TC has been typically assessed by semi-structured interviews, such as the Contemporaneous Assessment Instrument (CAI), which follows the six criteria established by Applebaum and Grisso that were pertinent to the four legal requirements for decision-making (understanding the nature of the act and its effect; the extent of the property the testator is disposing; the claims which the testator should give effect; to be free of a disorder of mind preventing the disposing of one’s property). These are that the functions assessed should be conceptually close to the standards of competence; the content of the instrument should be pertinent to the distribution of one’s estate; the instrument should be meaningful to the person assessed; the instrument should be sufficiently standardized to enable comparisons within and across research groups; the measures must have objective and reliable scoring criteria; and the instrument should be practical for use. By definition, some of the criteria must be specific to the person being assessed (2).

The Testament Definition Scale (TDS) is a short rating scale designed to evaluate and quantify the individual’s ability to give an exhaustive definition of what a testament is, thus addressing the first criterion derived from the Banks v Goodfellow case (10).

The Legal Capacity Questionnaire (LCQ), a screening instrument for attorneys, examines three of the four legal requirements of TC using true/false, multiple-choice, and open-ended questions. On the basis of limited normative data for older adults and correlational analyses with mental status and behavioral instruments, cut scores have been created for high, borderline, and low capacity. The LCQ is geared toward older clients with cognitive disabilities, is administered in a standardized manner and is norm-referenced but lacks reliability and validity information, the criterion of understanding what a Will is, and the investigation of psychotic symptomatology and undue influence (6).

The Testamentary Capacity Assessment Tool (TCAT) (11) is a brief instrument that has been validated in older adults with dementia by comparing the scores with the opinion of a forensic psychiatrist regarding the TC of the patient, and the judgment of a senior Judge. It has been standardized in an Italian neurologically healthy sample (12). The TCAT examines orientation, autobiographical memory, knowledge of financial parameters, intention and judgment. Behavior pertaining to presence of acute psychopathology is also documented. The instrument can be used to screen the clearly competent testator or as corroborative evidence of an expert’s judgment. Because it is based only on the interview with the patient independent of collateral information it is impartial to potential conflict within the patient’s family.

The more recent Testamentary Capacity Instrument (TCI), Martin et al. (13) is an interview-based instrument that assesses the aforementioned four criteria of TC in a standardized interview-based format, utilizing a variety of question formats depending on the section and provides objective scoring criteria. Preliminary data showed that it was a valid measure of TC with good inter-rater reliability and internal reliability for the criteria for which such measures could be calculated. No cut scores were proposed for determining capacity, because the TCI was intended to serve as a source of additional information for making clinical judgment.

Neuropsychological assessment is not a source of direct evidence for TC, although it can contribute to the clinical and diagnostic context for its assessment (14) and can provide important evidence where incapacity is alleged to invalidate a Will (3).

Cognitive screening measures such as the MMSE are not reliable and valid measures of TC and are not sensitive to impairments affecting judgment and reasoning despite their wide use in capacity evaluations. The cognitive domains of complex attention, executive function, learning and memory, language, and social cognition may affect TC in different ways, although there is limited information on how impairments in these domains impact on the Banks v Goodfellow criteria (3). A recent study that correlated performance on a neuropsychological battery to the TCI in a small number of cognitively healthy persons and persons with AD showed that measures of executive function, verbal memory, and semantic knowledge accounted for over 70% of shared variance with TCI performance and each measure correlated with each of the four TCI elements, suggesting a similar multidimensional cognitive basis for each TCI component (15).

Although a contemporaneous assessment is desirable, many situations require a retrospective assessment of TC. A proposed methodology along with pitfalls of conducting a retrospective neuropsychological assessment are described in Zago and Bolognini (16). Of interest is the inclusion of handwriting analysis, evaluated longitudinally, as a complementary tool for the evaluation of cognitive status, both with respects to formal aspects of the writing skills (spelling, grammar, layout) and logical consistency of the written text. The specifics of conducting such an analysis and deriving a “writing score” in the posthumous challenge of a will are provided in Balestrino et al. (17). Preliminary evidence points to the association of a low writing score with abnormal performance on mental status tests but the establishment of the validity of the method requires further research (16, 17). An assessment of medical history should precede such assessment to exclude testators with neurological conditions impacting on graphomotor skills, such as Amyotrophic Lateral Sclerosis or Parkinson’s disease, or other medical conditions such as rheumatoid arthritis.

Providing evidence-based assessment and reasoning for a TC opinion requires training that the person conducting the assessment may not have. Moreover, while effort is being made for establishing more objective criteria for TC judgment, assessment also needs to be tailored to the individual situation of the testator.

Evaluation of TC in aphasia and in neurodegenerative disorders characterized by progressive language impairment are a challenge to TC evaluation, as the testator’s ability to verbally express the components of TC and explain decisions is essential for establishing TC [e.g., case by (18)]. The assessment of TC in persons with severe aphasia is therefore problematic or even impossible, and techniques and formats that may facilitate the evaluation, such as prompted recall or multiple-choice formats for recognition memory, are not without drawbacks (13). It is proposed that persons with aphasia should have greater access to the courts through the facilitative support of speech-language pathologists (19, 20), which is in line with the supported decision-making trend of disability rights law (21).

In cases of severe aphasia, nonverbal neuropsychological assessment could serve as a proxy measure although it cannot replace the TC criteria. Linguistic deficits notwithstanding, other cognitive functions may be affected in aphasia (22), which may be independently impaired and may impact on mental capacity. To the best of our knowledge, there are no studies on the association of nonverbal assessment with neuropsychological tests of cognitive function that can be used in TC assessment. The establishment of such an association would facilitate the clinical and diagnostic context for TC assessment in aphasia and some types of dementia.

The new General Data Protection Regulation (GDPR) of the European Union enacted in 2018 prohibits automated decision-making, that is, making a decision solely by automated means without any human involvement, if a decision produces legal or similar consequences (23). The GDPR has also created a “right to explanation” of how an AI system works and how it reached a decision, which is especially pertinent for ML models that are inherently challenging to interpret (23, 40) and raise ethical issues, particularly in legal settings (24). It has been argued that the explanations that derive from correlational computations can lead to spurious correlations rather than cause-effect relationships, resulting in erroneous or biased explanations (41). International guidelines for trustworthy AI place explainability in the center of trustworthy systems and regulatory guidelines (42, 43). Therefore, there is a growing need for the design and development of transparent and interpretable AI models and explanations that reflect the needs of the users (44, 45).

Little attention has been paid to the potential contribution of AI to capacity assessment, although a theoretical paper has been written on the broader civil competency of financial capacity in cognitive disorders that includes AI (46). According to one perspective, capacity assessment does not lend itself to formalization; in mental health the notion of incapacity is a complex legal, clinical, ethical, and social construct. Testing for it varies according to the specific circumstance for making a decision, making it difficult to formalize mathematically as required by AI (47).

On the other hand, the use of ontologies for enhancing human understandability of global post-hoc explanations of black-box models, as presented in Confalonieri et al. (48) may be developed for machine analysis of human explanations. Using this reverse reasoning, it is the human explanations regarding the evaluation of TC that could be subjected to machine analysis. Thus, instead of humans evaluating machine-generated explanations as to their correctness, AI would be called upon for the evaluation of human-generated explanations of a person’s behavior, to facilitate the decision of the expert. The decision of the AI system can be explained using XAI methods in order to increase the trust of the human evaluator in it. The explanation that a person provides for a certain behavior or decision could be processed by a natural language programming (NLP) system in order to draw conclusions for the assessment of one’s TC.

Using NLP methods proposed for assessing artificial agents’ responses in similar situations (49), “what” questions may be used for the understanding of the nature of the act of making a Will and its consequences as well as knowledge of the extent of one’s assets, and “why” questions may help disclose the testator’s understanding of the claims of those who might benefit from the Will and the impact of the distribution of the testator’s assets. The computer analysis of the human explanations mentioned above may contribute to offering advice to the expert making the assessment as to their correctness. This analysis could be supported by an explanatory model of the kind proposed in Pfeiff (50) of explanatory practices in psychotherapy intended to make the patients assume responsibility for their symptoms. Two different notions of rationality are used: theoretical rationality, the adoption of beliefs that cohere with one’s relevant background knowledge; and pragmatic rationality, the adoption of beliefs that are consistent with one’s goals. The notion of dysfunctionality is proposed in which a decision is dysfunctional if it does not satisfy one’s needs and may even cause harm. By analogy, the notions could be adapted to the assessment of TC.

For a retrospective analysis of handwritten or written texts AI offers some tools for their automatic analysis that takes into account both the formal aspects of the writing skills (lexical, syntactic, semantic and discourse level) and the spatial orientation in cases of handwritten texts. There is a large body of research that presents methods for evaluating these aspects of natural language in texts using AI [e.g., (51, 52)]. The identification of emotional state through the analysis of handwriting features is a new area in AI (53) that may be extended to the identification of handwriting features that reflect cognitive state.

To develop and test such an XAI based decision support tool would require a corpus or dataset of explanations provided by persons of varying TC, together with expert assessment of the state of their TC. However, the presence of a language disorder would invalidate the method proposed for assessing reasoning capacity because an AI system analysis of the verbal explanations of the decisions of the testator presupposes lack of language disorder of the testator. It is almost impossible to construct an AI system assessing human explanations that do not exhibit correctness in all the linguistic aspects to be tested, as listed in Table 1. Therefore, a challenge for the NLP system supporting the experts in their assessment of explanations would be to detect a language disorder as distinct from problems with reasoning and filter out the aberrant explanations before they are submitted to the XAI decision support system (DSS). Table 1 lists the basic tests to be performed by the NLP system during the filtering process.

The design and implementation of an XAI DSS for TC assessment would have the following characteristics: (a) be a truly mixed human-AI system (b) be trustworthy, (c) utilize tacit knowledge of area ontology and simulation models, (d) be tested for its adaptation to the users regardless of their age, race, gender, or capabilities, and e) adapt its explainable decision support and explainable question answering “behavior” to the needs of the users. A mostly unexplored area is the variation of the explanation generation of the testator spanning several axes, such as from special to general ontology levels. This variation of the style of explanations would depend on the educational level and background of the testator. The explanation analysis software must be capable of processing a variety of styles of explanations, which is an open research question (see Appendix in Supplementary material for some possible variations to guide XAI experts).

Area under the curve (AUC) could be used as a test statistic to describe the ability of the tool to discriminate between capacity and incapacity and to compare the performance of different tools and methods, after Garb and Wood (24).