Perpetrator knowledge: a Bayesian account

Abstract

Perpetrator knowledge (also known as “guilty knowledge,” “insider knowledge,” “crime knowledge,” or “first-hand knowledge”) is an important, but undertheorized type of criminal evidence. This article clarifies this concept in several ways. First, it offers a precise, probabilistic definition of what perpetrator knowledge is. Second, the article provides a taxonomy of arguments relating to perpetrator knowledge. This classification is based on an analysis of 438 Dutch criminal cases in which this concept was mentioned. Third, it models these arguments using Bayesian networks. Fourth, the article explains a potential reasoning error relating to perpetrator knowledge, namely the fallacy of appeal to probability.

1. Introduction

As it is typically defined, perpetrator knowledge is “knowledge that only the (co-)perpetrator of a crime could have.”¹ Other terms that are used for this concept are “guilty knowledge,” “insider knowledge,” “crime knowledge,” or “first-hand knowledge.” Perpetrator knowledge is frequently crucial evidence in criminal cases. First, suppose that a suspect confesses to committing a crime. As we now know, not all confessions of crimes are truthful. One commonly suggested protection against false confessions is ensuring that the suspect also describes verifiable perpetrator knowledge in their confession.² Perpetrator knowledge can also be strong evidence in cases where a suspect denies that they committed a crime. We may, for instance, learn from messages that they sent to others that they knew intimate details about the crime, thereby showing that their denial is a lie. Police interrogators therefore often seek to trip up suspects so that they accidentally reveal that they have perpetrator knowledge (St-Yves and Meissner, 2014).

Despite its importance, the concept of perpetrator knowledge has, so far, not been explained with sufficient clarity. The goal of this article is to elucidate this concept and its role in criminal cases.³ More precisely, this article makes four contributions to the academic debate on perpetrator knowledge. First, it gives a (probabilistic) definition of the concept. Second, it develops a taxonomy of arguments relating to perpetrator knowledge to show how the concept is used in practice. Third, the article models these arguments using Bayesian networks in order to elucidate their structure.⁴ Finally, it explains a potential reasoning error relating to perpetrator knowledge, namely the fallacy of appeal to probability.⁵

To begin with the first aim, as I will argue, the above conception—which defines it as “knowledge that only the perpetrator could have”—is flawed. I develop a more nuanced, probabilistic account of perpetrator knowledge, employing Bayesian probability theory (Section 2). On this account, perpetrator knowledge, just like other evidence, can have varying levels of probative force. I also argue that in order to adequately assess this likelihood ratio, we need to incorporate insights from explanation-based accounts of rational criminal proof.⁶

To compile a taxonomy of arguments used in criminal legal practice, I searched⁷ for all Dutch criminal cases mentioning the term “perpetrator knowledge” and its synonyms⁸ on www.rechtspraak.nl, the official repository of judicial rulings in the Netherlands.⁹ After removing all duplicate and irrelevant cases, 438 rulings remained.¹⁰ I then created a systematic classification of the types of arguments that were used by the court, prosecution, and defense in these cases. More precisely, a codebook was created based on an in-depth analysis of a small selection of the rulings (50). This codebook was then used to code the selected rulings, in order to classify how perpetrator knowledge was used as evidence and/or how it was established that a party displayed such knowledge. The aim of this coding process was not to get quantitative insight into the analyzed case law,¹¹ but to provide an exhaustive overview of the different arguments relating to perpetrator knowledge that can be found. In particular, I distinguish two main categories of arguments. First, there are arguments that draw a conclusion from a person (not) displaying perpetrator knowledge (e.g. that their confession is therefore reliable). I discuss these arguments in Section 3. In Section 4, I turn to arguments that relate to whether or not a person has perpetrator knowledge. I provide one or more examples of each argument. Finally, in Section 5, I discuss the appeal to probability fallacy.

2. A probabilistic account of perpetrator knowledge

The typical definition of perpetrator knowledge is “knowledge that only the perpetrator of a crime could have.” This is serviceable enough as a rough approximation of the concept. However, as a definition, it is problematic. In particular, it treats perpetrator knowledge as an all-or-nothing issue: something is perpetrator knowledge if only the perpetrator could have known about it. If there is a possibility that a non-perpetrator knew about it, then it cannot be perpetrator knowledge and therefore (presumably) has no probative value. But we can never be truly certain that something is perpetrator knowledge. There is always a remote possibility that someone successfully guessed correctly about some fact. Conversely, suppose that, in a given case, we conclude that someone does display perpetrator knowledge. What does this prove? Taking the above definition literally, this person must be the perpetrator. After all, he knows something that only the perpetrator of the crime could know. But this seems too strong a conclusion. If that is the case, then no additional evidence, no matter how strong, should be able to overturn our belief that the defendant is the perpetrator of the crime.

The upshot of the above is that we should not regard perpetrator knowledge as an all-or-nothing concept. Instead, we want a definition on which we can be more or less certain that something is perpetrator knowledge, and on which apparent perpetrator knowledge can have more or less probative force, depending on this level of certainty. To put it differently, we want a probabilistic interpretation. The most straightforward way to express such an interpretation is by means of the Bayesian likelihood ratio, a well-known definition of probative force¹²:

P(E | H)

————

P(E | ¬H)

This ratio determines how receiving some item of evidence should impact our prior beliefs about a given hypothesis (e.g. that the suspect is the perpetrator of the alleged crime). P(E|H) refers to the probability that we would observe the evidence if we suppose that the hypothesis is true. P(E|¬H) refers to the converse: what is the probability that we would observe this evidence if the hypothesis were not true? The higher the value of the numerator and the lower that of the denominator, the more the given item of evidence should raise our belief in the hypothesis. If the two values are equal, because the evidence is as expected under the hypothesis as under its negation, the evidence has no probative force.¹³ If the ratio is lower than 1, because the evidence is more expected if the hypothesis is false than if it is true, then the evidence should lower our belief in the truth of the hypothesis.

The hypothesis H that we consider in the case of perpetrator knowledge is usually “the suspect is the perpetrator of the alleged criminal offense.”¹⁴ The evidence E for this claim is that “the suspect knows X,” where X refers to some set of verified details about the crime (e.g. the murder weapon used).¹⁵ So, the likelihood ratio expresses the probability that the suspect would know X given that they are the perpetrator versus the probability that they would know X given that they are not the perpetrator:

P(SUSPECT KNOWS | SUSPECT PERP)

—————————————————————————-

P(SUSPECT KNOWS | ¬ SUSPECT PERP)

Whether the suspect knows that X has to be shown using further evidence. Once it has been established that he does know X, the belief that the suspect is the perpetrator should be updated based on this ratio. However, one worry is how we determine the value of these conditional probabilities. This is difficult when we are concerned with non-statistical evidence such as witness testimony (Jellema, 2022). It is then not clear what these numbers express or where we should get them from. I want to suggest that we can determine the value of the likelihood ratio by considering a select number of sufficiently detailed explanations (or “scenarios”) of why this person offers this particular statement in this context. For the numerator, this means coming up with a scenario in which the suspect is the perpetrator and knows these details X. In the case of the denominator, we have to come up with one or more scenarios in which the suspect would (appear to) know these details despite not being the perpetrator. Such explanation-based thinking makes it easier to assess the relevant probabilities, as it helps us determine what evidence is relevant for assessing these probabilities and what the strength of this evidence is. Let me illustrate this point with an example:

In a murder case, the victim’s debit card was used to withdraw money after they were killed.¹⁶ The police wiretapped the suspect’s phone (the suspect’s mother would later testify that they were aware that their phone conversations were probably being monitored). During a call with his parents, the suspect talked about how the media reported on this money withdrawal. At one point he mentioned a sum of €350,- According to the prosecution, this was perpetrator knowledge, as this was the actual amount withdrawn and this sum was not mentioned in the media. However, the court rejected this argument, noting that the phone conversation took place 1.5 years after the events. Both the victim’s sister and mother claimed that police investigators had told them by then how much money was withdrawn, and that they spoke about this with various people. Rumors about this sum then started circulating in the village where both they and the suspect lived. However, the court did note that after the suspect mentioned the €350,- he quickly started to bring up different sums, and gave the impression that he was trying to hide a slip-up. The court argued that an innocent person who knew about this sum by accident would not have acted as if they just made a mistake. The suspect was ultimately convicted in part upon the basis of this observation.

In this case, the court considers two scenarios. The first is that the suspect was the perpetrator. In that case, he would have known how much money was withdrawn. The defense’s argument was that the probability was much higher that he would have mentioned this sum if he was the perpetrator, than if he was innocent. The court noted that an innocent person could also have known this sum, so that the likelihood ratio of this knowledge was (close to) neutral. However, given the nature of the suspect’s behavior, it was unlikely that an innocent person would have reported on this sum in this way (and it was likely that a guilty person would act this way). Hence, the suspect did display perpetrator knowledge. This conclusion could only be drawn based on a careful analysis of other ways in which the suspect could have learned about the sum and about what other facts we would (not) expect to see if this alternative explanation were true.

As we will see in this article, there are several types of arguments relating to perpetrator knowledge, each with their own structure. Furthermore, these arguments can quickly become complex. While all of them lead to conclusions that can be expressed in terms of the above likelihood ratio, this ratio does leave out many inferential steps and subtleties. A better way of formalizing these arguments is therefore by using Bayesian networks. A Bayesian network is a graphical representation of the probabilities and probabilistic connections between a set of propositions (for further explanation, see Nielsen and Jensen, 2009; Fenton, Neil and Lagnado, 2013). We can express the above likelihood ratio in a Bayesian network as shown in Fig. 1.

Here we have two nodes, SUSPECT PERP and SUSPECT KNOWS, which represent the two hypotheses. The arrow represents a probabilistic dependency between these two hypotheses. The node SUSPECT PERP is a so-called parent node and SUSPECT KNOWS is a child node because the arrow goes from SUSPECT PERP to SUSPECT KNOWS. The direction of the arrow can be read in causal terms—the suspect knows because they were the perpetrator. The table next to each node—the node probability table—gives the conditional probabilities for each child node given its parents. Parent nodes only have a prior probability. The direction of the arrow therefore ensures that we are specifying the conditional probabilities of SUSPECT KNOWS given different values of SUSPECT PERP, which is what we are after, as this model is intended to capture the likelihood ratio mentioned above. The probabilities from this likelihood ratio are on the left-hand side of the table next to SUSPECT KNOWS. Once we have established that the suspect knows that X, we can set the probability of SUSPECT KNOWS to 1¹⁷ and calculate the new value of SUSPECT PERP based on this ratio.¹⁸

Note that the conditional and prior probabilities that I used in this example are merely placeholders, intended to show how a Bayesian network works. In the next two sections, I model the structure of several arguments using Bayesian networks. However, I do not provide the node probability tables. Instead, I follow Wellman (1990) and only indicate whether the probabilistic link between hypotheses is positive, in the sense that they correlate positively, or negatively (denoted with a “+” or a “–” symbol, respectively).

3. Arguments from perpetrator knowledge

In this section, I distinguish six types of arguments relating to perpetrator knowledge. The first three involve drawing a conclusion based on a person displaying perpetrator knowledge. The other three arguments start from the opposite assumption—that someone does not display perpetrator knowledge.

3.1 Arguments from the presence of perpetrator knowledge

3.1.1 Suspect reveals perpetrator knowledge

This argument is by far the most common one of those discussed in this article. It was used in 280 of 438 analyzed cases. According to this argument, the suspect is the perpetrator of the alleged crime because they have perpetrator knowledge. It is typically used either in an attempt to establish the reliability of a suspect’s confession or against a witness who denies being the perpetrator of a crime, to show that they are lying. This argument is usually raised by the prosecution,¹⁹ which then typically points out that the suspect appears to know about specific, verified details of a crime that were not known to the general public or even facts that were not known to the police at that time.²⁰ In the previous section, I modeled perpetrator knowledge using two hypotheses: SUSPECT KNOWS and SUSPECT PERP. However, that model left out the fact that we cannot look inside the mind of the perpetrator. We can only know what the suspect knows on the basis of observable evidence such as a verbal or written statement. I will call this evidence a “report” (SUSPECT REP). We can include it in the Bayesian network (Fig. 2).

As said, a “+” denotes a positive correlation between two nodes. For instance, as the probability of SUSPECT KNOWS goes up, so does the probability of SUSPECT PERP (and vice versa).

While the suspect’s report is typically a verbal or written statement by the suspect, such as a confession during an interrogation or a text message to fellow suspects. However, the “report” can also be something else. To give an example, in an arson case from 2018, the suspect was accused of ordering a sports center to be set on fire.²¹ Part of the evidence was the suspect’s search history. Shortly before the building caught fire, he searched for the center’s name several times. Then, right after the fire started, he searched 24 times in a span of 25 min for the building’s name on websites that track emergency service notifications. The court concluded that he knew that the sports center would be set on fire and it convicted him in part based on this evidence.

In the above case, the suspect offered an alternative explanation for his searches. First, he claimed that he looked up for the building’s name before the fire because there was a chopper show going on there and he wanted to know who was attending. However, this show did not require prior registration and hence the information would not have been on the website. He also said that he often searched for emergency service notifications and that in this case, he did so because he received a text message about the fire. However, he did not specify from whom he received this text message, nor did he explain why he searched so often in such a short period. Hence the court concluded that these alternative explanations were implausible.

What these arguments illustrate is that there can be uncertainty regarding the link between REP and KNOWS. In this case, the suspect challenged the claim that based on his search history, we can conclude that he probably knew that the building would be set on fire. In other cases, the defense might challenge the link between KNOWS and PERP. For example, in a murder case, the suspect’s phone was tapped. During one conversation the suspect mentioned that the victim was Polish and that this victim had been thrown into the water in a tent. The court accepted that this counted as knowledge of the crime, but acquitted the suspect because this knowledge did not show that he was involved in the man’s death.²²

3.1.2 Indirect perpetrator knowledge

The argument that a suspect displays perpetrator knowledge is not always based on a report by the suspect themselves. It can also be based on the claim that someone else displayed perpetrator knowledge, which they got from the suspect. The structure of the argument can be modeled as shown in Fig. 3.

Each link represents a source of uncertainty that can sometimes undermine the argument. Compared to the previous argument, this structure adds one such source: the link between OTHER KNOWS and SUSPECT SOURCE. Even if the other person knows about details of the crime, then that does not have to be because they learned these from the suspect. There can be alternative explanations, such as that the other person is a co-perpetrator. To give one example, in a case where the suspect was accused of a deathly stabbing, their father testified that he heard his son come home on the night of the incident.²³ The father said that his son told him about the incident. One of the things he told his father was that the victim was carrying nun-chucks and that he wrestled with the victim during which the latter’s shirt was pulled over his head and that there was a witness at the scene. The court deemed this to be perpetrator knowledge.

3.1.3 A different perpetrator

According to this argument, someone other than the suspect displays perpetrator knowledge, which in turn means that they, rather than the suspect, are the perpetrator. We can model this argument as in Fig. 4.

One source of uncertainty stems from the connection between OTHER KNOWS and OTHER PERP. As we saw in the previous subsection, not everyone who has perpetrator knowledge is a perpetrator. The other person could, for instance, have learned this information from the suspect. Another possible source of uncertainty is the connection between OTHER PERP and SUSPECT PERP. The former negatively correlated with the latter. However, there can be cases in which both hypotheses are true; the two can be co-perpetrators.²⁴

One example of this argument occurred in a burglary case where someone called the police with an anonymous tip.²⁵ The caller used a phone that was close to a cellphone tower near the crime scene when the burglary took place. During this phone call, the unknown person displayed perpetrator knowledge. The court concluded that while the evidence against the two suspects in this 3.2Because there had been only two burglars, it could not established whether either of the suspects were part of this duo and therefore whether they perpetrated the crime. Hence, the court acquitted them.

This argument does not always rest on the assumption that the witness with perpetrator knowledge was the actual perpetrator. For instance, sometimes the defense simply argues that this witness is unreliable because they have perpetrator knowledge.²⁶ This leaves open the possibility that the witness was not involved in the crime but does know who the real perpetrator is. Another example of this comes from a rape case.²⁷ The suspect claimed that the victim falsely accused him. The defense argued that her statements were unreliable because they displayed perpetrator knowledge. In this case, the argument was not that someone other than the suspect was the perpetrator. Rather the claim was that there had not been a crime in the first place and that the supposed victim knew information that normally only the perpetrator of rape would know because they made the story up. However, the court rejected this claim. It concluded that her statement contained only information that a victim of a rape could know.

3.2 Arguments from the absence of perpetrator knowledge

The following arguments draw a conclusion based on the fact that a person does not have perpetrator knowledge.

3.2.1 No perpetrator knowledge, no perpetrator

This argument was raised in 32 rulings. According to this argument, the defendant is innocent because they do not possess perpetrator knowledge. In some cases, the defense brings up this line of reasoning as a counterargument to the argument discussed in Section 3.1.1. In other words, it is a response to the prosecution’s claim that the defendant is guilty because they have perpetrator knowledge. However, sometimes a lack of perpetrator knowledge on the suspect’s part is used as an argument in favor of acquittal by itself, without the prosecution making such a claim. It is on this latter type of cases that I focus here.

As I understand the argument, the fact that a suspect’s reports do not display perpetrator knowledge is taken to mean that they do not have such knowledge, because they were not the perpetrator. We can model the structure of this argument as shown in Fig. 5.

The structure of this argument is the same as the argument from Section 3.1.1, except that the link between SUSPECT REP and SUSPECT KNOWS is now negative. After we receive the suspect’s report(s), our belief that the suspect knows details of the crime goes down and because of this, so does our belief that the suspect is the perpetrator.

One potential worry is that this structure resembles a well-known fallacy called the “argument from ignorance.” As this worry is often phrased, absence of evidence is not evidence of absence. Ofshe and Leo (1996) apply this thought to perpetrator knowledge:

While a guilty suspect can prove his accurate knowledge of the crime, an innocent person cannot prove his ignorance. Because a suspect's post-admission narrative can never prove that he does not possess certain information. (…) Since a failure to demonstrate something is not proof of its non-existence, an evaluator of the statement cannot conclude that the absence of this information proves an absence of personal knowledge. The failure to demonstrate actual knowledge can only lead to the conclusion that the admission is suspect, and that both it and the narrative are of little or no value as evidence of guilt. (Ofshe and Leo, 1996: 994–997)

Ofshe and Leo correctly note that a failure to demonstrate that the suspect has perpetrator knowledge can be a reason to discount a confession. Indeed, most occurrences of this argument have this aim. Let me give an example. In a case from 2017, the suspect confessed to robbing several people and killing one person.²⁸ However, the defense argued that he was having a psychotic episode after he was arrested and that anything he said during that time was therefore unreliable. They backed this up with the argument that he did not share any verifiable details that would count as perpetrator knowledge, as all the information in his statements had also been in the media.

While this argument is typically used only to discount testimony, rather than as positive evidence of the absence of perpetrator knowledge, the latter is not as impossible as Ofshe and Leo make it out to be. As Bayesian probability theorists have pointed out, the argument from ignorance is not always a fallacy (Hahn and Oaksford, 2006; Taroni et al., 2019). For example, suppose that scientists conduct thorough tests to determine whether a drug has toxic effects. If no evidence of toxicity is found after repeated tests, then it seems reasonable to conclude that the drug is not toxic. Whether such an argument is a fallacy or not depends on how much positive evidence we would expect to find if the hypothesis were true and how thoroughly we have looked for it. As Hahn and Oaksford (2006) argue negative arguments are generally weaker than positive ones, but evidence is all the same. In some cases, the absence of perpetrator knowledge is (seemingly) used as such positive evidence of absence. To give an example, in one case the suspect was accused of producing and/or selling drugs.²⁹ The key evidence against him was a bag filled with cocaine with the suspect’s DNA on it. The defense argued that this DNA only showed that he touched the bags at some point, but not that he handled the drugs. They supported this scenario with the claim that the suspect did not display perpetrator knowledge in a recorded phone conversation with his brother.³⁰ I read the defense’s line of reasoning as a positive argument in favor of his innocence.

3.2.2 Lack of indirect perpetrator knowledge

On this argument, we infer that the suspect is not the perpetrator (or at least that evidence for his guilt is unreliable) because a person other than the suspect does not display perpetrator knowledge. It therefore combines the argument from the previous section with that of indirect perpetrator knowledge. This argument is sometimes used in response to a claim by the prosecution that this person does have perpetrator knowledge. However, here I focus on situations where no such prior claim was made. The structure of this argument is as in Fig. 6.

The structure of this argument is the same as the argument from Section 3.1.2, except that the link between SUSPECT REP and SUSPECT KNOWS is now negative. This means that when we receive the other person’s report, containing no perpetrator knowledge, the probability of OTHER KNOWS goes down, which propagates through the network, so that ultimately the probability of SUSPECT PERP also decreases.

Just like the argument from the previous subsection, this argument is typically used to argue that some piece of testimony lacks probative value. For instance, consider a murder case in which a witness claimed that the suspect told him that he was involved in the victim’s death.³¹ The defense pointed out that the police promised a reward of €15,000—for the tip that would lead to solving the case. The defense attorney argued that this should make us doubt the witness’s statement. After all, he might be trying to frame the suspect in order to receive this reward. The defense added to this that the witness’s testimony did not contain perpetrator knowledge as he could have read details of the crime in the case file. The court rejects this argument, noting that while the witness could have derived most of the information in his statement from the file, he also discussed details that he could not have obtained from it. For instance, he correctly reported that the suspect had been observing the victim’s apartment for a while before the murder.

3.2.3 Innocence knowledge

The term “innocence knowledge”³² was introduced by Israels and Rassin (2014). This term refers to knowledge that an innocent person is much more likely to have than a guilty person. Take the likelihood ratio from Section 2 again:

P(SUSPECT KNOWS | SUSPECT PERP)

—————————————————————————-

P(SUSPECT KNOWS | ¬ SUSPECT PERP)

Innocence knowledge occurs when this ratio is lower than 1. This means that if we learn that a suspect had such knowledge, then this should lower our belief that they were the perpetrator.³³ We can also capture this idea in a variation on the Bayesian network from Section 3.1.1 (Fig. 7).

The negative link expresses the idea that the more confident we are that the suspect has this innocent knowledge, the less confident we become that they are the perpetrator.

I was not able to find examples of the phrase “innocence knowledge” used in evidential reasoning. To the extent that it is mentioned in Dutch criminal proceedings, it is as a reference to the article by Israels and Rassin (2014), noting the theoretical possibility of such knowledge. However, Israels and Rassin also provide several examples themselves. For instance, they discuss the famous Deventer murder case, in which the suspect argued that he could not have committed the murder as he was stuck in a traffic jam some distance from the crime scene at that moment. The police confirmed that there was indeed a traffic jam at that time and that it was at the location where the suspect claimed it was. This could be an example of innocent knowledge, as the suspect would likely know about this traffic jam if he was indeed innocent, whereas it is arguably much less likely that he would have known about this if he was guilty.³⁴ More generally, we can think of innocence knowledge as any information that the suspect would likely know if their alibi was true, but would be unlikely to know if their alibi was false.

According to Nieuwkamp, Horselenberg and van Koppen (2022), the evidential strength of both innocence and guilt knowledge should depend on how difficult it would be to obtain from an alternative source. Diederik Aben,³⁵ the Procurator General of the Supreme Court, refers to the article by Israels and Rassin (2014) and notes that as with perpetrator knowledge, innocence knowledge will tend to be weaker evidence when it is provided at a late stage in the investigation, since the suspect could base their alibi on their knowledge of the results of the investigation.³⁶ This is an example of providing an alternative explanation for the knowledge, which I will discuss in more depth in Section 4.3.³⁷

4. Arguments about perpetrator knowledge

The six arguments discussed in the previous section all involved drawing a conclusion from the hypothesis that someone did or did not display perpetrator knowledge. In this section, I discuss arguments that aim to establish whether someone has or does not have perpetrator knowledge. I first look at arguments in favor of a person displaying perpetrator knowledge. After that I turn to two categories of arguments for why a person does not possess such knowledge. First, there can be alternative explanations for why a report initially appears to contain verified details of a crime, which do not imply that the person actually has intimate knowledge of this crime. Second, even if a person knows intimate details of a crime, this does not always imply that they are the perpetrator.

4.1 Arguments for perpetrator knowledge

I have assumed so far that we learn that someone has perpetrator knowledge from a report. Such a report will often be a written or verbal statement which contains one or more verified details of a crime. This can, for instance, be a confession to interrogators or text messages to fellow perpetrators. As we have seen, perpetrator knowledge can also reveal itself through the websites that someone visited or their online search history.

I want to mention four more sources for determining whether someone has perpetrator knowledge. First, perpetrator knowledge can show itself in gestures. This was for instance the conclusion in a murder case from 2009.³⁸ During an interrogation, the suspect was told that trace evidence made it clear where the victim’s body was held when it was dragged to a different location. When asked where he had held her, he made a grabbing motion toward the waistband of his trousers. This was indeed the place where his DNA was found on the victim’s body. The district court therefore concluded that the suspect had perpetrator knowledge.³⁹

Another possible source for determining whether someone has perpetrator knowledge is that they were able to perform actions that only someone with perpetrator knowledge could undertake. For instance, in a case from 2023, the court concluded that the suspect had such knowledge based on a glove being found in a drugs lab.⁴⁰ This glove had blood on it which matched the suspect’s DNA. The court argued that it was unlikely that the glove ended up there by accident and noted that only people with perpetrator knowledge—that is, those who knew where the drug lab was and how to enter it—could have left the glove there.

A third way of determining the existence of perpetrator knowledge is shown in a case from 2017 case, where the suspect claimed to suffer from amnesia.⁴¹ They were subjected to a so-called Forced Choice Recognition test by an expert. During such a test, a person is shown two objects: an object that they have seen before and an object that they have never seen before and they must determine which is which (Bayley et al., 2008). In this case, the test showed no sign of memory problems in the suspect. The expert concluded that the evidence pointed to them faking their amnesia. Based on this, the court wrote that this test provided “a strong indication that the suspect had perpetrator knowledge.”

A somewhat futuristic, though also controversial method of detecting perpetrator knowledge is through (EEG-P300) brain scans (see Bijlsma et al., 2022: 71–72). The hypothesis underlying this method is that such scans can differentiate between the perpetrator and an innocent person seeing, for example, a photo of the murder weapon among photos of other weapons. This method was used to convict in a case in India, though it was later criticized and declared unconstitutional by the Indian Supreme Court (Van Toor, 2017). Brain scan techniques are currently only part of procedural criminal law in Japan (Van Toor, 2017). In general, the use of such techniques to detect perpetrator knowledge is highly controversial and heavily criticized (Meijer et al., 2013).

4.2 Alternative explanations for the report

Sometimes a report may appear to contain details of a crime that only the perpetrator could have known about. However, further examination may reveal that there is an alternative explanation for the contents of this report that does not imply that the person knew details of the alleged crime. We can represent such an alternative explanation as shown in Fig. 8 (using the argument from Section 3.1.1).

The alternative explanation competes with the explanation that the suspect knows details of the crime. This means that as the alternative becomes more probable, SUSPECT KNOWS decreases in probability. I encountered five types of these alternative explanations in the analyzed cases.

4.2.1 Report insufficiently detailed

According to this argument, a person does not display perpetrator knowledge because their report lacks detail. An example comes from a violent robbery case in which the prosecution claimed that a witness had perpetrator knowledge and that they got this from the suspect, who confessed the crime to them.⁴² The defense argued against this claim on the grounds that the witness’ account only gave a broad, vague account of the alleged events.⁴³

4.2.2 Details not accurate

Aside from not confirming the details of a report, the evidence can also show that this report is inaccurate. To give an example, a suspect was accused of murder.⁴⁴ According to the prosecution, one of the witnesses displayed perpetrator knowledge because he was able to name the brand of the car used as a getaway vehicle in a murder case. However, the court rejected this argument because the witness claimed that the car was green or silver, while it was actually blue.⁴⁵

4.2.3 No evidence for the details

Even when a person reports on specific details, these details need to be verified to count as perpetrator knowledge. In other words, the events mentioned in the report have to be supported by the available evidence. But this is not always the case. For example, in one case, the defense argued that even though it might appear as if the suspect displayed perpetrator knowledge in their confession, police investigators failed to investigate whether this knowledge might have been obtained from the media, the interrogators, or other sources.⁴⁶ The defense also claimed that the investigators should have checked whether the details mentioned fit with the evidence in the case and to gather further evidence to verify or falsify the claims made in the confession.

4.2.4 Misinterpreted

Sometimes a person with presumed perpetrator knowledge argues that this presumption rests on a mistaken interpretation of their report. For instance, in a murder case, the suspect searched for terms such as “deadly blow to the head” and “cleaning up after crime.” However, the defense argued that these searches could also be explained by the fact that the suspect had a general interest in crime shows.⁴⁷

4.2.5 Guessing and general knowledge

Sometimes a person can correctly guess details of a crime even though they do not have actual knowledge of this crime. Ofshe and Leo (1996) note that “the likelihood of an unlucky guess diminishes as the number of possible answers to an investigator’s questions grows larger.” While this is true, guessing what happened during a crime is rarely a matter of randomness, as if one were guessing a number between 1 and 100. Instead, the guesses in question are typically based on the person’s “general knowledge.” In other words, the person tries to come up with what they believe such crimes usually look like (often also taking what they do know about this specific case and what the investigators have told them into account). To give an example, in a case about dumping waste from drug production, the suspect retracted an earlier confession which seemed to contain perpetrator knowledge such as that the waste was dumped in a hole.⁴⁸ The suspect claimed that he was covering for someone else by confessing, but had not realized that he was confessing to a crime. He said that he had guessed details of the crime based on the fact that such drugs are usually disposed of by being dumped in a hole.⁴⁹

4.3 Alternative explanations for the knowledge

Aside from alternative explanations for the report, there can also be one or more alternative explanations for why a person has detailed knowledge of a crime, other than that this person is the perpetrator. It is extremely important that courts and others are mindful of these alternative explanations. For example, in a content analysis of 20 false confessions, Appleby, Hasel, and Kassin (2011) found that all of them contained visual and auditory details about the crime and how it was committed, about the time and location, and about the victim. Such confessions can easily mislead us if we are not cognizant of the fact that there are many different ways in which an innocent person could come to know the details of a crime. We can represent the possibility of such alternative explanations of knowledge (Fig. 9).

Most of the literature that discusses perpetrator knowledge focuses on one type of alternative explanation for knowledge, namely that the suspect received this information from their interrogators (this is called contamination)⁵⁰ (e.g. Ofshe and Leo, 1997; Garrett, 2009; Leo, 2013; Alceste, Jones, and Kassin, 2020). However, there are many other types of alternative explanations for knowledge.

4.3.1 Attorney

A suspect’s attorney will often have access to the case file or they might learn details of the crime from the police or in some other way. They could share these details with the suspect. An example of this occurred in a case where the suspect confessed to murder. The suspect (unsuccessfully) argued that he received information about the cause of death and other details about the crime from his attorney.⁵¹

4.3.2 Case file

Suspects (and also others) will often receive the case file at some point during the criminal proceedings. They may learn about details of the crime from this. For example, in an arson case, the prosecution argued that the suspect sent letters with threats to different people and institutions and that these letters contained perpetrator knowledge.⁵²

4.3.3 Contamination during interrogation

Interrogators may (consciously or subconsciously) share the details of the crime with the person being interrogated. This is the most commonly mentioned alternative explanation in the analyzed cases, which fits with findings from other legal systems. For instance, Garrett (2011, 2015) wrote two articles in which he studied a number of known false confessions and concluded that, respectively, 95% and 91% of them included “surprisingly rich, detailed, and accurate information,” which were likely disclosed to them during the interrogation process. To give an example, Israels and van Koppen (2006) discuss the famous judicial error in the case of Ina Post, who was convicted of manslaughter in the case of death of an elderly woman.⁵³ The evidence in this case was mainly her own confession, which contained false perpetrator knowledge. As Israels and van Koppen point out, the confession changed over time in a way that coincided with new knowledge obtained by the investigators and it contained errors that matched up with false beliefs held by the police at that point in the investigation.⁵⁴ Furthermore, at one point in her confession, Ina discussed the purchase of an item several years prior. She claimed to remember that this item cost €34, but not what it was that she had bought. This was surprising under the assumption that she was guilty as people will typically remember what they bought but not the exact cost. However, it matched up with what the police knew, as investigators had found a receipt of this purchase which only contained the price of the item but not what it was.

4.3.4 Media

Details about crimes are sometimes mentioned on television,⁵⁵ in newspapers, in books, or on the Internet. Often these details are shared with the media by the police.

4.3.5 Other, known

Sometimes a party will mention one or more persons that the perpetrator knowledge could have been learned from, without specifying how the person(s) in question obtained this knowledge themselves. This leaves open possibilities such as that they were a witness or had access to the case file or learned it from others. The two most commonly mentioned categories of such people are family members and fellow suspects. However, to give a different example, in one case, the defense argued that the suspect could have obtained perpetrator knowledge from a neighborhood police officer, but did not specify whether this was, for instance, because this officer was involved in the case, was a witness, or had heard about it from fellow officers.⁵⁶

4.3.6 Other, unknown

In contrast with the previous explanation, someone can also claim to have heard the information from others, without naming a specific person that they got this information from. For instance, in one case, the suspect displayed perpetrator knowledge during recorded conversations. He later claimed that he was just telling a story that he had heard from someone else during these conversations.⁵⁷

4.3.7 The perpetrator

Suspects sometimes claim that they learned details of the crime from the actual perpetrator. For instance, in a robbery case, the suspect claimed to have received information about the injuries sustained by the victims from “someone involved in the crime,” though he refused to name this person because he claimed to be scared of retaliation.⁵⁸

4.3.8 Victim

In many cases, the victim and the suspect know one another, for instance, because they are family. The suspect could learn details of the crime by talking to the victim about the events. While this alternative was not mentioned in the analyzed cases, I have included it because it is easy to imagine cases where this would be a relevant explanation.

4.3.9 Was participator, but did not perpetrate

Sometimes a suspect will argue that they do have perpetrator knowledge but that this was not obtained from being involved in the alleged crime. Instead, they claim to have gained this knowledge from a different, less serious crime. For example, they might argue that while they were part of a criminal organization that smuggles cocaine, their involvement in the alleged smuggling was minor and insufficient to be convicted for co-perpetration. According to the defense, the suspect obtained their perpetrator knowledge through their participation in this organization (a less serious crime than smuggling drugs).⁵⁹

4.3.10 Was victim

Often it is not just the perpetrator who knows the details of a crime but also the victim. In a rape case, the suspect argued that the victim’s testimony was unreliable, as it contained a suspicious amount of perpetrator knowledge. However, the court concluded that the details that she shared could reasonably have been known by a rape victim.⁶⁰

4.3.11 Was witness

A suspect can learn details of the crime because they were a witness of that crime. For example, in one case, a man was accused of murder. The prosecution claimed that he had perpetrator knowledge because he knew how the victim was killed. However, the suspect argued that he never denied being at the victim’s home that evening but that he was a witness to the events.⁶¹

4.3.12 Witnesses

Sometimes a person can hear details of a crime from people who were witnesses. For instance, they might learn the cause of death of the victim this way.⁶²

5. The appeal to probability fallacy

As I noted earlier, if we define perpetrator knowledge as “knowledge that only the perpetrator could have,” we view the concept as an all-or-nothing matter. On this definition, if there is an alternative explanation for the (apparent) knowledge of the crime, then this knowledge would, presumably, have no probative value. After all, it is then not only the perpetrator could have known these details of the crime. But the problem with this way of looking at perpetrator knowledge is that it is based on a reasoning error. Merely because an alternative explanation is a reasonable possibility does not mean that we should treat it as if it is (probably) true. To do so means to commit the fallacy of “appeal to probability,” where we take something to be the case (or to very probably be the case) because it is possible (Petric, 2020). Instead, the probative force of the (apparent) perpetrator knowledge will depend on our probabilistic assumptions.

To illustrate which probabilistic assumptions are relevant and how these would change our conclusions, let us consider a simple, fictional case in which a suspect is interrogated about a murder. During this interrogation, he mentions the type of gun used in this murder. At trial, the prosecutor argues that this is perpetrator knowledge. The defense offers an alternative explanation, namely that the suspect got this information from the interrogators (i.e. there was contamination). We can model this as in Fig. 10.

The numbers in this Bayesian network are purely for illustrative purposes. I have assumed that the prior probability of guilt is 1%.⁶³ In other words, we assume that there is a 99% chance that the suspect is innocent before taking into account any evidence (including the suspect’s perpetrator knowledge). The other numbers represent the conditional probabilities that the suspect would know what gun was used in the murder under various hypotheses (e.g. they were not the perpetrator and there was contamination, they are the perpetrator and there was no contamination). I have not yet entered a prior probability for the hypothesis that there was contamination—that is, the interrogators leaked information about the murder weapon to the suspect.

If we set the value of REP GUN TYPE to “True” (i.e. we learn that the suspect mentioned the type of gun used in the killing), this raises the probability that the suspect is the perpetrator. But the exact impact of this evidence will depend on the probability of contamination. Table 1 shows how the impact of receiving the suspect’s report changes given various values of CONTAMINATED.

As we can see, if we know with certainty that the suspect was not contaminated, then the report is extremely strong evidence, raising the probability that the suspect is the perpetrator from 1% to over 80%. But even if we are in doubt about whether there was contamination, the report can still make a significant change. For example, if we assign a 50% probability to CONTAMINATED, the report still more than doubles the probability that the suspect is the perpetrator. When combined with further evidence, this report could realistically make the difference between conviction and acquittal. Of course, the opposite is also true, simply because we are uncertain about an alternative explanation, it does not mean that we should ignore this possibility. We can see that even a 1% chance of contamination lowers the impact of the suspect’s report enormously.

Contamination is not the only variable that can influence the probability of guilt. For instance, even if we have strong evidence of contamination, it matters how strong an effect this contamination would produce on an innocent person. In other words, how likely would the contaminated suspect be to know the type of gun used? Suppose that the interrogators accidentally mentioned this information once during a long interview. This would be much less likely to have an effect on the suspect than if the police kept bringing up these details. We can model this in the Bayesian network. Suppose that we set the prior probability of CONTAMINATED to 90%. Table 2 shows how different assumptions about the probability that an innocent, contaminated person would know about these details influence the probability that the suspect is the perpetrator.

As we see again, it matters a great deal how confident we are that the remarks by the interrogators would lead to knowledge on the part of the suspect. This again shows the importance of thinking probabilistically. But note that to think probabilistically does not have to mean using Bayesian networks to calculate probabilities. What it means is that we need to understand how different sources and levels of uncertainty should roughly impact our resulting beliefs.⁶⁴ For instance, simply because there is a plausible alternative explanation for apparent perpetrator knowledge does not mean that we should assign no probative value to this knowledge, nor should we ignore the possibility of something like contamination entirely, even if the probability that this occurred in a given situation is small.

It is difficult to say whether courts tend to think about perpetrator knowledge as an all-or-nothing concept. Considerations on perpetrator knowledge are usually only a small part of a large set of evidence and courts typically do not spell out what role individual items of evidence played in their ultimate decision to acquit or convict. Nonetheless, there are cases in which it at least seems as if a court fails to think about perpetrator knowledge probabilistically. To give an example, in a manslaughter case from 2017, the court of appeal acquitted the defendant.⁶⁵ The district court had originally convicted the defendant in part on the basis of witness testimony. An important argument for the reliability of this witness was that he displayed perpetrator knowledge regarding the scene of the crime and the victim’s wounds. According to the prosecutor and the district court, he could only have learned this information from the suspect. However, the court of appeal concluded that “it could not be excluded” that the witness got this information from a television program. The witness argued that he never watched television because he suffered from epilepsy. The court of appeal wrote that this claim was not plausible, however. It concluded that the witness’s statements were, in part because of this, “insufficiently reliable to be used as evidence.” It is difficult to say whether this means that the court assigns no probative value to the witness’s statement based on the ruling, but this does appear to be the case. This would indicate an all-or-nothing view of perpetrator knowledge, where such knowledge is considered to have no probative value merely because there is a plausible alternative explanation.

6. Conclusion

This article concerned the phenomenon of “perpetrator knowledge.” The classic example of perpetrator knowledge is that of a suspect who confesses to a murder during an interrogation and, for instance, knows details about the murder weapon that are not publicly known. Those who discuss such cases often do so in the context of wrongful confessions that involve false perpetrator knowledge. In those cases, the apparent perpetrator knowledge is typically the result of “contamination”—that is provided by the interrogators. However, as this article shows, perpetrator knowledge (or a lack thereof) is relevant evidence in a wide variety of cases and can be possessed by other people than just the suspect. Moreover, there are more ways of displaying such knowledge than merely in an oral or written confession. For instance, perpetrator knowledge can also be revealed through a person’s text messages or search history. Finally, contamination is far from the only alternative explanation for apparent perpetrator knowledge. For example, a suspect may have gained knowledge of a crime from the media or the victim. One contribution of this article is therefore to illustrate the richness of the concept of perpetrator knowledge.

Aside from showing how perpetrator knowledge is used in practice, this article also offered a precise definition of this concept in terms of the Bayesian likelihood ratio. On this definition, we can be certain to varying degrees that information reported by a person is perpetrator knowledge. Apparent perpetrator knowledge can have more or less probative force, depending on this level of certainty. Yet while the likelihood ratio helps us define the concept of perpetrator knowledge, it leaves out too many details to adequately capture the different ways in which such knowledge can be utilized as evidence. Bayesian networks offer an elegant framework for solving this issue. They allow us to model the structure of the various arguments used in criminal cases. These networks can also help explore how specific probabilistic assumptions impact the probative force of apparent perpetrator knowledge.

The central claim of this article is that it is essential to reason probabilistically about perpetrator knowledge. When we fail to do so we run the risk of committing the appeal to probability fallacy—where we over- or underestimate the strength of apparent perpetrator knowledge because we think of it as an all-or-nothing matter.

Footnotes

Acknowledgements

I would like to thank, Ron Allen, Christian Dahlman, Dave Lagnado, Ludi van Leeuwen, Anne Ruth Mackor, Henry Prakken, Paul Roberts, Emily Spottswood and Marouschka Vink, the members of UCALL, and the two anonymous reviewers for their helpful comments on earlier drafts of this article.

Funding

No funding was obtained for the preparation of this manuscript.

REFERENCES