2. Artificial intelligence

Recent innovations in both hardware and software have advanced AI technology dramatically in the last few years. The most important ones being a dramatic increase in processing power using specialised computational hardware (such as computer processing units, graphics processing units, tensor processing units,Footnote ²  and even early-stage qubit computing) (Zhu et al. Reference Zhu, Yu, Xu, Chen, Dustdar, Gigan, Gunduz, Hossain, Jin, Lin, Liu, Wan, Zhang, Zhao, Zhu, Chen, Durrani, Wang, Wu, Zhang and Pan2023) and groundbreaking discoveries in machine learning using artificial neural networks (OpenAI 2023). This rapid development of AI has resulted in unprecedented breakthroughs in image processing and recognition (OpenAI 2024b), natural language processing (OpenAI 2023), audio manipulation (OpenAI 2024c), and autonomous systems more generally (OpenAI 2024d).

It will be useful moving forward to define narrow AI and general AI. Narrow AI commonly encompasses AI systems that are built for specific tasks or applications that are well defined. They execute precise functions within a limited domain that cannot be generalised to tasks beyond that domain, although general AI is commonly understood as having fully developed human cognitive capabilities (Sheikh Prins and Schrijvers Reference Sheikh, Prins, Schrijvers, Sheikh, Prins and Schrijvers2023).

ANI can be thought of as a task-specific optimised system in its relevant domain, whereas AGI can be said to display general cognitive autonomy as one would expect from normal epistemic agents (i.e. humans). AGI, as of writing this, does not exist. ANI, however, is rapidly integrating itself into our lives, and although image and audio generative AI systems have gotten increasingly more attention, LLMs, such as DeepSeek, Gemini, Claude, and ChatGPT,Footnote ³ have become almost synonymous with AI.

LLMs are computational devices used for natural language processing that are able to generate human-like text and complete other language-related tasks with high accuracy (Kasneci Sessler Küchemann and Bannert Reference Kasneci, Sessler, Küchemann and Bannert2023: 1). They are designed to generate sequences of words, code, or other data, from a source input (commonly referred to as a prompt) (Floridi and Chiriatti Reference Floridi and Chiriatti2020). The most recent versions of LLMs at the time of writing, such as Anthropic Claude 5.x, Google Gemini 3.x, DeepSeek-V4, and ChatGPT-5.x, embody both the fantastical advancements of AI, as well as its inherent limitations. Although these advanced LLMs are able to instantly generate coherent and contextually relevant responses, easily passing the Turing test (Mei et al. Reference Mei, Xie, Yuan and Jackson2024), they also generate perplexing non-sequiturs, tell brazen lies, and misunderstand the simplest of tasks (Hicks, Humphries, and Slater Reference Hicks, Humphries and Slater2024).

Furthermore, even when LLMs produce coherent and accurate text, they are often unable to provide evidence for the claims they make. Instead, they make up imaginary sources and facts. This tendency of LLMs to fabricate evidence has been called ‘AI hallucinations’ (Hicks, Humphries, and Slater Reference Hicks, Humphries and Slater2024: 38).Footnote ⁴ LLMs do not reflect on the text they produce, which can make these hallucinations (or confabulations) snowball,Footnote ⁵ generating further errors (Zhang et al. Reference Zhang, Press, Merrill, Liu and Smith2023: 1). Interestingly, in some cases, when LLMs are asked to justify their previous hallucinations, they generate false claims that they can recognise as being incorrect when they are presented with the same false claims in a separate interaction session (Zhang et al. Reference Zhang, Press, Merrill, Liu and Smith2023: 2). Recently, Hicks, Humphries, and Slater (Reference Hicks, Humphries and Slater2024) have argued that ChatGPT is a bullshit machine, and instead of hallucinations, ChatGPT’s erroneous responses should be called bullshit because the terms ‘hallucinations’ and ‘confabulations’ suggest that ChatGPT can perceive (if it can hallucinate), or rely on, memory in a traditional sense (if it can confabulate) (Hicks, Humphries, and Slater Reference Hicks, Humphries and Slater2024: 8). Furthermore, both terms suggest that ChatGPT is generally attempting to convey accurate information, when it is simply predicting the next word in a sentence.

Although we will be comparing AI attempts at a performance with human attempts to a degree, current LLMs only indirectly and incidentally track truths, as the design function of an LLM is to produce text that sounds plausible.Footnote ⁶ However, it is not clear what follows from this concession. Consider, for example, that the relevant design function of paradigmatic epistemic competences like perception and memory is to help survival and more generally evolutionary fitness of the organism. This is not an epistemic aim, but an evolutionary adaptive aim keyed to organism fitness. Granting this fact, however, does not stand in tension with the idea that our perceptual and memory functions reliably deliver accurate information.

The compatibility of epistemic faculties having non-epistemic design or etiological functions, thus non-epistemic function-generated aims, offers a vantage point to reassess what it means for LLMs to be optimised for producing plausible-sounding content. If that is their design function, they might have a function-generated aim that is not epistemic, but that in and of itself does not necessarily preclude them from having the kind of reliable connection with truth (a question determined by seeing how well they actually deliver true information) that is demanded by competence, at least in appropriate conditions. For instance, if we specify that the kind of situational component of competence that is normal for LLMs to operate includes largely reliable training data, then it looks like we can make sense of a virtue-theoretic competence structure,Footnote ⁷ forming the basis of what looks like reliable artificial competence, even if we continue to grant that LLMs are intentionally designed to produce plausible human-like text.Footnote ⁸

Note that the inner workings of AI can be akin to ‘black boxes’, so their justification for what led to a given prediction might not be interpretable by humans. Consider that there might be thousands of variables that contribute significantly to a single prediction. Even if they could all be inspected and the way the AI system decided to weigh each variable was accessible,Footnote ⁹ it is not reasonable to expect users to understand why the AI system came to a prediction (Ribeiro Singh and Guestrin Reference Ribeiro, Singh and Guestrin2016: 1137). To combat this, AI can be imbued with explicability,Footnote ¹⁰ which entails explainability and interpretability, often referred to as explainable AI, or explainable artificial intelligence (XAI). This enables humans to retain a level of intellectual oversight, generally by providing access to the reasoning behind the decisions and predictions made by the XAI, in the form of comprehensible explanationsFootnote ¹¹ that use concepts that we understand (Cappelen and Dever Reference Cappelen and Dever2021: 25; Longo et al. Reference Longo, Brcic, Cabitza, Choi, Confalonieri, Del Ser, Guidotti, Hayashi, Herrera, Holzinger, Jiang, Khosravi, Lecue, Malgieri, Páez, Samek, Schneider, Speith and Stumpf2024). Currently, many deployed AI systems (including LLMs) are not explainable, and no known methods exist to make them so without losing either information or accuracy (Rao Reference Rao2025: 47). The opaque nature of AI systems makes XAI development focused on the explainability and interpretabilityFootnote ¹² of AI vitally important to our understanding of present and future AI systems (Cappelen and Dever Reference Cappelen and Dever2021: 26).Footnote ¹³ The literature on AI mirrors epistemology in some important ways. Sosa (Reference Sosa2021) gives an example of an eye-exam, where one starts to incorporate guessing as the letters get smaller. In his case, the guesses are correct, and reliably so (Sosa Reference Sosa2021: 144). Guessing reliably has obvious parallels to LLMs. We will argue that, in both cases, these guesses are alethic affirmations, but they are not judgements, because the guesser aims at truth without aiming at aptness. The guesser’s affirmation might be apt (correct because of manifested competence), but it is not aptly apt, because that would require aiming at aptness itself, guided by a second-order grasp that one’s affirmation would be apt. This distinction between alethic affirmation and judgement will structure our account of the ANI/AGI divide.

Furthermore, the literature on AI is concerned with performances and domains, which fits well within a virtue epistemological framework such as the one Sosa has developed, most recently in his book Epistemic Explanations (Reference Sosa2021). Additionally, Sosa’s approach to competences takes them to be special cases of dispositions, i.e. dispositions of agents to perform well in a given domain. Thinking about competences as special cases of dispositions is advantageous to us while writing about AI, as machines can possess such dispositions in a functional-teleological sense, that is, dispositions to attain aims functionally provided by their design and training, which is independent of whether they possess consciousness or the kind of biological architecture on which human cognition supervenes. Before we consider AI competence from a virtue theoretical perspective, we need to elaborate on some of the key ideas found in Sosa’s framework.

3. Virtue-theoretic competence

Sosa’s telic virtue-epistemological view argues that epistemic normativity should be understood as one form of telic normativity and that performances can be understood as attempts aimed at some goal. An attempt can thus become an achievement when the attempt is successful because of sufficient competence (Sosa Reference Sosa2021: 18). As we have already established, competences are a special kind of dispositions, namely dispositions of agents to perform well in a given domain. To assess whether an agent’s performance is good, Sosa proposes the following structure:

Aptness is a necessary and sufficient part of achievements, as without it one can be accurate and adroit because of luck, allowing for the possibility of lucky successes such as Gettier cases constituting cases of knowledge. An oft-cited example to grasp these distinctions is of an archer. The archer shoots at a target and their shot hits the target. The archer’s performance is accurate because they hit the target. It is apt iff the archer hit the target because their performance was adroit, i.e. the archer manifested their competence when they hit the target.

An important question is what it means to manifest competence. In the archer example, the performance can be said to be completely competent if the shot manifested the archer’s intrinsic archery skill (constitutional competence),Footnote ¹⁴ the archer was in good shape when shooting (inner competence), and the situation when shooting was appropriate (Sosa Reference Sosa2010: 465). Competences of an agent can thus be seen as dispositions they have to perform well and are comprised of the agent’s skill, shape, and situation (Sosa Reference Sosa2010: 465).

To further clarify, if the archer is to perform competently, they would first of all have to have the skills required to hit their target reliably enough when they are in proper shape and properly situated, as if they would not possess the necessary skills to do so they are incapable of competently hitting the target as luck would be the salient reason for them hitting it.

Second, they would have to be in a position to access those skills, i.e. the agent must be in shape. The details of what exactly constitutes good shape can vary in practice, but one can imagine that in the archer’s case, being in shape includes being sober enough to utilise their skills, keep their eyes open, be mentally unperturbed, alert, and so on. Note that these requirements depend on the notion that without them the archer could not perform up to the intrinsic skill they possess, e.g. if they could reliably shoot their target while drunk with as much skill as if they were sober, then the state of being drunk would be compatible with the shape needed to be competent.

Finally, the situation must be appropriate for the archer to perform. Here, appropriate can be understood as specifying the range of conditions under which the archer’s competence is properly attributable (i.e. conditions where their skill is meant to apply). In this specific case, the relevant range of situations might include, for example, wind levels and lighting conditions within the thresholds set by archery tournaments.

Now that we have explained what an apt performance entails, we must make a further distinction between apt performance, and aptly apt performance. Even when a performance is apt, in the sense that the agent is accurate because of adroitness, its aptness is not necessarily itself apt. For a performance to be aptly apt, the agent must be attempting not only to be accurate but apt. This becomes an especially meaningful distinction when we move from athletic performances to intellectual performances, where it maps onto the distinction between alethic affirmation and judgement.

When we apply the AAA structure to epistemic performances, we find certain parallels between the epistemic agent and the archer. On Sosa’s (Reference Sosa2021) view, the relevant epistemic performance is the alethic affirmation, an affirmation made in the endeavour to get it right on a given question (pp. 23–24). Animal knowledge can thus be understood within the AAA structure as an apt alethic affirmation: the agent affirms, the affirmation is accurate (true), adroit (issued from competence), and apt (true because of the manifested competence).Footnote ¹⁵ Once again, we use the SSS structure to see what competence in this example entails. First, the epistemic agent must possess the relevant constitutional epistemic skills, such as cognitive abilities (seat). Second, they must be in a good epistemic state to apply those skills, e.g. by being sober, awake, and alert (shape). Third, the situation must fall within a range that is not epistemically hostile, for example the epistemic agent is not being deceived by tricky lighting or illusions (situation).

To finalise our discussion of the virtue-theoretic framework that will be used to analyse AI competence, we need to categorise the kinds of knowledge that are derived from this framework. In Sosa’s (Reference Sosa2021) hierarchy, the lowest level of knowledge is animal knowledge, constituted by an apt alethic affirmation, one whose truth is due to the first-order competence of the thinker. This includes sub-credal cases like Sosa’s eye-exam subject, who reliably gets the letters right at a row where they take themselves to be guessing (Sosa Reference Sosa2021: 144). Above this level is reflective knowledge full well, constituted by a fully apt judgement: an alethic affirmation whose aptness is itself attributable to the agent’s meta-competence, that is a second-order sensitivity to when one’s first-order affirmations would be apt (Sosa Reference Sosa2021: 186; Carter Reference Carter2018: 285).Footnote ¹⁶

4. Artificial competence

Sosa (Reference Sosa2021: 23) extends the telic structure for evaluating performances to systems whose aims are merely functional-teleological, which includes ANI systems. Before turning to the more contested case of LLMs, however, it is worth constructing a clear view of ANI more generally. An AI chess engine’s constitutional competence in the domain of chess is well defined and frequently exceeds the best human performers. Its situational requirements are tractable; a properly configured engine with sufficient computational power, given a legal position, manifests its disposition to evaluate the position and find the strongest move. The engine’s aim here is thus not an epistemic aim directed at truth. Importantly, it has no capacity to step back and assess whether its evaluation function is inadequate for the position at hand or whether its opening book is being exploited. As the chess engine is rarely mistaken, it generally performs aptly within the domain, but it does not assess its own aptness.

If AI outputs can manifest the AI’s competence to some degree, then it seems fruitful to examine AI competence further using the SSS model. Recall that SSS stands for seat, shape, and situation. In many ANI systems, we find that the seat, or constitutional competence, is not only in place, but oftentimes exceeds the highest performing humans in the relevant domain. What about ANI systems that are able to reach across domains? LLMs are capable of writing poetry and speeches, fluent in nearly every programming language, and have access to immense amounts of data, which they can quickly sort through to generate responses to prompts. One way to answer whether LLMs have constitutional competence is by examining whether the LLMs would be able to perform reliably enough Footnote ¹⁷ if they were in proper shape and in an appropriate situation (Sosa Reference Sosa2010: 473).Footnote ¹⁸

So, let us examine LLMs’ shape and situation and see how applicable they are to AI before we assess LLMs’ constitutional competence. The shape, or inner competence, of ANI in general, including LLMs, seems unlike the inner competence humans possess because many of the examples generally used to describe this kind of competence, such as sobriety and alertness, do not apply to ANI. We can still imagine that a proper shape for an AI might consist of being configured correctly with access to sufficient computational infrastructure to take advantage of its capabilities. This means the AI would not only have the appropriate nodes on the neural network after the relevant machine learning processes and the computational power to utilise them, but also the kind of configuration that enables it to correctly determine the appropriate response. To illustrate, think of an author that, when sober, can distinguish between fact and fiction, but when drunk they confuse the two and claim something as true that only happened in their book. In much the same way, LLMs have access to a vast amount of data, some of which is factual and some of which is fictional, and their shape determines whether they are configured to make proper use of the skills they possess.

Another way in which AI’s shape could be compromised is when it gets stuck in a local minimum, unable to successfully perform. For clarification, think about a chess engine trying to find the best move. It sees the potential moves on the board and immediately dismisses the moves that do not seem promising. One such move is sacrificing the queen for apparently no compensation. As it turns out, that queen sacrifice is the best move at a very high depth. Unfortunately, the ANI is stuck with the candidate moves it initially decided on, trying to figure out which of them is most promising. In the case of LLMs, their structure allows them to access, transform, and employ, whatever data are needed. However, without a protocol to clearly distinguish between the two, it is not in proper shape to successfully perform in the epistemic domain of transmitting truths.

Regarding the range of appropriate situations in which an LLM’s abilities are relevant, I suggest that it is defined by the quality and type of data used to train the LLM. If the situation is appropriate, i.e. the data fall within the type and quality the LLM’s training is meant to accommodate, and the LLM is properly configured, it is clearly capable of performing well, with the caveat that their performance must be within their domain of expertise. In sum, ANI systems can possess complete competence if we define shape in terms of a system’s ability to access and apply the skills it possesses, without having consciousness be a necessary property.

Even when LLMs reliably make true predictions within their domain of competence, the question of whether they are positioned to do so, whether their SSS are arranged to facilitate aptness, is itself an epistemic question. The requirements for reflective knowledge are inherently epistemic, regardless of whether the first-order competence is situated in an epistemic domain. A baseball player that is attempting to perform in the domain of baseball is not aiming at truth, but the epistemic domain is still relevant when they consider whether they are triple-S competent, as epistemic reflection is needed for this kind of a second-order competence. This does not mean that epistemic domains are uniform in nature, as standards between epistemic domains differ greatly depending on the setting (Sosa Reference Sosa2021: 14).

Such second-order competence is not present in LLMs either, even though their aims are epistemic to a degree. When an LLM generates a justification for its prediction, it is performing another first-order task. It is predicting what a good justification would look like given its training data, but it is not assessing whether the conditions of its own competence were such that its original performance was apt. Judgement, on Sosa’s account, requires aiming at aptness (not merely at truth) (Sosa Reference Sosa2021: 24–25), along with a guiding presupposition that one’s affirmation would be apt (Sosa Reference Sosa2021: 144–145). An LLM’s evaluation of its own predictions lacks both, as it aims at truth on a new question (namely, ‘was my prediction accurate?’) without any presupposition concerning the aptness of the original affirmation. The second-order competence required for judgement is not a separate performance but a condition on the original performance itself (Sosa Reference Sosa2021: 25, 186), and no amount of recursive self-monitoring can provide that.Footnote ¹⁹

At a certain point ANI might become advanced enough that its predictions resemble those of Norman the clairvoyant, who in BonJour’s (Reference BonJour1980) thought experiment possesses a perfectly reliable clairvoyance faculty but lacks any reflective endorsement of that faculty. We would have little to no understanding of why the ANI makes a given prediction, but we could still be rather certain of its accuracy.Footnote ²⁰ No matter how reliable such an ANI becomes, no increase in first-order competence can provide the second-order competence required for judgement. The difference between Norman and ANI is that although they are both making predictions without reflective endorsement, only the ANI, especially when equipped with high interpretability and transparency, can provide post hoc rationalisations supported by evidence and reasoning on account of the data it holds and the neural network processing it performs. However, this should not be mistaken for second-order epistemic reflection.

Recall the initial distinction between ANI and AGI and their defining characteristics, namely that ANI are specialised AI systems that perform tasks within a limited domain or a set of domains, whereas AGI have human-like cognitive capabilities across a diverse range of domains. Now, one might try to distinguish ANI from AGI by claiming that ANI performances are simply not apt, whereas AGI performances would be apt. But, as we have seen, this is too hasty. To illustrate, see the following example from Sosa (Reference Sosa2021, Reference Sosa2010) about Simone the fighter pilot, who could easily not be in a real cockpit but in a near-perfectFootnote ²¹ simulation:

Simone’s shots are apt: her shooting competence manifests in the accuracy of her shots. But, as Sosa points out, ‘what of her intellectual shots, her judgments and beliefs?’ (Sosa Reference Sosa2010: 468). Imagine that Simone forms the alethic affirmation that she successfully shot a target, and her affirmation is accurate and adroit. Sosa asks whether her alethic affirmation can also be apt, given that she could easily have been placed into a simulation (Sosa Reference Sosa2021: 169–170; Sosa Reference Sosa2010: 468). It is difficult to deny that it can be, given how clearly apt her physical shot is despite the fragility of her situation (Sosa Reference Sosa2010: 468). Simone’s ‘competence manifests in the accuracy of her shot’, and the same holds for her epistemic performance (Sosa Reference Sosa2010: 468). On Sosa’s recent view (Reference Sosa2021: 169–170), Simone can even attain reflective knowledge (but not secure knowledge).Footnote ²² Nevertheless, the case shows how one can possess first-order competence while lacking the situational security that higher grades of knowledge demand. We can imagine ANI systems existing in a kind of simulation, trained on datasets that cannot fully represent the world, leaving them incapable of recognising the incompleteness of their information.Footnote ²³ ANI systems, unlike Simone, cannot reach the level of reflective knowledge, because they lack the second-order competence required for judgement.Footnote ²⁴

Sosa divides the various epistemic competences into two main categories: non-global competences, which host seemings, and global epistemic competences, which are basic judgement-forming competences (Sosa Reference Sosa2021: 148).Footnote ²⁵ Global epistemic competences differ from non-global competences in the sense that they are used to determine how to judge, all things considered.Footnote ²⁶ This means that nearly anything could be relevant to any epistemic attempt (given proper stage-setting). Sosa claims that this kind of ‘holistic competence is global in that it is required in properly making any judgment or forming any belief’ (Sosa Reference Sosa2021: 148).

To demonstrate the difference between global and non-global competences, Sosa gives an example of Norm, the normal perceiver, and Abnor, the mental ward patient (Sosa Reference Sosa2021: 148).Footnote ²⁷ On any given day when Abnor wakes up, he could have been experimented on which would deprive him of some, or all, epistemic competences that day (both non-global and global). If he would wake up without non-global epistemic competence, then he would not be capable of reflective knowledge unless someone would tell him that his competence, be it his shape or situation, was absent (and he would be in a position to make proper use of that information) (Sosa Reference Sosa2021: 151). If he would wake up without global epistemic competence, then there is no way for him to properly reflect on any tell-tale signs that this has happened, and he would lack the ability to assess its presence (Sosa Reference Sosa2021: 149, 151).Footnote ²⁸ According to Sosa, Norm is positioned to acquire both animal and reflective knowledge, whereas Abnor is limited to animal knowledge (Sosa Reference Sosa2021: 151).

For an AI to possess the sort of holistic global epistemic competence, Sosa identifies, it would need to be able to make judgements all things considered. One might think that Norm and Abnor show ANI could potentially attain reflective knowledge if its architects were to inform it that its competence had been compromised. But this would not amount to reflective knowledge. Even if the ANI were so informed, it would still lack the global epistemic competence (which entails second-order competence) that would be required to integrate that information into a revision of its own epistemic standing. And in any case, the architects would be the ones reflecting, not the ANI itself.

When it comes to the loss of global competence, we see that ANI is in the same or worse position as Abnor. At best it can only tell when its global epistemic competence has not been compromised with no way of knowing when it has, at worst it cannot even tell when it has not been compromised, because it lacks the second-order competence to assess the status of its own global competence (Sosa Reference Sosa2021: 152). For an example, consider that ANI systems can hallucinate while being unaware they are doing so because they lack second-order competence. In that case, adding data does not make up for this deficit, as ANI cannot verify that data to be correct, and is incapable of reflecting on that unreliability. ANI’s deficit is therefore constitutional; it lacks second-order competence not because of an unfavourable environment, but because its seat does not include the capacity to assess its own first-order competence.

I propose the following virtue-theoretic definitions of ANI and AGI from competence: