Understanding and interpreting algorithms: toward a hermeneutics of algorithms

Understanding

Gadamer (1989) points out that

A person who is trying to understand is exposed to distraction from fore-meanings that are not borne out by the things themselves. Working out appropriate projections, anticipatory in nature, to be confirmed ‘by the things’ themselves, is the constant task of understanding. (p. 267)

Thus, the task of understanding algorithms in everyday life is the process of ‘working out appropriate projections’ in accordance with the perceptible algorithmic activity at hand. To be confirmed by the things (i.e. algorithms) themselves is to acknowledge their ordinary presence and, with our own fore-meaning, trying to make sense of how they are doing what they are doing, in what ways and why. However, we cannot make sense of algorithms in whatever way we like:

But understanding realizes its full potential only when the fore-meanings that it begins with are not arbitrary. Thus it is quite right for the interpreter not to approach the text directly, relying solely on the fore-meaning already available to him, but rather explicitly to examine the legitimacy – i.e., the origin and validity – of the fore-meanings dwelling within him. (Gadamer, 1989: 267)

Understanding here is a way of explicating our fore-meanings to ourselves in our process of approaching algorithms, not only relying automatically on fore-meanings, and scrutinizing the suitability of our fore-meanings for making sense of algorithms. However, this does not mean that we should and do necessarily ‘agree’ with the algorithms and the results of their actions because,

. . . what another person tells me, whether in conversation, letter, book, or whatever, is generally supposed to be his own and not my opinion; and this is what I am to take note of without necessarily having to share it . . . Of course this does not mean that when we listen to someone or read a book we must forget all our fore-meanings concerning the content and all our own ideas. All that is asked is that we remain open to the meaning of the other person or text. But this openness always includes our situating the other meaning in relation to the whole of our own meanings or ourselves in relation to it. (Gadamer, 1989: 268)

What Gadamer reminds us of here is that whatever is being communicated to us in a process of communication by someone, we do not have to subscribe to what is communicated and we do not have to eliminate our fore-meanings. However, we have to take note of the meaning of the other when trying to understand what is being communicated to us and what the communication is about. We cannot understand texts and communication in whatever way we like. In a process of understanding, there must be something to understand and that ‘something’ is the other meaning, which must be kept ‘in relation to the whole of our own meanings or ourselves in relation to it’. As for algorithms in communication, this implies that we cannot dismiss them on the grounds that they are awkwardly coded, that they come up with odd results, or that they produce failures as ‘The emergence of failures has to do with the complexity of interactions . . . Consequently, the everyday use of algorithms results in a mixture of surprise and disappointment’ (Roberge and Seyfert, 2016: 14). Algorithms are the meaning of the other which we must remain open to but without a necessity to share that meaning. We see that when people are trying to circumvent algorithms or at least align with them in digital communication by attaching particular metadata, and not others, in order to be captured by search engines (e.g. SEO’s). In these cases, users recognize and embrace that algorithms are there and that they say something, but users interpret and question them in particular ways to make the algorithms do something on behalf of them and then ‘The hermeneutical task becomes of itself a questioning of things’ (Gadamer, 1989: 269). However, we are not able to do that if we are not open to the ‘otherness’ of the algorithm and if we are not aware of our own bias. Gadamer (1989) calls this the ‘hermeneutically trained consciousness’:

a person trying to understand a text is prepared for it to tell him something. That is why a hermeneutically trained consciousness must be, from the start, sensitive to the text’s alterity. But this kind of sensitivity involves neither ‘neutrality’ with respect to content nor the extinction of one’s self, but the foregrounding and appropriation of one’s own fore-meanings and prejudices. The important thing is to be aware of one’s own bias, so that the text can present itself in all its otherness and thus assert its own truth against one’s own fore-meanings. (p. 269)

In trying to understand algorithms in digital communication, we must be prepared for them to tell us something. Algorithms do tell us something when they recommend, filter, or order digital products or data for us, implying that algorithms are more than ‘Logic + Control’ (Roberge and Seyfert, 2016: 17). Moreover, it is important in our daily digital interactions with algorithms that we recognize their alterity, their otherness. Due to their omnipresence in many forms of daily digital communication, we cannot let go of the presence of algorithms and their otherness. They are simply there being inscribed in our software-based media life. But what does that mean? As when we are confronted with any other sort of text, it means that we have to, in the first place, let the algorithm present itself in all its otherness; let the algorithm say what it has to say (its truth), that it has something to disclose, before we ‘match’ the algorithm with our own fore-meanings:

It is not at all a matter of securing ourselves against the tradition that speaks out of the text then, but, on the contrary, of excluding everything that could hinder us from understanding it in terms of the subject matter. It is the tyranny of hidden prejudices that makes us deaf to what speaks to us in tradition. (Gadamer, 1989: 269–270)

So too it is with algorithms; they speak out of a tradition charged with the calculation techniques, choices, values, and motives built into them in order for them to perform in their appropriate settings (Cardon, 2016). Algorithms do communicate something (their otherness) to us. In order for us to make sense of algorithms and their communicative activity and apply algorithms in our own communications, we must take note of how they seem to perform (‘the subject matter’) in the first place before we begin to trick, tweak, or communicatively relate with algorithms in our digital interpretative activities. However, we do not neutrally observe how algorithms perform. We bring with us our own fore-meanings and prejudices. Otherwise, we are not able to frame our responsive interpretative actions accordingly. This leads us to another key assumption in Gadamer’s work: that all understanding involves prejudices and ‘. . . gives the hermeneutical problem its real thrust’ (Gadamer, 1989: 270).

Prejudices

The idea of prejudices and their role in understanding is crucial to Gadamer’s argument. According to Gadamer (1989), it is the Enlightenment that has produced a ‘. . . prejudice against prejudice itself, which denies tradition its power’ (p. 270). In rehabilitating the role of prejudices in understanding, Gadamer (1989) reminds us that a prejudice is a preliminary judgment and its negative connotation familiar to us today depends on ‘. . . the positive validity, the value of the provisional decision as a prejudgment’ (p. 270). Gadamer presents a critique of the Enlightenment because of its belief in reason as the only authority in matters of truth and understanding and its subsequent displacement of prejudices. Enlightenment is a reaction to dogmatic interpretations of the Bible. As such, Enlightenment is concerned with the hermeneutical problem of interpretation and tradition – but in a particular way: ‘It wants to understand tradition correctly – i.e. rationally and without prejudice’ (Gadamer, 1989: 272). Gadamer (1989) confronts this idea:

. . . there is a special difficulty about this, since the sheer fact that something is written down gives it special authority. It is not altogether easy to realize that what is written down can be untrue. The written word has the tangible quality of something that can be demonstrated and is like a proof. It requires a special critical effort to free oneself from the prejudice in favor of what is written down and to distinguish here also, no less than in the case of oral assertions, between opinion and truth. (p. 272)

Pointing out that what is written down has a special authority (e.g. contracts, receipts, manuals, or algorithms) and that we have no reason to or interest in questioning it as this is a rather difficult effort, is important here. Applying this kind of thinking about systems, services, or infrastructures in current digital culture implies that just by experiencing, enacting, or engaging with them, we also articulate an initial prejudice in favor of them.

For our discussion here about a hermeneutics of algorithms, the questions of prejudices and authority are central. As algorithms are obviously not neutrally written and coded, so it is with us domesticating and engaging with algorithms in our everyday worlds: when trying to make meaning out of algorithms in our everyday lives in the sense of applying them, we bring our prejudices into play. The prejudices we encounter algorithms with are a product of the kinds of authorities (e.g. family and school) we have engaged with in our lives and our experiences with them. This means again that understanding algorithms is a meeting between what has formed and justified our prejudices (i.e. authority and tradition) and the algorithms loaded with interpretations and traditions too, implying an active role on behalf of us as interpreters.

Tradition

Moving on to tradition, Gadamer (1989) points out that

At any rate, our usual relationship to the past is not characterized by distancing and freeing ourselves from tradition. Rather, we are always situated within traditions, and this is no objectifying process – i.e. we do not conceive of what tradition says as something other, something alien. It is always part of us, a model or exemplar . . . (p. 282)

Tradition is always part of us and we are situated within traditions. Regarding algorithms as a kind of tradition, a kind of past, means that we should not see them as something alien. Engaging with algorithms in digital actions is not an engagement at a distance, that algorithms are objects out there to engage with. Algorithms are tradition, or are culture (Gillespie, 2016a; Seaver, 2017). They are part of us as they address us and we address them (Gadamer, 1989: 282). Therefore, understanding an understanding algorithms

. . . is to be thought of less as a subjective act than as participating in an event of tradition, a process of transmission in which past and present are constantly mediated. (Gadamer, 1989: 290)

Such a view on algorithms and our actions with them suggests a particular way of understanding them. Algorithms are not objects to be observed in action, to be processed, or as objects of critique. Hermeneutically, that is not how we come to understand algorithms. In everyday life, our knowledge of the existence and actions of algorithms in digital media points to how we understand them as part of our lives. They are part of us and they carry tradition. Writing our web pages using particular words in order to be captured by search engines and their algorithms is an example of how algorithms are part of us, something we count on and not something alien.

Furthermore, as for understanding algorithms, we learn from Gadamer (1989) that

When we try to understand a text, we do not try to transpose ourselves into the author’s mind but, if one wants to use this terminology, we try to transpose ourselves into the perspective within which he has formed his views. But this simply means that we try to understand how what he is saying could be right. (p. 292)

Thus, understanding an algorithm is not a matter of going into the mind of its designer. Rather, we try to position ourselves in the perspective that produced the ideas or motives shaping the design and activity of the algorithm. Placing ourselves in such a perspective implies, then, that we see the algorithm as trying to communicate something, that it could be right. We have no reason not to if our goal is understanding. However, this is not the same as subscribing to what the algorithm might communicate. To understand, accordingly, ‘. . .is not a mysterious communion of souls, but sharing in a common meaning’ (Gadamer, 1989: 292). As for understanding algorithms, this further implies that the goal is not a form of mental unity with the algorithm and its designer. Rather, it is to be part of a common meaning. What does that mean? It suggests that sharing in a common meaning, when we try to make sense of an algorithm, is not about locating the meaning in the mind of its creator. It is not about finding the meaning in an algorithm. Nor is it about the meaning the interpreter ascribes to the algorithm. It is about understanding the intersection of meaning as the algorithm and the interpreter enact it. For instance, when we are provided with a given search result based on our search keywords, we are not free to interpret this result in whatever way we like if we want to make hermeneutic sense of the result. The very meaning of the search result is a product of what we type in the search query and how that matches with the range of equivalent keywords and geo-location data appearing in the range of resources in the search engine. If my search query consists of the search string ‘HERMENEUTICS AND ALGORITHMS’, I am provided with a list of results that may or may not fit what I am looking for. This list of results is not just an answer to my query. The search algorithm has interpreted my query in a particular way because it is designed to do so. Therefore, in a hermeneutical sense, my understanding of the provided search result is not a judgment about whether or not the result is correct or articulates a degree of correctness as the usual measures of retrieval performance, recall, and precision, suggest (Lancaster, 1978). Rather, my understanding of the search result is the interplay between that result, itself a product of interpretation, and the meaning I ascribe to the result. I may object to the result that it was not what I was looking for because I do not know how to apply it to my situation. However, this is still an act of understanding, not a judging of the degree of correctness.

With the idea of understanding as ‘sharing in on a common meaning’, Gadamer presents a critique of the classic hermeneutic notion of the hermeneutic circle. According to classic hermeneutic theory, Gadamer (1989) argues, ‘. . . the circular movement of understanding runs backward and forward along the text, and ceases when the text is perfectly understood’ (p. 293). It is a view of understanding ‘. . . by means of which one places oneself entirely within the writer’s mind and from there resolves all that is strange and alien about the text’ (Gadamer, 1989: 293). With reference to Heidegger, Gadamer (1989) rejects this view on the hermeneutical circle by reminding us that

. . . the understanding of the text remains permanently determined by the anticipatory movement of fore-understanding. The circle of whole and part is not dissolved in perfect understanding but, on the contrary, is most fully realized. (p. 293)

Dissolvement of understanding would mean that what understanding is made up of is erased, while understanding as a full realization encapsulates what understanding is made up of, that is, fore-understanding of the interpreter and tradition. This is necessary in order to maintain understanding as ‘sharing in on a common meaning’. Therefore, the hermeneutic circle is,

. . . neither subjective nor objective, but describes understanding as the interplay of the movement of tradition and the movement of the interpreter. (Gadamer, 1989: 293)

Such kind of interplay is another way of saying that we are part of tradition as interpreters. However, being part of tradition does not mean being determined by tradition in a teleological way. It means that interpreters also produce tradition when involved in acts of understanding. Therefore, the meaning we confront texts with is not solely our own:

The anticipation of meaning that governs our understanding of a text is not an act of subjectivity, but proceeds from the commonality that binds us to tradition. (Gadamer, 1989: 293).

Again, understanding a text is not a subjective activity, but is a product of interplay between tradition and the interpreter, ‘the commonality that binds us to tradition’. In order to understand something or someone, we must have something in common and that commonality is secured by tradition. Likewise, we can say with algorithms if we start to see them as part of tradition like any other cultural product. Algorithms grow out of a tradition of computation, calculation, and software; a tradition that is not an auxiliary to modern digital culture, but is in a range of ways forming our social and cultural interactions with digital media through computational processes. This tradition is what comprises an algorithmic culture and it shapes ‘. . . the habits of thought, conduct, and expression that arise in relationship to those processes’ (Hallinan and Striphas, 2016: 119); that is, the presence of algorithms produces a corresponding mode of action in us; actions that are computational in the sense that we make our activities algorithmically recognizable (Gillespie, 2014, 2017). Gadamer (1989) further underscores the idea of how commonality binds us to tradition when he argues that

. . . this commonality is constantly being formed in our relation to tradition. Tradition is not simply a permanent precondition; rather, we produce it ourselves inasmuch as we understand, participate in the evolution of tradition, and hence further determine it ourselves. (p. 293)

Here, Gadamer points to the necessity of acknowledging ourselves as active makers and shapers of tradition, not just products of tradition. Thus, algorithms as tradition are not a permanent precondition. We produce them ourselves ‘inasmuch as we understand’, for instance, when circumventing them (Gerrard, 2018). With this, Gadamer (1989) concludes about the hermeneutical circle that it is not a methodological one, but rather that it describes ‘. . . an element of the ontological structure of understanding’ (p. 293), thus pointing to understanding as a matter of ontology, a mode of being, rather than a matter of procedure or method. From here, Gadamer (1989) moves on to his notion of the ‘fore-conception of completeness’ (pp. 293–294). This is a helpful concept in a discussion about understanding (as a mode of being) algorithms.

Fore-conception of completeness

The fore-conception of completeness goes back to Gadamer’s contention that an author and her or his text could be right (‘we try to understand how what he is saying could be right’), and that a text is always intelligible in all of its signifying aspects. The fore-conception of completeness states that

. . . only what really constitutes a unity of meaning is intelligible. So when we read a text we always assume its completeness, and only when this assumption proves mistaken – i.e., the text is not intelligible – do we begin to suspect the text and try to discover how it can be remedied. (Gadamer, 1989: 294)

What constitutes a unity of meaning has of course been questioned by both post-structuralist and deconstructivist thought. However, if we accept that algorithms are a unity of meaning, no matter if and how they might mutate or are changed along the way, we also have to assume their completeness in our particular interaction with them. In doing this, we acknowledge algorithms as being intelligible. As social beings in a digital environment, surrounded and saturated by a variety of automated communication infrastructures, we are forced to see algorithms as being intelligible and assume their completeness in our efforts to make sense of these very infrastructures (Couldry and Hepp, 2017). As social beings, we always try to make sense of the particular realities we are part of and in doing so we ascribe meaning to these realities. Commencing such acts of meaning-making with the assumption that these realities are false, incomplete, and not intelligible in terms of meaning, would hinder us in understanding anything at all. Achieving understanding here ‘. . . depends on understanding the content’ (Gadamer, 1989: 294); that is, understanding what has meaning. Therefore,

The fore-conception of completeness that guides all our understanding is, then, always determined by the specific content. Not only does the reader assume an immanent unity of meaning, but his understanding is likewise guided by the constant transcendent expectations of meaning that proceed from the relation to the truth of what is being said. (Gadamer, 1989: 294; italics added)

What the fore-conception of completeness tells us is to acknowledge that in understanding a specific sort of content, we set out in such an act of understanding from the premise that what is being communicated (the content) is true and understanding means understanding specific content and not another’s (e.g. an author’s) opinion. As a result, understanding algorithms in the light of the fore-conception of completeness means an expectation and assumption that the meaning of algorithms and what they communicate (i.e. what they sort, filter, recommend, retrieve, etc.) is intelligible and true. We must be open to what the algorithms have to say, their specific content, because we are,

. . . fundamentally open to the possibility that the writer of a transmitted text is better informed than we are, with our prior opinion. (Gadamer, 1989: 294)

When algorithms are successful in predicting tastes or preferences and come up with, for example, music recommendations, then to some extent we can say that algorithms are actually ‘. . . better informed than we are’ because they are able to see patterns that humans are unable to (Hallinan and Striphas, 2016). However, when we come to realize that what the algorithm said was not true, then ‘. . . we try to “understand” the text . . . as another’s opinion’ (Gadamer, 1989: 294). In other words, the prejudice of completeness implies that what a text says, beyond expressing its meaning, ‘. . . should be the complete truth’ (Gadamer, 1989: 294). From this follows that understanding primarily means to ‘. . . understand the content of what is said, and only secondarily to isolate and understand another’s meaning as such’ (Gadamer, 1989: 294). Thus, understanding algorithms means to understand the content of what they communicate. We do not understand algorithms by trying to locate and detect the designer’s meaning or intention, practically an impossible task.

With the fore-conception of completeness, we have established a position from where to understand algorithms and what they communicate, in the first place, as true. However, according to Gadamer, to understand what is true is not to reproduce that moment of truth. This is the idea in romanticism where understanding is the reproduction of an original production. Gadamer (1989) argues,

Every age has to understand a transmitted text in its own way, for the text belongs to the whole tradition whose content interests the age and in which it seeks to understand itself. The real meaning of a text, as it speaks to the interpreter, does not depend on the contingencies of the author and his original audience. It certainly is not identical with them, for it is always co-determined also by the historical situation of the interpreter and hence by the totality of the objective course of history. (p. 296)

Thinking of an algorithm as a transmitted text helps us see how to understand an algorithm as a question of interpreting it in the light of our own situation. This interpretation does not depend on the designer of the algorithm and their original audience and intentions. The meaning of the algorithm is co-determined by the historical situation of us, the users, public, audiences, or consumers. Therefore, the meaning of algorithm ‘Not just occasionally but always . . . goes beyond its author’ (Gadamer, 1989: 296) and understanding is therefore productive rather than reproductive.

Productive understanding

Productive understanding means understanding in a different way:

Understanding is not, in fact, understanding better, either in the sense of superior knowledge of the subject because of clearer ideas or in the sense of fundamental superiority of conscious over unconscious production. It is enough to say that we understand in a different way, if we understand at all. (Gadamer, 1989: 296–297, italics added)

Arguing in the same way with algorithms, we can come to realize that when we understand, we understand them differently because we know, or are getting to know how to apply them to our own historical situation, in our own practices so to speak. This is exactly what is happening when we try to be captured by search engines using particular keywords, relying on their recommendations, or when trying to tweak algorithms in our various sorts of communicative activities (Bucher, 2017). We produce our understanding of algorithms when acting with them in terms of searching, tweaking, to be captured, or when we question their recommendation or trending list (Gillespie, 2016a). If understanding were reproductive, we would have to understand algorithms in line with the original intentions of the designers or developers, but how such an understanding is supposed to look is difficult to imagine. Since algorithms mutate or are molded along the way, understanding the original intentions, for instance, would not help when trying to tweak them. We have not understood them at all but just reproduced their historical meanings, as romanticism would have it:

Such a conception of understanding breaks right through the circle drawn by romantic hermeneutics. Since we are now concerned not with individuality and what it thinks but with the truth of what is said, a text is not understood as a mere expression of life but is taken seriously in its claim to truth. (Gadamer, 1989: 297, italics added)

Once again, we are reminded that there is a difference between understanding a text ‘as a mere expression of life’ and ‘its claim to truth’. The former is a psychological game and the latter is taking seriously what a text wants and has to say and we have to consider this as true in the first place if the goal is understanding. Gadamer here teaches us to focus on the text because the text is what we have access to, what is available to us when trying to understand it, not the author and their intentions. The same with algorithms. If our goal is to understand them, we have to take seriously their claim to truth in terms of their statement of some affair in the world. It is not the same as saying what they communicate is the truth, to subscribe to that truth, or truth as in correspondence theory. However, it is to acknowledge that algorithms address us in terms of recommendations, showing trends, predictions, or providing search results. These are what we have hermeneutic access to and can perceive, not their code, black box, or intentions. These are what we respond to, question, deliberate about, or feel uneasy about. The beginning of understanding is when we are addressed by something; the first condition of hermeneutics (Gadamer, 1989: 299). However, this condition and the text’s claim to truth are not fulfilled if interpretation is equal to understanding things historically as we give up on the text’s claim to truth. This entails that we have not found anything in the text intelligible to us as we do not know how to apply the text to our own situation. We have only discovered another person’s standpoint (Gadamer, 1989: 303). From here, we can move on to Gadamer’s key argument about understanding as application.

Understanding as application

The argument about understanding as application continues Gadamer’s critique of the author/intention-oriented and psychological form of hermeneutics articulated by, for instance, Schleiermacher, where application is separated from understanding.

Gadamer gives an example with understanding a law or a gospel. Neither of them are understood if they are understood historically. Rather, they must be concretized in their legal validity or saving effect by being interpreted in concrete situations different from their origin:

This implies that the text, whether law or gospel, if it is to be understood properly – i.e., according to the claim it makes – must be understood at every moment, in every concrete situation, in a new and different way. Understanding here is always application. (Gadamer, 1989: 309)

Understanding a text in a new and different way entails applying it in new ways in one’s own situation. Therefore, understanding is application. As for algorithms, this means that we not only understand them differently, when/if we understand, we understand algorithms in a new and different way because we apply them in new ways in concrete situations, and not as historical artifacts. A case in point here is Gerrard’s (2018) recent study of how online pro-eating (pro-ED) disorder communities are developing practices of circumventing hashtag moderation on social media. Here, the users embraced the algorithmic nature of various social media sites and through this managed to communicate about pro-ED by means of signaling, even though the hashtags classifying pro-ED content were moderated by the sites.

However, understanding as application does not mean that the role of the artifact is irrelevant. We cannot perform a play or read a poem in some sort of performance and apply it to own our situation without understanding the original meaning. However, we have not gone very far in understanding if it stops here:

No one can stage a play, read a poem, or perform a piece of music without understanding the original meaning of the text and presenting it in his reproduction and interpretation. But, similarly, no one will be able to make a performative interpretation without taking account of that other normative element – the stylistic values of one’s own day – which, whenever a text is brought to sensory appearance, sets limits to the demand for a stylistically correct reproduction. (Gadamer, 1989: 310)

Once again, Gadamer emphasizes the relationship between the meaning of the historical artifact and the contemporary interpreter bringing in her or his time and place in the process of understanding. In modern terms, we may say that Gadamer actually recognizes the interactive nature of understanding. Such kind of interactive-ness is precisely also the nature of algorithms and our dealings with them. Therefore, understanding is at once productive and an application in that a different interpretation offered is based on applying the artifact in a concrete situation. In addition, therefore, understanding is not a unification of the creator and the interpreter. Understanding consists,

. . . in the fact that no like-mindedness is necessary to recognize what is really significant and fundamentally meaningful in tradition. We have the ability to open ourselves to the superior claim the text makes and to respond to what it has to tell us. (Gadamer, 1989: 311)

In understanding and interpreting algorithms, this quote is an important reminder. We, the interpreters or users of algorithms, should not forget that we actually have and are obliged to open ourselves to the claims and statements algorithms make in order to make meaning out of them by responding and applying them in our own everyday actions. In fact, what many users experience in algorithmic actions is no like-mindedness at all; yet, they have the ability of making meaning out of algorithms in terms of encoding, tweaking, circumventing, aligning with, or questioning the actions of algorithms.