Understanding computational hermeneutics: Making meaning between the past and the present

A large group of scholars led by Cody Kommers and Drew Hemment at the Alan Turing Institute recently published a paper on ‘computational hermeneutics’. They mention Hans-Georg Gadamer and Wilhelm Dilthey, two godfathers of hermeneutics, and talk about situated meaning, ambiguity and the plurality of meaning. How intriguing, I thought.

The paper brought back memories of my past work on meaning and interpretation in general linguistics and the history of linguistics. It made me wonder whether modern thinking about computational hermeneutics and past thinking about context and meaning could inspire each other even more than one might expect from the paper?

In the following, I’ll first set the scene with a very brief introduction to traditional hermeneutics, followed by some comments on an older version of computational hermeneutics, before summarising what I see as the core ideas contained in the new computational hermeneutics paper. Then I present both some older theories about situated meaning developed by Philipp Wegener at the end of the 19th century and some ideas on purposive ambiguity developed by myself and David Clarke at the turn of the millennium. Throughout, I will point out possible links between the older contextualist and hermeneutic approaches to understanding and meaning-making and this recent foray into ‘computational hermeneutics’.

This post is a bit of an experiment, as it emerged from writing about computational hermeneutics through computational hermeneutics, that is, a virtual conversation between me, a 19th-century linguist (Philipp Wegener) and a 21st-century AI system (Claude). All mistakes are mine!

Traditional hermeneutics

According to the Stanford Encyclopedia: “Hermeneutics as the methodology of interpretation is concerned with problems that arise when dealing with meaningful human actions and the products of such actions, most importantly texts”. As one of its main proponents, Wilhelm Dilthey, wrote, it is used to understand ‘the outer expressions of inner life’. 

Towards the end of the 19th century, Dilthey (and others) introduced a distinction between Verstehen and Erklären or understanding and explanation which they partially mapped onto the methods used in the human and natural sciences. The lynch pin of the hermeneutical method used in the human sciences is not causal explanation, based on Erklären, drawn from experimental or observational evidence, but understanding, Verstehen, based on the hermeneutic circle. 

The hermeneutic circle is the iterative recontextualisation and refinement of the meaning and understanding of a text or cultural artefact by going from part, say word or chapter, to whole, say scroll or novel and back again. The important thing is not so much replication but contextualisation in a sort of continuous feedback loop. 

As the new hermeneutics paper says, the hermeneutic circle “describes the interpretation of an artifact as an iterative process between understanding the meaning specific part of the artifact and the meaning of the artifact as a whole.” (Kommers et al., 2025: 3) 

Let’s now turn to ‘computational hermeneutics’, which, as we will see, has two meanings!

Digital hermeneutics

When I first heard the phrase ‘computational hermeneutics’ I thought we were dealing with the way that modern computational tools can help (digital) humanities scholars in their work – digital hermeneutics. I assumed it meant applying computational methods to hermeneutic and interpretative inquiries, such as using algorithms to identify (semantic or thematic) patterns in large corpora of texts, for example. But I was wrong.

Computational hermeneutics

In the paper “Computational Hermeneutics: Evaluating Generative AI as a Cultural Technology”, the authors don’t advocate using computation to aid human interpretation; rather, they argue that we need hermeneutic theory to better understand and evaluate generative AI systems, especially LLMs, themselves.

Here the hermeneutic circle comes in. As the authors say: “We must look at both the system-level generalizations and context-specific outputs in interpreting the outputs of these models“ (Kommers et al., 2025: 3) – and this iteratively. It’s not so much about using computers to aid human interpretation, but about using hermeneutics, as a framework for interpretation, to better understand and evaluate AI systems themselves.

The core insight seems to be that generative AI systems are fundamentally interpretive machines, what the authors call ‘context machines’, that operate within cultural contexts, not neutral tools that can be evaluated through culture-blind right/wrong metrics. By framing GenAI as context machines, the authors highlight that these systems don’t just process information, but that they (and those who operate them, I suppose) constantly make interpretive choices about meaning, context, and cultural significance.

The paper posits three principles of hermeneutic evaluation:

• Situatedness: meaning always emerges from specific contexts – this challenges the idea that AI can produce universally ‘correct’ outputs divorced from cultural setting. (A culturally blind AI could, for example, give inappropriate medical advice)
• Plurality: multiple valid interpretations can coexist – this calls into question the usual evaluation frameworks that assume single ‘right’ answers and is more in tune with human creativity.
• Ambiguity: conflicting interpretations are natural and inevitable – these are not bugs to be eliminated but an inherent, even productive, part of how we make meaning together.

Overall, instead of asking “Is this AI output accurate”, they ask, “What cultural work is the AI doing and how does it negotiate meaning-making in specific contexts?”

In the following, I’ll pick up on three topics, namely situatedness, meaning-making in context and ambiguity and try to uncover some historical roots for these concepts which might still have useful lessons for modern LLMology.

Situation and attention

The focus on situatedness in the paper reminded me of the work of Philipp Wegener who published a seminal but mostly forgotten book Untersuchungen über die Grundfragen des Sprachlebens [Investigations into the fundamental questions of linguistic life] in 1885, around the time that hermeneutics came into fashion. (This was translated in 1971 under the title Speech and Reason).

Wegener was one of the first linguists, I believe, who emphasised that meaning always emerges from the totality of communicative situations in cooperation between speakers and hearers. This tradition continued in the work of the anthropologist Bronislaw Malinowski, the Egyptologist and general linguist Alan Henderson Gardiner, and the general linguist John Rupert Firth, who all read Wegener.

Referring to Firth and Ludwig Wittgenstein, Ted Underwood, one of the co-authors of the hermeneutics paper, wrote recently that, like today’s language models, such scholars “sought meaning not in universal laws but in the patterning of everyday language and cultural  practice”. He refers in particular to Firth’s famous saying: “you shall know a word by the company it keeps”. This is, in way, the essence of how LLMs work as they seek meaning in patterns – in the company that words keep – in massive amounts of text.

According to Wegener, the overall situation consists in the situation of perception (including actions, gestures, etc.), the situation of remembrance or consciousness (including prevailing interests and ideas), and the cultural situation (Wegener, 1885: 22-27; 1971: 135-139).

As Clemens Knobloch, a Wegener specialist, pointed out in 1991; “Crucial for Wegener’s understanding of ‘situation’ is […]  the form and direction of attention that prevails or is habitually present in a person’s actions and experiences. Attention is attached for example to ongoing or just completed activity, and this is where the utterance usually fits in and becomes comprehensible.” 

Wegener stresses that what he calls verbal ‘exposition’, what one may call co-text, can replace situational context and be used in the negotiation of meaning (see next section). And with this we sort of come to LLMs, context and something called the attention mechanism. This revolutionised LLM transformer architecture in 2017 and lets LLMs deal with long-range dependencies between words. “Unlike traditional methods that treat words in isolation, attention assigns weights to each word based on its relevance to the current task. This enables the model to capture long-range dependencies, analyze both local and global contexts simultaneously, and resolve ambiguities by attending to informative parts of the sentence.” 

Might it be that what revolutionised LLM architecture in 2017 was a computational approximation of something Wegener identified as fundamental to human meaning-making in 1885? Context and attention together seem to be the foundations for the emergence of meaning in ‘context machines’, be they human or artificial. But how does that work?

Negotiating meaning in context

As I have hinted at above, Wegener proposed that verbal exposition can replace concrete situational context when constructing meaning between speaker and hearer. He elaborates on this in his work on exposition and predicate (posthumously published in 1911), which might also be interesting to LLM analysts.

How do speakers and hearers manage to construct meaning together? For this to happen there needs to be a balance between exposition and predicate, i.e. between what the speaker and the hearer know already or need to know to produce and understand an utterance and the new thing the speaker wants to convey. 

An exposition, in Wegener’s sense of the term, is everything which prepares the ground (the known or given) for the appearance and the understanding of the predicate (the unknown or new). The situation in which communication happens prepares the non-verbal ground or context for understanding to happen, while the exposition (in the more restrictive sense) prepares the verbal ground or the co-text for the hearer to understand the predicate. To construct meaning from an utterance, the hearer must draw conclusions from the nature of the predicate itself, as well as the verbal exposition or co-text and the situational context. It’s an intricate dance of meaning-making.

Wegener uses as an example the one-word sentence ‘boots’ which might function as a command to a servant to bring one’s boots, but how does that happen? “In the word-sentence ‘my boots’, the pure word-image does not trigger the representation of the facts that (1) somebody orders an action; (2) what that action is; (3) who should execute the action. All this can only be inferred from the situation and the gestures. The word-image only evokes the representation of a definite thing that the speaker has in mind as an object (Wegener, 1921: 9-10; 1971: 120).

The relation between the situation of communication and the verbal exposition is a complicated one: the more situation there is, so to speak, the less verbal exposition we need; the more complex or opaque the situation, the more verbal exposition we need. This applies in particular to novels, old texts, texts from different cultures, etc., where the exposition must replace as much as possible the situational clues – a situation well-known to LLMs!

The predicate, on the other hand, which is what the speaker wants to convey or communicate to the hearer, is the new, interesting, important, valuable bit of our efforts to communicate. Keeping the right balance between the exposition (help for the hearer) and the predicate (the speaker’s news) is a valuable communication skill, one that structures all our communicative interactions. 

This balance also changes in the evolution of language itself, where the means of linguistic representation gradually evolve from language used in situated action. Words only become usable as verbal exposition tools after having gone through a phase where they were predicates needing active interpretation in situational context.

Now, one might say that LLMs are constantly performing this same balancing act between what can be taken as given (exposition) and what needs to be explicitly generated or predicted (predicate). When an LLM processes a prompt, it essentially performs Wegener’s interpretive work – inferring from context what the exposition is (what the human already knows, what the situation calls for, what cultural assumptions apply) and then generating the appropriate predicate (the new information or response).

Take Wegener’s ‘boots’ example. When someone types ‘boots’ into an LLM, the system has to do exactly what Wegener described: infer from available contextual clues (the conversational history, the apparent intent, cultural assumptions) what kind of response is being solicited. Is this a request for information about boots? A creative writing prompt? Part of a larger conversation? LLMs must, of course, construct the ‘situation’ from purely textual cues – cotext is LLMs context!

Wegener’s observation that the more situation there is, the less verbal exposition we need might explain why LLMs often work better with richer context (longer prompts, conversation history, explicit framing), as they are compensating for the lack of physical/embodied situation with more verbal exposition.

And finally, Wegener’s point about language evolution – words moving from predicates requiring active interpretation to becoming expositional elements – feels like it might illuminate how LLMs learn patterns where frequently co-occurring elements become more ‘expositional’ in their internal representations.

Context is, of course, also needed to resolve or, as we’ll see, exploit ambiguity, and we’ll come to that now. If you say ‘bank’ instead of ‘boots’, you need to know whether you are on river and want to approach it or on a city street trying to get money.

Ambiguity and plurality of meaning

In the previous sections we have encountered the word ambiguity several times. Hearing the word ‘ambiguity’ one might think that this is something to be avoided, something to be eliminated, a failure of clear communication. However, in a paper I wrote with my husband in 2001, I tried to make the point that ambiguity isn’t just an obstacle to communication, but instead that it can do important communicative work. People don’t always want to disambiguate; sometimes ambiguity is itself the point, as especially when intentionally or unintentionally joking, punning, being diplomatic, letting multiple interpretations hang in the air and so on – what one may call purposive, deliberate or constructive ambiguity.

This challenges a core assumption in AI evaluation, that ambiguity is a problem to be solved rather than a resource to be exploited. It chimes with the computational hermeneutics approach to plurality (that multiple valid interpretations coexist) and ambiguity (that interpretations naturally conflict).

The authors of the hermeneutics paper identify ambiguity as one of the three interpretive challenges GenAI must navigate. Current benchmarks often penalise ambiguous responses or treat them as failures of clarity. But if ambiguity has, as we argued in our 2001 paper, genuine semantic and pragmatic functions – reinforcing semantic networks, strengthening social bonds, negotiating conversational turns – then maybe evaluations use the wrong criteria.

It is interesting in this context to observe how LLMs sometimes give responses that feel deliberately open to multiple interpretations, or how they might preserve useful ambiguities rather than forcing premature disambiguation. In Wegener’s terms, maybe sometimes the predicate should remain partially ambiguous to allow the human interlocutor to complete the meaning-making process collaboratively.

If we combine Wegener’s exposition/predicate framework with the work on purposive ambiguity, this might suggest that effective human-AI interaction might require systems that cannot just manage ambiguity but strategically deploy it. Rather than always striving for the most unambiguous response, maybe LLMs should sometimes maintain productive ambiguities that invite further interpretive, indeed hermeneutic collaboration. This then reframes LLMs from being pure answer machines to being more like hermeneutic negotiation partners.

Conclusion

In the computational hermeneutics paper, the authors claim that LLMs are ‘context machines’ or ‘hermeneutic machines’, constantly engaged in contextual interpretation. This means always trying to make sense of incomplete and ambiguous information by drawing in or drawing on various types of contexts, which, in humans, include our bodies, embodied contexts, our physical environments, our lived experience.

This invites questions like: How does interpretation differ between human and non-human hermeneutic machines operating with or without an embodied contextual component? How does interpretative or hermeneutic work differ between purely text-based pattern matching or pattern recognition and embodied pattern matching? What does all this imply for how we should interact with AI/LLMs in the future?

I am not a computational linguist, let alone computational modeller, but I hope that my historical musings might contribute to emerging research into LLMs as cultural and social technologies. Things are moving fast in this space. This has just come out on LLMs and ambiguity for example…..

Further reading

Dobson, J. E. (2023). On reading and interpreting black box deep neural networks. International Journal of Digital Humanities, 5(2), 431-449.

Elffers-van Ketel, E. (1993). Philipp Wegener as a proto-speech act theorist. Linguistics in the Netherlands, 10(1), 49-59.

Farrell, H., Gopnik, A., Shalizi, C., & Evans, J. (2025). Large AI models are cultural and social technologies. Science, 387(6739), 1153-1156.

Firth, J. R. (1957). Studies in Linguistic Analysis. Oxford: Blackwell.

Fuenmayor, D., & Benzmüller, C. (2018, October). A computational-hermeneutic approach for conceptual explicitation. In International Conference on Model-Based Reasoning (pp. 441-469). Cham: Springer International Publishing.

Gadamer HG (2004) Truth and Method. London and New York: Continuum.

Gardiner, A. H. (1932). The Theory of Speech and Language. Oxford: Clarendon Press.

Henrickson, L., & Meroño-Peñuela, A. (2022). The hermeneutics of computer-generated texts. Configurations, 30(2), 115-139.

Kommers, Cody and Ahnert, Ruth and Antoniak, Maria and Benetos, Emmanouil and Benford, Steve and Bunz, Mercedes and Caramiaux, Baptiste and Concannon, Shauna and Disley, Martin and Dobson, James and Du, Yali and Duéñez-Guzmán, Edgar and Francksen, Kerry and Gius, Evelyn and Gray, Jonathan and Heuser, Ryan and Immel, Sarah and So, Richard and Leigh, Sang and Livingston, Dalaki and Long, Hoyt and Martin, Meredith and Meyer, Georgia and Mihai, Daniela and Noel-Hirst, Ashley and Ostherr, Kirsten and Parker, Deven and Qin, Yipeng and Ratcliff, Jessica and Robinson, Emily and Rodriguez, Karina and Sobey, Adam and Underwood, Ted and Vashistha, Aditya and Wilkens, Matthew and Wu, Youyou and Yuan, Zheng and Hemment, Drew, Computational Hermeneutics: Evaluating Generative AI as a Cultural Technology (August 01, 2025). Available at SSRN

Makkreel, R. (2011). “Wilhelm Dilthey”, The Stanford Encyclopedia of Philosophy Edward N. Zalta (ed.).

Malinowski, B. (1923). The problem of meaning in primitive languages. In: J. K. Ogden and I. A. Richards, The Meaning of Meaning. Supplement I (pp. 296-336). New York: Harcourt, Brace & World. Inc.

Mallery, J. C., Hurwitz, R., Duffy, G (1986) Hermeneutics: From Textual Explication to Computer Understanding? A.I. Memo No. 871. MIT artificial intelligence laboratory.

Mohr, J. W., Wagner-Pacifici, R., & Breiger, R. L. (2015). Toward a computational hermeneutics. Big Data & Society, 2(2), 2053951715613809.

Nerlich, B. (2002). Change in Language: Whitney, Bréal and Wegener. London: Routledge.

Nerlich, B., & Clarke, D. D. (1996). Language, Action and Context: The early
history of pragmatics in Europe and America, 1780-1930. Amsterdam: John
Benjamins.

Nerlich, B., & Clarke, D. D. (2001). Ambiguities we live by: Towards a pragmatics of polysemy. Journal of Pragmatics, 33(1), 1-20.

Picca, D., Schnyder, A., & Romele, A. (2024). Computational hermeneutics of emotion: a comparative study of emotional landscapes in the Dostoevsky’s novel “Crime and Punishment”. Humanities and Social Sciences Communications, 11(1), 1-10.

Romele, A., Severo, M., & Furia, P. (2020). Digital hermeneutics: from interpreting with machines to interpretational machines. AI & SOCIETY, 35(1), 73-86.

Underwood, T. (2025). The impact of language models on the humanities and vice versa. Nature Computational Science, 1-3.

Wegener, P. (1885). Untersuchungen über die Grundfragen des Sprachlebens. Newly edited by Konrad Koerner (University of Ottawa), with an introduction by Clemens Knobloch 1991. Amsterdam: John Benjamins.

Wegener, P. (1911). Exposition und Mitteilung: Ein Beitrag zu den Grundfragen des Sprachlebens. Geschichte des Gymnasiums und der Realanstalt zu Greifswald von 1861-1911, hrsg. von Dr Max Schmidt, Teil: Wissenschaftliche Aufsätze, 1-21. Greifswald: Julius Abel.

Wittgenstein, L. (1953). Philosophical Investigations. German text and English transl. By G.E.M. Anscombe. Oxford: Blackwell.

And there is a podcast on Wegener’s work!

Image: Picryl: Vassily Kandinsky (1866–1944) Circles in a Circle, 1923