The previous article was about valuing what you made. This one is about handing something to someone who might not give it back, which is the other half of most commercial life and which has a cleaner arithmetic than anyone uses.
Key Takeaway
In the standard game the amount sent is tripled[1], so on our own arithmetic the trustee must return more than one third of what they receive for trust to pay. A citing account of the meta-analysis reports that roughly 40 percent of the variation in trust across experiments is attributable to differences in experimental protocols[2]. And a 2025 paper argues "we are not entitled to interpret the return of money as an indication of trustworthiness."[3]
The Verdict, Stated First
Five claims, in descending order of confidence.
One. The paradigm is dominant and its interpretation is under serious attack. A 2025 paper in a peer-reviewed journal argues the second mover's motive is indeterminate, so returning money cannot be read as trustworthiness.
Two. A large share of what the games measure is the games themselves. Roughly 40 percent of the variation in trust and 30 percent in trustworthiness is attributed to protocol differences, on one citing account of the meta-analysis.
Three. People demand better odds from a person than from a lottery. That is betrayal aversion, and on the stated payoffs of one variant a risk-neutral person should trust at any repayment probability above 28.6 percent.
Four. A finding this series would have reported as a failure was reversed. A 2000 study established that survey trust questions do not predict trusting behaviour; a 2018 reinvestigation reports that they do, when the standard game is used.
Five. The laboratory threshold does not transfer to business, and the direction matters. The game triples the money. Trade credit does not, and on our own arithmetic a firm on a 30 percent margin needs a 70 percent payment probability merely to break even.
Our Grades For These Claims
Applying the scheme from the first article in this series.
Grade A for the game's structure, obtained from an academic preprint reconstructing the original with its sample sizes, and corroborated by two further sources.
Grade A for the 2025 critique, from an abstract obtained verbatim from the publisher.
Grade C for the meta-analytic figures, which reach us through a citing preprint rather than the meta-analysis, and whose study count in that account we could not verify.
Grade B for betrayal aversion and for the 2018 reversal, both from repository pages reproducing text rather than from the papers.
Grade A for our own arithmetic, which is exact on the reported payoff rules.
Our position: the experiment is elegant, the interpretation is contested, and the commercial lesson comes from the payoff structure rather than from any measured quantity.
A Note On Method
Everything here is verified to August 2026.
We did not obtain the 1995 paper. Its design and sample sizes come from an academic preprint reconstructing it[1], and its citation from five independent reference lists[4].
We did not obtain the meta-analysis. Its findings reach us through one citing preprint[2], which reports a study count we could not corroborate elsewhere.
We obtained the 2025 critique's abstract verbatim from the publisher[3] and not the paper.
The 2018 reversal and the betrayal aversion account come from repository pages reproducing text from those papers[5][2], and we obtained neither paper.
The MAP variant's payoffs come from an academic preprint describing it[6], and our benchmark calculation rests entirely on those figures being correct.
All arithmetic is ours. The threshold and benchmark are exact given the reported rules; the credit section uses invented margins.
This article discusses research on trust. It is not credit, procurement or legal advice.
The 1995 Paper
The source.
Berg, J., Dickhaut, J., and McCabe, K. (1995), Trust, Reciprocity, and Social History, Games and Economic Behavior, 10(1), 122–142, DOI 10.1006/game.1995.1027[4].
A 2025 paper describes its standing: "The ubiquitous 'Trust Game' of Berg, Dickhaut, and McCabe is the dominant experimental paradigm for the study of trust and trustworthiness in experimental economics, social psychology, and neuroscience of the last two decades. This elegant framework, promising a quantifiable behavioral measure of trust and trustworthiness, is at the center of a strikingly broad range of research programs."[3]
Three observations, ours.
The paper is also known as the investment game, and a source describing it notes the first player transfers "without any assurance of reciprocation"[2], which is the defining feature.
The phrase "promising a quantifiable behavioral measure" is doing careful work in that 2025 abstract. Promising, not delivering, and the rest of the paper is about the difference.
And we did not obtain the paper itself, which for a work this central is a limitation we flag now rather than later.
The Design
The rules, from a preprint reconstructing the original.
It records: "Berg et al. (1995) studied trust and reciprocity in a two-stage investment game with University of Minnesota undergraduates (N = 120): No-History (n = 64, 32 pairs) and Social-History (n = 56, 28 pairs). Room A chose how much of a $10 endowment to send ($0-$10); the amount was tripled; Room B decided how much to return. Social history consisted of a report summarizing the prior 32 pairs' outcomes."[1]
Four observations, ours.
The tripling is the whole engine. It creates surplus from the act of transferring, which is what makes trust worth extending rather than merely generous.
The two conditions test something specific. Social history means the second cohort saw what the first cohort did, which is the experimental version of a reputation.
The samples are modest. Thirty-two pairs and twenty-eight pairs, which is a normal size for 1995 and small by any standard applied since.
And the design is anonymous and one-shot, which is what makes any return at all a finding, and which is also its largest distance from commercial life.
The One-Third Threshold
The arithmetic nobody states, and it is exact. Ours, on the reported rules.
You send X. It becomes 3X. The other party returns a fraction f of what they received, so you get 3fX back against the X you gave up. You profit if and only if f exceeds one third.
Per dollar sent, you receive back: at 20 percent returned, $0.60. At 25 percent, $0.75. At 33.3 percent, exactly $1.00. At 37 percent, $1.11. At 40 percent, $1.20. At 50 percent, $1.50. At 60 percent, $1.80.
Four observations.
The threshold is exactly one third and does not depend on how much you send, which is the entire commercial content of the game in one line.
Below it you lose regardless of amount; above it you gain regardless of amount. There is no optimal stake, only an optimal counterparty.
That is a surprisingly clean result for a literature this contested, and it holds whatever anyone concludes about motives.
And it identifies what the experiment is really asking. Not whether people are generous, but whether they clear a specific arithmetic bar, which is a much more answerable question.
Fairness And Profit Do Not Conflict Here
A consequence of the tripling that is worth drawing out. Ours.
Four observations.
A trustee who divides the tripled pot evenly returns 50 percent of what they received. On the arithmetic above, that pays the trustor $1.50 per dollar sent.
So the split most people would call fair is comfortably profitable for the person who took the risk, and by a wide margin.
Which means the interesting behaviour is not generosity. A trustee returning less than a third is keeping more than the entire surplus their counterparty created, and that is a much narrower thing to look for.
And it reframes what a business reader should watch for. Not whether a counterparty is fair, but whether they are willing to be worse than an even split, which is a far lower bar and easier to observe.
The Meta-Analysis
The synthesis, which we can report only at one remove.
Johnson, N. D., and Mislin, A. A. (2011), Trust games: A meta-analysis, Journal of Economic Psychology, 32(5), 865–889, DOI 10.1016/j.joep.2011.05.007[4].
A citing preprint records: "in an earlier meta-analysis based on metadata from 84 trust game experiments, (Johnson and Mislin, 2011) found that approximately 40% of the variation in trust and 30% of the variation in trustworthiness could be attributed to differences in experimental protocols. In addition, they identified significant associations between socio-cultural variables and trust-related outcomes; for example, greater ethnic diversity was associated with lower levels of trustworthiness."[2]
We did not obtain the meta-analysis, and we could not corroborate the study count of 84 anywhere else.
Forty Percent From Protocol
What that figure means, and it is the most important thing in this article after the arithmetic. Ours.
Four observations.
Forty percent of the variation in measured trust is attributable to how the experiment was run, not to the people in it. That is a very large share for a measurement instrument.
The same source names two protocol features affecting the amount sent: whether payment is random and whether play is with a simulated counterpart[7]. Neither is a fact about trust.
So a substantial part of what fifty years of published trust levels record is the equipment rather than the subject. This series found the same structure in the sixty-second article, where a discount rate absorbed everything that was not time preference.
And note the asymmetry between the two figures. Forty percent for trust and thirty for trustworthiness, so the trustor's behaviour is the more protocol-sensitive of the two, which matters because the trustor is the party a business reader is usually being.
What That Does To Comparison
The practical consequence for anyone reading a trust statistic. Ours.
Three observations.
Cross-country and cross-industry comparisons of trust levels are common, and if two fifths of the variation is protocol, then comparing studies run differently is largely comparing methods.
The meta-analysis's own socio-cultural findings are therefore the ones to weight, since they were computed after the protocol variation was accounted for. We report one of them, on ethnic diversity, without endorsing any interpretation of it, and we did not obtain the analysis.
And the general form is a caution this series keeps arriving at. A number that varies more with the instrument than with the subject is a property of the instrument, and forty percent is a long way toward that.
Betrayal Aversion
The finding that explains why people do not trust as often as the arithmetic says they should.
A citing preprint records: "(Bohnet and Zeckhauser, 2004) introduced the concept of betrayal aversion and showed that individuals require a significantly higher probability of return when trusting other humans than they do in equivalent random lotteries."[2]
The source is Bohnet, I., and Zeckhauser, R. (2004), Trust, risk and betrayal, Journal of Economic Behavior and Organization, 55(4), 467–484, DOI 10.1016/j.jebo.2003.11.004[8]. We did not obtain it.
Four observations, ours.
The comparison is the finding. Identical odds, identical payoffs, and people demand more when a person is on the other side than when a machine is.
That isolates something specific. The reluctance is not risk aversion, because the lottery has the same risk. It is an additional cost attached to being let down by someone who chose.
The same preprint records a related position: "(Fehr, 2009) argued that trust is not merely a form of risk-taking, but is instead grounded in social preferences, such as betrayal aversion."[2]
And a study-aid site summarising the 2004 paper records two results, less sent in the trust case and little sign of altruism[7]. That is a study-aid site, not an academic source, flagged here and at every use.
The Benchmark, Computed
What the odds actually need to be, using a variant whose payoffs are stated. Our own arithmetic, exact on those figures.
A preprint describes the variant: "If the trustor chooses not to trust the trustee, each will receive $10; If the trustor and the trustee both choose to trust, each will receive $15; If the trustor chooses to trust, but the trustee does not, the trustor will receive $8 and the trustee will receive $22." The minimum acceptable probability is "the minimum value of p at which the trustor would" choose to trust[6].
Setting the expected value of trusting equal to the certain ten dollars: 15p plus 8 times (1 minus p) equals 10, which reduces to 8 plus 7p equals 10, giving p equals two sevenths, or 28.6 percent.
Four observations.
A risk-neutral person should trust at any repayment probability above 28.6 percent, which is a low bar and considerably lower than most people's intuition.
The betrayal aversion finding is that people demand more than that from a person and less from a lottery on identical odds, which means the gap is not explained by the arithmetic.
The benchmark is exact given the payoffs as reported, and if the preprint has them wrong then so do we. We did not obtain the 2004 paper to check.
And the structure is worth noticing. Trusting risks two dollars to gain five, which is why the required probability is well below a half.
What Demanding More Costs
Pricing the reluctance. Our own arithmetic, on the same stated payoffs.
Expected value of trusting, by the probability someone insists on before doing it, against the ten dollars they could have had for certain:
At 28.6 percent: $10.00, forgoing nothing. At 35 percent: $10.45, forgoing $0.45. At 45 percent: $11.15, forgoing $1.15. At 55 percent: $11.85, forgoing $1.85. At 65 percent: $12.55, forgoing $2.55. At 75 percent: $13.25, forgoing $3.25.
Three observations.
Someone who requires 65 percent confidence before trusting walks away from $2.55 of expected value on every such decision, which is a quarter of the safe payoff.
The cost rises steeply because each point of required probability discards a whole band of profitable opportunities, which is the same shape the sixty-fifth article computed for ambiguity premiums.
And the figures are per decision. A firm making this choice weekly compounds the forgone value, and nothing in its accounts records it.
Do Survey Questions Predict It?
A question that matters far beyond this experiment, because a single survey item underpins much of the social capital literature.
Glaeser, E. L., Laibson, D. I., Scheinkman, J. A., and Soutter, C. L. (2000), Measuring Trust, Quarterly Journal of Economics[9].
Its abstract: "We combine two experiments and a survey to measure trust and trustworthiness, two key components of social capital. Standard attitudinal survey questions about trust predict trustworthy behavior in our experiments much better than they predict trusting behavior. Trusting behavior in the experiments is predicted by past trusting behavior outside of the experiments. When individuals are closer socially, both trust and trustworthiness rise."[9]
The paper states the consequence plainly: "standard survey questions about trust do not appear to measure trust. However, they do measure trustworthiness, which is one ingredient of social capital. This means that most work using these survey questions needs to be somewhat reinterpreted."[9]
Four observations, ours.
That last sentence is one of the more consequential claims this series has quoted. A large body of research measures trust with a survey question, and this paper says the question measures something else.
The asymmetry is precise and worth holding. Asking whether people can be trusted predicts whether the respondent will repay, not whether they will lend.
The paper also reports that trusting behaviour "is predicted by past trusting behavior outside of the experiments"[9], which is the seventy-first article's conclusion arriving from a different literature: behaviour predicts behaviour, and self-report does not.
And it reports a finding we note without developing: "Trustworthiness declines when partners are of different races or nationalities."[9] We did not obtain the analysis behind it and it is outside this article's scope.
The Reversal
What happened when someone put the original design back, eighteen years later.
Aksoy, B., and colleagues (2018), Measuring trust: A reinvestigation, Southern Economic Journal, DOI 10.1002/soej.12259[5].
Its abstract: "This important study established that the behavior in an incentivized trust game is not correlated with the responses to the most widely used survey measures of trust. We conduct a replication and a reinvestigation of GLSS. In the replication, we use the GLSS protocol and we reproduce their results."[5]
Then the change: "In the reinvestigation, we introduce one major change: We replace their variation of the standard Berg, Dickhaut, and McCabe (1995) investment game with the original unmodified game. The standard game endows both players, while the modified version endows only the first mover. After endowing both movers in the reinvestigation experiment, we find a significant correlation between the two measures, suggesting that trust is a single construct, whether measured by the survey questions or by an incentivized trust game."[5]
Four observations, ours.
This is the cleanest structure a reappraisal can have. The replication reproduced the null. The reinvestigation changed exactly one thing and the null disappeared.
The one thing was whether the second mover started with money of their own. That is a detail nobody would list as important, and it turned a famous null into a significant correlation.
It also fits the earlier sections mechanically. If both players are endowed, returning money is a transfer between two people who each have something, and if only the first is endowed, returning money is closer to charity, which the sixty-third article's literature would predict behaves differently.
And the general lesson is the one the seventy-second article's meta-analysis stated as a finding. Around 40 percent of the variation in this measure comes from protocol, and here is a single protocol choice reversing a headline conclusion.
The 2025 Critique
The most serious objection, and it is conceptual rather than empirical.
A 2025 paper in a business ethics journal argues that "widespread adoption of the Trust Game marks an unnoticed yet fundamental shift in how trust is understood in social science."[3]
Three observations, ours.
The claim is not that the experiments are badly run. It is that adopting a measure changed the concept being measured, which is a claim about a field rather than a study.
That is the same structural worry the fifty-seventh article examined under a different name. A proxy becomes the thing, and here the proxy is an experimental paradigm and the thing is trust.
And the paper is constructive rather than dismissive, concluding with "positive suggestions for empirical researchers who aim to nuance and deepen our understanding of trust and trustworthiness."[3] We did not obtain the suggestions.
The Motive Is Indeterminate
The specific charge, and it lands squarely on the second half of the game.
"In the classic BDM trust game, the second mover's motive for returning money is indeterminate. This means that we are not entitled to interpret the return of money as an indication of trustworthiness."[3]
Four observations, ours.
The argument is simple once stated. A second mover who returns money might be trustworthy, or generous, or averse to unequal splits, or acting on a norm, and the observed number is the same in each case.
That matters because the return figure is what the whole literature calls trustworthiness, and a great deal rests on the label.
It also has a commercial edge that the paper does not draw. A supplier who performs might be reliable, or merely watched, and a firm reading performance as character is making the same inference the critique forbids.
And note that the critique leaves the first mover's number alone. Sending money still measures a willingness to accept exposure, whatever the second mover's motives turn out to be.
Trustworthiness Is A Norm, Trusting Is Not
A related position from the same reference list, which we can name and not evaluate.
The 2025 paper's reference list identifies Bicchieri, C., Xiao, E., and Muldoon, R. (2011), Trustworthiness is a social norm, but trusting is not, Politics, Philosophy and Economics, 10(2), 170–187[3].
We did not obtain it and report the title, which states its thesis.
Four observations, ours.
The asymmetry claimed is the same one the 2000 survey paper found empirically. Trustworthiness behaves like a norm, and trusting does not.
If that is right, the two halves of the game need different treatment. Norms respond to expectations, observation and reputation; a decision to expose yourself responds to incentives and information.
Which would explain why the protocol variance is so large. A norm is sensitive to how a situation is framed, and the frame is exactly what a protocol sets.
And it supplies the commercial prescription this article ends on. You can raise the other party's trustworthiness by changing what is observed and expected, and you cannot raise your own willingness to be exposed the same way.
Eighty-Four Or A Hundred And Thirty-Seven
A discrepancy in the record about the central quantitative source in this article, which we could not resolve. Ours.
One academic preprint describes the meta-analysis as "based on metadata from 84 trust game experiments"[2]. A working paper refers to "the 137 experiments covered by Johnson and Mislin (2011)"[7].
The same working paper supplies the figure we most wanted: the average return across those experiments was 0.37 of the tripled amount[7].
Four observations.
We did not obtain the meta-analysis, so we cannot say which count is right, and both sources are citing the same 2011 paper.
The most likely explanation is benign and we cannot confirm it. A meta-analysis often reports several counts, being studies included, experiments coded and effect sizes extracted, and two citing authors may have picked different ones.
What the discrepancy does establish is a limit on how much weight the 0.37 figure can carry here. It reaches us at one remove, from a source that disagrees with another source about the study it is describing.
And we would rather report that than present the number cleanly. Placed against break-even, 0.37 says trust pays by 3.7 percentage points out of 37, which is the tightest margin in this article and the figure we would least want to be wrong about.
What Actually Survives
Our reading, stated directly.
Five statements.
People send substantial amounts and receive substantial amounts back, against a theoretical prediction of zero, across hundreds of experiments and a meta-analysis.
The result is robust to attempts to break it. A replication modified information presentation and prompted strategic reasoning, and reported that none of its treatments reduced the amount invested.
Around 40 percent of the variation is protocol rather than people, on a citing account of the meta-analysis, which is a very large share for a measurement instrument.
The margin over break-even is thin. At a reported average return of 0.37 against a break-even of one third, trust pays by 3.7 percentage points, and that figure reaches us at one remove from a disputed description.
People treat betrayal as worse than equivalent bad luck, which is the finding with the clearest commercial reading and whose magnitude we could not obtain.
And what the second half of the game measures is contested, with a 2025 paper arguing the return of money cannot be read as trustworthiness at all.
Why The Threshold Does Not Transfer
The most important thing in this article, and it is the reason a business owner should not take the laboratory number away with them. Our own arithmetic, exact.
In the game, money sent is tripled. That is a 200 percent return on the amount at risk, which is why break-even needs only a third back.
A business triples nothing. It earns its margin. Deliver work at cost, bill a price, and let m be the gross margin. Paid with probability p, the expected value is p times price minus cost. Setting that to zero gives p equals one minus m.
The break-even payment probability is one minus your gross margin.
At a 10 percent margin you need to be paid 90 percent of the time. At 20 percent, 80. At 30 percent, 70. At 40 percent, 60. At 50 percent, 50. At 70 percent, 30.
Four observations.
A firm on a 30 percent margin needs a 70 percent payment probability merely to break even, against the laboratory's 33 percent. The gap is entirely the multiplier.
The rule is exact, memorable and requires no research. Your tolerable bad-debt rate equals your gross margin, and most owners have never computed it.
It also explains why low-margin businesses are rightly ruthless about credit. At a 10 percent margin, one unpaid invoice in ten wipes out the profit on the other nine, which is not caution but arithmetic.
And it reframes the whole literature for a commercial reader. The experiments describe a world where trusting has a 200 percent expected return. Your world is not that world, and the behaviour they measure is generous in a setting where generosity is cheap.
One qualification, because the comparison is not quite fair to the experiments. The multiplier exists to create surplus worth arguing over, and without it there would be nothing to measure, so it is a feature of the design rather than a flaw in it.
The error is in the transfer, not in the research. Nobody in that literature claims a third is the threshold for a business, and a reader who takes it away has imported a number from a setting that was built to be generous.
And there is a version of the multiplier that does exist commercially, which is worth naming so the analogy is not dismissed entirely. Where extending credit wins a customer who would otherwise go elsewhere, the value at stake is the relationship rather than the invoice, and the effective multiplier is however many repeat orders follow. That is a real case and it is not the case most owners are in when they wonder whether to chase a late payer.
Extending Credit
The first application. Ours, untested, and not credit advice.
Four points.
Every invoice issued before payment is the first move in this game. You have delivered, the value is with them, and their return is voluntary in the short run.
The threshold to compute is the one above. One minus your gross margin, which for most service firms sits between 50 and 80 percent and which almost nobody states before extending terms.
The replacement arithmetic is the version that changes behaviour. On a 20 percent margin, a ten thousand dollar write-off destroys eight thousand of cost and requires forty thousand dollars of new sales to recover. Our own figures, invented invoice size.
And the betrayal finding predicts where you will misjudge. A default by a client you liked will feel worse than an identical loss from a bad forecast, and that feeling is real and should not be allowed to set the policy for the next client.
Choosing Suppliers
The second application, where you are the exposed party in a different direction. Ours.
Four points.
Paying a deposit, ordering ahead, or committing to a sole supplier is the same structure. You have moved first and their performance is voluntary until it is not.
The 2025 critique applies directly here and is uncomfortable. A supplier who has performed might be reliable or might have been watched, and the observed record cannot distinguish them.
Which suggests a specific test that costs nothing. Has this supplier performed when nobody was checking? If every past instance was monitored, you have no evidence about character, only about your own supervision.
And the protocol finding gives the practical lever. If 40 percent of the variation in this behaviour comes from how the situation is set up, then the setup is where a firm's effort belongs, rather than in judging people.
Delegation
The third application, and the one owners find hardest. Ours.
Four points.
Handing work to someone else is a first move with a payoff structure most owners never write down. You give up control, the value is created by them, and you receive what comes back.
The betrayal finding predicts systematic under-delegation for a specific reason. Being let down by a person is worse than an equivalent loss from circumstance, so an owner comparing delegation against doing it themselves is not comparing like with like.
The asymmetry is worth naming, because it is invisible. A job done badly by you is a bad week. The same job done badly by someone you chose is a betrayal, and the arithmetic does not distinguish them.
And the norm finding suggests the productive response. Trustworthiness responds to expectations and observation, so the useful question is not whether this person is reliable but what they will believe is expected.
Promises
A finding that gives the cheapest intervention in this article, reported at the strength our sourcing allows.
A citing preprint records: "(Charness and Dufwenberg, 2006) examined the influence of non-binding promises in the trust game and found that making a promise, even in the absence of enforcement mechanisms, can significantly affect player behavior."[2]
We did not obtain the paper and report this characterisation.
Four observations, ours.
The operative words are "non-binding" and "even in the absence of enforcement mechanisms." This is not a contract; it is a statement.
If it holds, it is the cheapest instrument available. Asking someone to state what they will do costs nothing and creates no obligation anyone could enforce.
And it fits the norm account precisely. A promise makes an expectation explicit, and if trustworthiness is a norm responding to expectations, making the expectation explicit is exactly the lever.
We would flag the obvious limit. A promise that changes behaviour also changes who says yes, and someone unwilling to state an intention has told you something before any money moves.
The second limit is that we are recommending an instrument on one characterisation of one paper we did not read, which is thinner sourcing than we would normally act on. We include it because the cost of trying it is nothing and the cost of being wrong is nothing either, which is a standard we would not apply to anything expensive.
What To Do
Compute your break-even payment probability as one minus your gross margin. A firm on a 30 percent margin needs to be paid 70 percent of the time to break even, against a laboratory figure of 33.
Convert a write-off into replacement sales before deciding your credit policy. On a 20 percent margin, a ten thousand dollar loss requires forty thousand dollars of new sales to recover.
Do not import the laboratory threshold. The experiments triple the money at risk, which is a 200 percent return, and your business does not.
Ask whether a supplier performed when nobody was checking. A 2025 critique argues the return of value cannot be read as trustworthiness, because the motive is indeterminate.
Ask for a stated intention rather than relying on judgment. A non-binding promise reportedly affects behaviour without any enforcement mechanism, which makes it the cheapest instrument here.
Work on the setup rather than on judging people. Around 40 percent of the variation in this behaviour is attributable to protocol, on a citing account of the meta-analysis.
Expect to under-delegate and know why. Being let down by a person is treated as worse than equivalent bad luck, so your comparison between doing it yourself and delegating is not like for like.
Do not use a survey question to measure trust. The 2000 paper reports standard attitudinal questions predicting trustworthy behaviour much better than trusting behaviour, and says most work using them needs reinterpreting.
The Limits Of This Analysis
Several caveats matter. This article discusses research on trust and is not credit, procurement or contracting advice; the applications are our own reasoning and untested. Everything is verified to August 2026. We did not obtain the 1995 paper, and its design and sample sizes come from a paper reconstructing it rather than from the original. We did not obtain the meta-analysis, and every figure attributed to it here reaches this article through papers citing it; two of those sources disagree about how many experiments it pooled, as the body records. We did not obtain the betrayal aversion paper and report its finding through two citing descriptions without any magnitude, which is the most important missing number in this article. We did not obtain the 2018 reinvestigation, the 2000 survey paper beyond its abstract and one passage, the 2025 critique beyond its abstract, or any of the works named from reference lists, including the promises paper whose finding supplies our cheapest recommendation. One source is a study-aid site, flagged at every use. All arithmetic is ours. The break-even threshold and the replacement-sales figures are exact given the definitions stated, and the invoice size and margins are invented. The minimum acceptable probability computed earlier is exact given payoffs reported in a preprint we could not check against the original. And the whole literature concerns one-shot anonymous encounters between strangers, which is close to the opposite of a commercial relationship with a repeat client, a reputation and a contract; where those exist, the experiments describe a floor rather than a forecast.
Frequently Asked Questions
What is the trust game?
What return rate makes trusting worthwhile?
What is the equivalent threshold for my firm?
What is betrayal aversion?
Is the measure sound?
Can I just ask people how trusting they are?
What is the cheapest thing I can do?
References
- Academic preprint on simulating research participants, reconstructing the 1995 study and recording its design: that Berg and colleagues studied trust and reciprocity in a two-stage investment game with University of Minnesota undergraduates, N equal to 120, split into a No-History condition of 64 participants in 32 pairs and a Social-History condition of 56 participants in 28 pairs; that Room A chose how much of a ten dollar endowment to send, from zero to ten dollars; that the amount was tripled; that Room B then decided how much to return; and that the social history condition consisted of a report summarising the prior 32 pairs' outcomes. Note: an academic preprint reconstructing the study, not the study itself. Our only source for the design and sample sizes; we did not obtain the 1995 paper. arxiv.org
- Academic preprint on trust-building games, surveying the literature and recording: that the classical trust game was first introduced by Berg, Dickhaut and colleagues in 1995 as the investment game and has been widely adopted as a foundational framework for analysing trust behaviour under uncertainty; that in an earlier meta-analysis based on metadata from 84 trust game experiments, Johnson and Mislin found approximately 40 percent of the variation in trust and 30 percent of the variation in trustworthiness attributable to differences in experimental protocols; that they identified significant associations between socio-cultural variables and trust outcomes, with greater ethnic diversity associated with lower trustworthiness; that Bohnet and Zeckhauser introduced the concept of betrayal aversion and showed that individuals require a significantly higher probability of return when trusting other humans than they do in equivalent random lotteries; that Fehr argued in 2009 that trust is not merely a form of risk-taking but is grounded in social preferences such as betrayal aversion; that Charness and Dufwenberg examined the influence of non-binding promises in the trust game and found that making a promise, even in the absence of enforcement mechanisms, can significantly affect player behaviour; and that Cochard and colleagues reported both trust-giving and return rates increasing over time in repeated versions, with a sudden drop toward the end of finitely repeated games. Note: an academic preprint surveying the literature. Our source for the meta-analytic figures, betrayal aversion and the promises finding, none of which we obtained in the original; note that this source gives the meta-analysis as covering 84 experiments where another source gives 137. arxiv.org
- Publisher record for a 2025 paper in the Journal of Business Ethics titled What Does the Trust Game Measure?, reproducing its abstract: on the ubiquitous Trust Game of Berg, Dickhaut and McCabe being the dominant experimental paradigm for the study of trust and trustworthiness in experimental economics, social psychology and neuroscience of the last two decades; on this elegant framework, promising a quantifiable behavioural measure, being at the centre of a strikingly broad range of research programmes; on the paper bringing Trust Game research into conversation with literature in the philosophy of trust to show that widespread adoption of the Trust Game marks an unnoticed yet fundamental shift in how trust is understood in social science; on the second mover's motive for returning money being indeterminate in the classic game, meaning we are not entitled to interpret the return of money as an indication of trustworthiness; and on the paper concluding with positive suggestions for empirical researchers. The same record carries a reference list confirming Bicchieri, C., Xiao, E., and Muldoon, R. (2011), Trustworthiness is a social norm, but trusting is not, Politics, Philosophy and Economics, 10(2), 170–187; Bohnet, I., and Zeckhauser, R. (2004), Journal of Economic Behavior and Organization, 55(4), 467–484; and Burnham, T., McCabe, K., and Smith, V. L. (2000), Journal of Economic Behavior and Organization, 43(1), 57–73. Note: the publisher's record. We obtained the abstract only, and not the positive suggestions the paper concludes with, which would be the most useful part. link.springer.com
- Five independent reference lists confirming Berg, J., Dickhaut, J., and McCabe, K. (1995), Trust, reciprocity, and social history, Games and Economic Behavior, 10(1), 122–142, July, DOI 10.1006/game.1995.1027; and Johnson, N. D., and Mislin, A. A. (2011), Trust games: A meta-analysis, Journal of Economic Psychology, 32(5), 865–889, DOI 10.1016/j.joep.2011.05.007. The same lists variously confirm Cox, J. C. (2004), How to identify trust and reciprocity, Games and Economic Behavior, 46, 260–281; Houser, D., Schunk, D., and Winter, J. (2010), Distinguishing trust from risk: an anatomy of the investment game, Journal of Economic Behavior and Organization; and Schechter, L. (2007), Traditional trust measurement and the risk confound: An experiment in rural Paraguay, Journal of Economic Behavior and Organization, 62, 272–292. Note: reference lists used to confirm citations, volumes, pages and DOIs independently. We obtained none of the works named. pmc.ncbi.nlm.nih.gov
- Publisher record for Aksoy and colleagues (2018), Measuring trust: A reinvestigation, Southern Economic Journal, DOI 10.1002/soej.12259, reproducing its abstract: on the authors reinvestigating the question first posed by Glaeser and colleagues in 2000, whether survey measures about trust predict actual trusting behaviour; on that important study having established that behaviour in an incentivized trust game is not correlated with responses to the most widely used survey measures of trust; on the authors conducting a replication using the original protocol and reproducing those results; on the reinvestigation introducing one major change, replacing the modified game with the original unmodified Berg, Dickhaut and McCabe investment game, the standard version of which endows both players where the modified version endows only the first mover; and on the authors finding, after endowing both movers, a significant correlation between the two measures, suggesting that trust is a single construct whether measured by survey questions or by an incentivized trust game. Note: the publisher's record. We obtained the abstract only and not the paper, so we report no correlation coefficients or sample sizes. onlinelibrary.wiley.com
- Academic preprint describing a binary trust game variant with stated payoffs: that if the trustor chooses not to trust the trustee, each receives ten dollars; that if both choose to trust, each receives fifteen dollars; that if the trustor trusts but the trustee does not, the trustor receives eight dollars and the trustee twenty-two; and defining the minimum acceptable probability as the minimum value of the repayment probability at which the trustor would choose to trust. Note: an academic preprint describing a variant. Our benchmark calculation is exact given these payoffs and we could not check them against any original source. arxiv.org
- Working paper on disentangling trust from risk-taking, recording that on average the trustees in its own experiment returned 0.96 of the transfer, representing 0.32 of the tripled amount, and that the average return in the 137 experiments covered by Johnson and Mislin (2011) was 0.37; and recording that about one third of its subjects chose the low return level in all four games while only one sixth always chose the high return level. Note: a working paper citing the meta-analysis. Our source for the 0.37 average return; note that this source gives the meta-analysis as covering 137 experiments where reference 2 gives 84. mpra.ub.uni-muenchen.de
- Publisher reference list confirming Bohnet, I., and Zeckhauser, R. (2004), Trust, risk and betrayal, Journal of Economic Behavior and Organization, 55(4), 467–484, DOI 10.1016/j.jebo.2003.11.004; alongside a separate journal reference list giving the same paper's page range as 467–485. Note: reference lists used to confirm the citation; recorded also as the source of a one-page discrepancy in the reported page range. journals.plos.org
- Copy of Glaeser, E. L., Laibson, D. I., Scheinkman, J. A., and Soutter, C. L. (2000), Measuring Trust, Quarterly Journal of Economics, hosted on the second author's own university page, together with a repository record reproducing its abstract: on the authors combining two experiments and a survey to measure trust and trustworthiness, two key components of social capital; on standard attitudinal survey questions about trust predicting trustworthy behaviour in the experiments much better than they predict trusting behaviour; on trusting behaviour in the experiments being predicted by past trusting behaviour outside the experiments; on both trust and trustworthiness rising when individuals are closer socially; on trustworthiness declining when partners are of different races or nationalities; on high status individuals eliciting more trustworthiness in others; and on the paper's own statement that standard survey questions about trust do not appear to measure trust but do measure trustworthiness, which is one ingredient of social capital, meaning most work using these survey questions needs to be somewhat reinterpreted. Note: a copy hosted by a co-author's own university page plus a repository record of the abstract. We obtained the abstract and selected passages and not the analyses. scholar.harvard.edu
This article discusses research on trust and is not credit, procurement or contracting advice. None of the underlying papers was obtained in full. The 1995 study's design comes from a paper reconstructing it; every meta-analytic figure reaches this article through citing papers, two of which disagree about how many experiments were pooled; the betrayal aversion magnitude could not be obtained. One source is a study-aid site, flagged at every use. All arithmetic is the authors' own: the break-even threshold and replacement-sales figures are exact given the definitions stated, with invented invoice sizes and margins. The literature concerns one-shot anonymous encounters between strangers, which is close to the opposite of a commercial relationship with a repeat client, a reputation and a contract.