutti (编辑。), Pro-
ceedings of the Fifth International Conference on Computational Models. http://
comma2014.arg.dundee.ac.uk/res/pdfs/07-pease.pdf

Prakken, H. (2006). Formal systems for persuasion dialogue. Knowledge Engineer-
ing Review Archive, 21(2), 163–188. http://www.cs.uu.nl/groups/IS/archive/
henry/dgreview.pdf. 10.1017/S0269888906000865

奎因, J。, & Bederson, 乙. 乙. (2011). Human computation: A survey and taxonomy
of a growing field. In Proceedings of the SIGCHI Conference on Human Factors in
Computing Systems (PP. 1403–1412). 10.1145/1978942.1979148

Siitari, K., 马丁, J。, & 泰勒, 瓦. 瓦. (2014). Information flow in fisheries man-
代理: Systemic distortion within agency hierarchies. Fisheries, 39(6), 246–250.
10.1080/03632415.2014.915814

Stratton, K., Ford, A。, Rusch, E., & 克莱顿, 乙. 瓦. (Eds). (2012). Adverse effects of vac-

cines: Evidence and causality. 美国国家科学院出版社.

Stratton, K., Wilson, C. B., & McCormick, 中号. C. (编辑。). (2002). Immunization
safety review: Multiple immunizations and immune dysfunction. 国家科学院
按.

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d

我

F
/

3
5
3
5
3
6
2
0
7
1
8
3
2
n
e
C
哦
_
A
_
0
1
5
6
2
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

TruthSift

553

TruthSift. (n.d.a). On controversial topics who is more often right, 大多数
or a minority? [diagram]. https://tssciencecollaboration.com/graph/On%20
Controversial%20Topics%2C%20Who%20is%20More%20Often%20Right%2C%
20the%20Majority%20or%20a%20Minority/466/0/-1/-1/0/0#lnkNameGraph
TruthSift. (n.d.b). Does Jane Doe have breast cancer given a positive test? [dia-
https://tssciencecollaboration.com/graph/Does%20Jane%20Doe%20

公克].
have%20cancer%20%20given%20a%20positive%20test/964/0/-1/-1/0/0#lnk
NameGraph

TruthSift. (n.d.c). It would be far better to replace journal refereeing with TruthSift.
https://tssciencecollaboration.com/graph/It%20would%20be%20
[diagram].
far%20better%20to%20replace%20journal%20refereeing%20with%20TruthSift/
819/0/-1/-1/0/0#lnkNameGraph

Wilson, A. M。, & Likens, G. 乙. (2015). Content volatility of scientific topics in
维基百科: A cautionary tale. PLOS One, 10(8). 10.1371/journal.pone.0134454

Received April 28, 2022; accepted July 21, 2022.

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
n
e
C
哦
A
r
t
我
C
e
–
p
d