AI Sprints
David M. Berry
 |
| Figure 1: Gemini generated image |
The emergence of Large Language Models (LLMs) presents a remarkable opportunity for humanities and social science research. As I have argued elsewhere, we face an “algorithmic condition” where computational systems increasingly mediate not just our analytical tools but how we understand culture and society more generally (Berry 2025). This also opens the possibility for new methodological approaches capable of exploiting AI’s augmentative capacities whilst maintaining the critical reflexivity that is key to humanistic inquiry. In this article I want to suggest that “AI sprints” might be a possible method for using those capacities, building upon my recent experiments in critical code studies where I have developed an educational application through iterative dialogue with generative AI (Berry 2025b). That experiment, which adapts the idea of “vibe coding” following Karpathy’s (2025) neologism, revealed the suggestive potential of human-AI collaboration when undertaken in tight loops of iterative development. Here I want to outline this approach as a broader research method, one that adapts the data and book sprint idea whilst acknowledging the profound transformations generative AI introduces (see Berry et al 2015).
As Rogers (2013) argues, digital methods allow us to “follow the medium”, studying digital objects through their native protocols and using the strengths of the digital in order to develop research projects and methods. The AI methods I propose here extend this principle, using computational systems not merely to process digital traces but to analyse materials that traditionally required extensive manual coding or transcription. However, this extension risks “cognitive delegation”, which I describe as the offloading of interpretive judgement to systems that lack human understanding. A critical methodology must therefore operate in both technical and reflexive registers, utilising AI’s strengths whilst critiquing its epistemological assumptions and outputs.
From Data Sprints to AI Sprints
The “data sprint” methodology emerged from the notion of book sprints which were an agile practice for the rapid writing of bookish objects (see Berry et al 2015). In digital methods research, this has been taken further through the Digital Methods Initiative (DMI) at the University of Amsterdam. Traditional data sprints typically involve teams with mixed technical and domain expertise, intensive time-bounded work creating productive constraints, iterative cycles between data collection, processing, visualisation, and interpretation, production of “intermediate objects” such as graphs, networks, and timelines that mediate between raw data and analytical claims, and public presentation of provisional findings (Rogers 2013).
Figure 2: the development of sprint forms: book sprint (2008), data sprints (2014) and AI sprints (2025)The DMI data sprint approach relies particularly on what Rogers calls “issue mapping” and “controversy analysis”, tracing how matters of concern circulate through digital media. These sprints have the potential to democratise digital research by creating structured environments where non-technical researchers can engage computationally-mediated inquiry through collaboration with technical partners. As Marres (2015) observes, this collaborative division of labour can be useful in digital sociology by distributing expertise across teams.
This potential shift in collaborative expertise towards individual researchers working with LLMs raises important questions about the social epistemology of digital research. Whilst AI augmentation may democratise access to computational methods by reducing dependence on technical collaborators, it might also threaten the productive friction that emerges when researchers with different epistemic commitments must negotiate shared analytical approaches or methods. The data sprint’s strength lay not simply in distributing tasks but in requiring continual dialogue between differently positioned knowers. AI sprints risk losing this dialogical dimension, replacing it with what might be called a monological augmentation where a single researcher’s interpretive approach remains unchallenged by alterity. This suggests AI sprints might work best when combined with traditional collaborative practices rather than replacing them entirely, using LLM augmentation to extend individual analytical capacity whilst maintaining spaces for collective interpretive work.
AI sprints thus represent both a continuation and a break with the data sprint approach. We might also add that AI sprints respond to what might become a potential “methods crisis” in digital humanities and social sciences. This crisis emerges as LLMs become increasingly capable of performing many tasks that previously required bespoke computational tools, potentially undermining the rationale for developing field-specific digital methods expertise. If an LLM can generate code to analyse Twitter data, map controversies, or extract themes from interviews without requiring specialised scraping tools, network analysis software, or coding schemes, what is likely to become of the methodological fields built around such tools? AI sprints address this crisis not by abandoning digital methods but by integrating LLM capabilities within established methodological approaches that maintain disciplinary traditions that privilege interpretation and critique. The continuity lies in AI sprints emphasis on time-bounded intensive work, iterative refinement through feedback loops, production of intermediate mediating objects, and commitment to making research processes open and teachable. The difference emerges through the changed nature of human-computer collaboration. Where data sprints distribute expertise across human teams, using technical researchers to handle scraping and processing, and domain experts guide interpretation, AI sprints potentially collapse this distribution towards individual researchers working in tight loops of dialogue with LLMs that handle each of these previous specialisms.[1]
This is not simply automation but what I term “cognitive augmentation”, where the LLM extends research capacity whilst the human researcher maintains top-level control over the inquiry (Berry 2025b). As I discovered through developing my own software application using Google Gemini Pro, I found that this augmentation operates through three distinct modes which I call: (1) cognitive delegation, (2) productive augmentation, and (3) cognitive overhead. These cognitive modes are important for understanding the best ways of approaching an AI sprint method and determining when collaboration can be successful.[2]
Drawing on traditions established through FLOSS Manuals and the Book Sprint method, particularly through the work of Adam Hyde, the idea is that AI sprints build on these approaches to developing rapid, collaborative knowledge production. As detailed in On Book Sprints (Berry et al 2014), these intensive writing workshops emerged from the free software documentation movement, developing patterns for extractive knowledge capture and generative concept development. The Book Sprint method demonstrated that compressed timeframes, when properly facilitated, could produce high-quality outputs through intensive collaboration. AI sprints extend this process into computationally-supported research, replacing some human collaborators with LLM interlocutors whilst maintaining the emphasis on intermediate objects, iterative refinement, and critical reflexivity that characterises Book Sprint practice.
The Three Modes of Cognitive Augmentation
My experimentation in (extended) critical code studies revealed that cognitive augmentation is not a single approach but operates through three distinct modes which I now outline in more detail (see also Berry 2025b). Understanding these modes is helpful for designing effective AI sprint practices. Here I further develop this framework by connecting it to broader questions of computational ontology and the politics of abstraction.
Cognitive Delegation
Cognitive delegation represents the primary issue to be avoided in AI-augmented research. This occurs when researchers offload interpretive judgement to the LLM, accepting its analytical framings uncritically. As I observed in Phase One of my software development experiment, the ease of offloading problems to the LLM created a false confidence in using the AI. I had assumed the LLM could parse complex, semi-structured documents because previous prompts had succeeded with simpler tasks. The LLM, for its part, generated increasingly complex solutions without signalling that the approach was actually unworkable. In effect, the AI produced increasingly buggy, broken code, which it nevertheless cheerfully recommended to me to solve my problem.
This dynamic reveals how cognitive augmentation can easily become cognitive delegation so that rather than improving research project decisions, the LLM’s positive responses encouraged me to persist with what turned out to be an unworkable strategy. In research contexts this becomes particularly hazardous as it wastes considerable amounts of time and takes the research down lines that will be fruitless. Where buggy code fails visibly, the analysis may nonetheless appear superficially plausible. The researcher must therefore take a stance of critical distance from the LLM and its apparent capabilities, recognising that its willingness to generate a solution or digital tool does not indicate the approach is sound.
This process of abstraction operates, following Philip Agre (2014), through both “capture” in the acquisition of data and in the creation of ontologies that model the research domain (see Berry 2017). The danger in cognitive delegation lies not simply in offloading tasks to the LLM but in unconsciously accepting the epistemological and ontological presumptions embedded in its processing. As I have argued, “procedures of abstraction make different knowledges comparable, calculable, and subject to re-engineering and reconstruction: they radically reshape the world in terms of the model that was originally abstracted” (Berry 2017, p. 104). Therefore, when researchers delegate analytical decisions to LLMs, they implicitly accept the ontological assumptions encoded in the model’s training and architecture and these may subtly undermine humanistic and social scientific research paradigms.[3]
This danger is an example of what Weizenbaum (1976) identified as the “ELIZA effect”, where humans readily attribute understanding to simple pattern-matching scripts. As my colleagues and I discovered through uncovering ELIZA’s original source code, Weizenbaum’s program operated through remarkably minimal linguistic transformations, yet users invested it with therapeutic insight and emotional intelligence (Berry and Marino 2024; see also Ciston et al 2026). The kind of "vibe coding" approach I adapt here for research purposes (“vibe research”) represents a particularly interesting example of this phenomenon, as the tight iterative loops that make this method productive also create conditions where a positive assumption about AI competence can most easily take hold. Where the ELIZA effect described users misrecognising computational processing as understanding, similarly with LLMs, a “competence effect” emerges when the positive over-confidence of an LLM chatbot obscures or hides the absence of ability (Berry 2025).
The competence effect is potentially more problematic than the ELIZA effect because the evidence appears to support the researcher’s anthropomorphism. The LLM produces analysis that reads plausibly, citations that look correct, and interpretations that seem insightful, creating a feedback loop that reinforces the users assumptions about the system’s capabilities. Unlike ELIZA users who had to ignore the system’s obvious repetitiveness and mistakes, researchers using AI sprints often receive constant positive reinforcement that the LLM “gets it”. The competence effect thus represents a qualitatively different epistemological trap, one where generative ability becomes confused with real comprehension. I argue, therefore, that the competence effect operates differently from traditional notions of anthropomorphism in human-computer interaction because it emerges not from designers deliberately creating human-like interfaces (as with ELIZA’s therapeutic approach) but from the LLM's training on vast human-generated text corpora. The system’s outputs bear surface similarity to human reasoning not through programmed imitation but through statistical learning from human language patterns. This might be said to create anthropomorphism as an emergent property rather than designed feature, albeit shepherded using guard-rails and “system prompts.” This might therefore make it more difficult for researchers and other users to maintain a critical distance from the software.
Productive Augmentation
It is important therefore to develop a working method with AI that enables what I call productive augmentation. This is where researchers maintain strategic control over research questions, analytical frameworks, and interpretive claims whilst using LLMs to rapidly process large datasets, generate intermediate objects such as visualisations, summary tables, and extracted themes, and implement analytical procedures that would otherwise require extensive technical expertise or research assistance.
In Phase Two of my coding experiment, this mode emerged clearly. My realisation that the LLMs text extraction code was faulty, enabled me to direct the LLM towards a better solution. The LLM then augmented my implementation suggestions, rapidly generating the interface code I proposed but would have taken considerable labour to write manually. This, I think, usefully demonstrates the optimal human-LLM division of labour, where a human's capacity for creative thinking and seeing the whole combines with AI generation of implementation options.
For AI sprints, productive augmentation could mean that the researcher controls the research design, identifies appropriate analytical frameworks, determines what intermediate objects will make findings legible, and maintains interpretive control over the results. The LLM then handles the implementation through creation of digital tools or direct processing of texts, generating provisional visualisations, creating computer code, and producing summary information. By these means the augmentation remains supplementary rather than substitutive, with each actor, human or AI, contributing capabilities the other lacks.
As Hayles (2012) observed, we think through, with, and alongside media. Productive augmentation brings this idea to the fore, allowing the LLM to become a medium for thought that extends human analytical capacity without displacing human judgement. However, maintaining this mode requires constant attention not to drift into cognitive delegation.
Cognitive Overhead
The third mode I want to discuss I call cognitive overhead and represents the scaling limits of AI augmentation. As I discovered in Phase Three of my software development project, beyond certain complexity levels the mental resources required to manage LLM context may exceed resources saved by not writing code directly. For example, the context window for my conversation with Gemini became increasingly unwieldy as the software grew more complex. The LLM began to get confused about the version we were working on, attempted to reimpose previously rejected features, or rolled back to earlier versions. Careful prompting was required to remind it which version was current, phrases like “fix only this, update the version number”, “add the 2.4 comment to this file”, etc. Indeed, I had to suggest to the AI the use of version numbering as it couldn't keep track of different iterations that it was generating. The cognitive resources freed by not writing implementation code were partially consumed by managing the LLM’s context and preventing reversions in the project.
This is distinct from both delegation and productive augmentation because rather than augmenting capabilities, the LLM now requires active human management and oversight. This included tracking version numbers, constraining scope with phrases like “do only these changes”, and preventing unsolicited modifications. The creative freedoms opened up by not writing implementation code are gradually consumed by project managing the LLM’s work. This shows that cognitive augmentation is not simply additive but involves real trade-offs where automating research and tools must be set against the cost of AI management.
For an AI sprint method, cognitive overhead suggests scaling limits beyond which the LLM will simply be unable to cope and require human-support to continue to work effectively. The approximately 1000-line program I developed reached these limits surprisingly quickly. For complex research projects involving multiple analytical threads, managing LLM context across extended timeframes may become more exhausting for the researcher than the gains from AI augmentation. This suggests AI sprints work best for bounded, intensive research tasks of perhaps 1-5 days duration, where the entire project remains manageable within an LLM context window. Extended scholarly projects spanning weeks or months would require either breaking the work into discrete sprint episodes, each producing intermediate objects that feed subsequent sprints, or adopting hybrid approaches that combine AI-augmented analysis with traditional inquiry. But we should also remember that the LLMs are developing quickly in their capacities, and this kind of meta-project knowledge will surely be a key objective for them to be able to manage in some way. At present we see attempts through Retrieval-Augmented Generation (RAG) and Model Context Protocol (MPC) technologies, together with interface tools like Cursor, to help manage these issues, but they remain a very real problem as the project grows beyond a certain threshold which is effectively an AI context collapse.
Key Principles
Before outlining the key principles, it is important to acknowledge the material conditions of possibility for an AI sprint practice. Whilst I have tried to show how AI augmentation can reduce dependence on technical expertise, this apparent democratisation may conceal new forms of infrastructural dependence. Access to frontier LLMs typically requires institutional subscriptions, API credits, or acceptance of platform terms of service that may restrict research freedom. The apparent reduction in human collaboration costs is thereby offset by computational costs that may be unevenly distributed across institutions and geographical regions. Moreover, the proprietary nature of most frontier LLMs means researchers depend on corporate platforms whose priorities often are very different from scholarly values. A critical AI sprint methodology must therefore be reflexive about these political economic issues, considering how platform dependencies shape what research becomes possible and for whom.
I now want to briefly outline what I think are the key principles that need to be kept in mind when developing an AI sprint. Firstly, it is crucial that some sense of architectural control is maintained by the researcher to ensure that strategic authority over research questions, analytical frameworks, and interpretive claims is ensured. The idea is that LLMs are used to create or generate research elements but do not determine the research direction. As I argue elsewhere, this requires treating the LLM as “a sophisticated pattern-completion engine rather than a genuine collaborator with understanding” (Berry 2025b). Secondly, it is important to foreground the value of producing intermediate objects, such as visualisations, summary tables, coded excerpts, thematic mappings, extracted JSON files, etc. that make LLM processing legible and contestable by the researcher. I think this might be one of the most important insights I have learnt from undertaking an AI Sprint. Researchers often assume LLMs can jump this stage and produce comprehensive research outputs – but in fact this is far from the truth due to limitations in the AIs we currently have. Creating these intermediate objects also prevent cognitive delegation by making algorithmic analysis visible and subject to critical analysis. As Latour (1999) demonstrates through his concept of “circulating references”, scientific facts emerge through transformations that progressively abstract from material phenomena to inscriptions. AI sprints can make this process very explicit through a process of raw data becoming intermediate objects (via the AI) then becoming analytical claims, with each transformation open to examination by the researcher.
We can understand intermediate objects in AI sprints not merely as transitional representations but as “materialised abstractions” that actively co-structure the research process. These objects emerge through what I have described as “a grammatisation process that discretises and re-orders” research materials (Berry 2016, p. 108), transforming complex qualitative materials into computationally usable forms and enabling new possibilities for analysis. The concern lies in what I call the technical “derangement of knowledge” where computational processing introduces distance from the source materials so that human intelligibility and interpretability is undermined (Berry 2016, p. 109). For example, when an LLM is used to process interview transcripts to generate a thematic summary table, the grammatisation operates through tokenisation, embedding, and pattern extraction that necessarily distances the intermediate object from the semantic thickness of the original conversation. The table generation makes certain patterns computationally legible whilst rendering other dimensions of meaning opaque or absent. The researcher has to remain critically aware not simply of what the table shows but what processes of abstraction produced it. This is to take an interpretative stance that I argue is key within critical digital humanities (Berry 2023).
Figure 3: hermeneutic-computational loop between human understanding and computational generationThirdly, it is important to develop a process of iterative revision where AI sprints use tight feedback loops between human interpretation and AI processing (see Figure 3). The researcher needs to pay careful attention to the intermediate objects generated, and identify patterns, anomalies, and questions requiring further investigation, then revise the prompts being used. This can involve introducing different analytical approaches, alternative visualisations, targeted probes into the data, or reorganising the data created (e.g. differently ordered JSON files). This enables what I term a hermeneutic-computational loop between human understanding and computational generation (Berry 2025b).
The next key principle is the importance of methodological transparency which is made possible by the textual outputs of LLMs which enables the documenting of prompts, LLM interactions, and key decision points or failures. This enables the possibilities of reproducibility, critique, and collective methodological learning by sharing and reflecting on these materials. As Marino (2020) has argued for in relation to the method of critical code studies, we must attend not just to what systems do but what they mean, examining how analytical choices shape possible interpretations. The last key principle is the need for continual reflexive critique throughout an AI sprint. This requires constant care to guard against this notion of the competence effect. It is important that researchers direct critical attention to the seductive effect of speed and scale, which may come at a cost of nuance, scholarly rigour and attention to context in the research. Indeed, an LLM’s capabilities are usually unevenly distributed across different analytical tasks, which makes it more suited to some than others. In addition an AI's training data may shape what becomes analytically visible or invisible in unexpected ways due to biases, absences and overfitting of the data.
Conclusions
AI sprints are a specific method for AI-augmented research, but I think their wider contribution lies in what I am calling “critical augmentation” as a research ethos.[4] I’ve tried to show that computational systems can extend human analytical capacities whilst insisting that such augmentation requires constant reflexive vigilance against instrumental rationality’s encroachment on the research process. As I discovered through developing an example of educational software via “vibe coding”, collaboration with LLMs produces useful insights whilst requiring constant care and project management combined with critical interpretability (Berry 2025b). LLMs can process texts at scales impossible for individual scholars, generate visualisations revealing unexpected patterns, create interfaces without requiring technical expertise, and so produce bespoke computational tools from natural language prompts.[5] Yet they cannot determine whether these processes are intellectually sound, contextually appropriate, or interpretively useful. That judgement remains with the human researcher.
** Headline image generated using Google Gemini Pro. November 2025. The prompt used was: "draw a high-resolution, photographic image capturing the concept of an 'AI Sprint.' The scene is set in a classic Oxford senior common room, characterised by dark wood paneling, floor-to-ceiling bookshelves with leather-bound books, and large, leaded glass windows. Around a large, polished wooden table, a diverse group of six students and researchers are deeply engaged in collaborative work. They are using a mix of modern and traditional research tools: laptops, smartphones, paper notebooks, and books are all visible and in use. The group is in discussion, pointing at screens and notes, embodying an atmosphere of intense, focused, and critical inquiry. There is no large central presentation screen or monitor; the focus is entirely on the group's interaction with each other and their personal research materials." Due to the probabilistic way in which these images are generated, future images generated using this prompt are unlikely to be the same as this version.
Notes
[1] The emergence of frontier LLMs poses what might become an existential challenge to the digital humanities. Much of the field has justified itself through appeals to a “methodological commons” of shared digital tools, techniques, and archives that require specialised technical literacy to access and deploy. If LLMs can generate bespoke analytical tools from natural language prompts and process archival materials without requiring such specialised knowledge, what becomes of these justifications for DH? The DH field seems largely unaware of this potential crisis, continuing to work on tool development and infrastructure building without adequately considering how generative AI might undermine the very rationale for such activity. This reinforces the importance of a critical digital humanities capable not merely of applying digital methods but of continually interrogating their epistemological foundations and their embeddedness within computational capitalism. A critical approach pays close attention to how these AI methods emerge from and reinforce particular formations of power and capital, rather than treating computational tools as neutral instruments for humanistic inquiry. The AI sprints methodology I propose here attempts to embed this into critical practice, developing reflexive critique in relation to the use of computational augmentation.
[2] The collaborative process I describe here is applied to text-based LLMs used for text and coding tasks. But mutatis mutandis they also apply to image generation systems like Stable Diffusion even though they use different technical architectures (e.g. iterative denoising processes rather than token prediction). The cognitive modes I identify, cognitive delegation, productive augmentation, and cognitive overhead operate in a slightly different register when the intermediate objects are images rather than code or structured data as visual outputs are more difficult to apply iterative revisions to and lack the version control that textual materials more readily accommodate.
[3] This is linked to concerns that connect to Frankfurt School critiques of instrumental rationality more broadly (Berry 2014). As Adorno and Horkheimer argued in the Dialectic of Enlightenment, the subsumption of qualitative difference under quantification represents a key moment in the dialectic of enlightenment itself. This is where reason’s emancipatory promises become means for new forms of domination. LLM-mediated research risks accelerating this tendency, as the ontological commitments required for computational processing favour those forms of knowledge amenable to quantification, formalisation, and algorithmic manipulation (see Berry 2014, chapter 2, Berry 2023).
[4] The AI sprint approach might also be applied reflexively to the study of artificial intelligence systems themselves, where AI augmentation becomes both method and object of inquiry. For instance, researchers might use LLMs to process large corpora of AI safety literature, technical documentation, or platform terms of service (ToS), generating intermediate objects that help to reveal patterns in how AI systems are governed, marketed, or constrained. Similarly, AI sprints could analyse datasets of LLM outputs to study the competence effect empirically, using computational methods to identify where generative output regresses away from accuracy or coherence. Researchers might also use AI sprints to examine training data biases, prompt engineering practices, or the circulation of AI-generated content across platforms, with the LLM serving as both analytical tool and research object.
[5] Particularly promising is the possibility to use AI sprints to map the probabilistic outputs of AI outputs through sophisticated version variation type approaches. This might operate in a number of ways. For example, researchers could use one LLM (the “research AI”) to systematically control and question another LLM (the “subject AI”), running identical prompts iteratively to capture the stochastic variation in outputs of synthetic media or AI slop (see Berry 2025). This would make visible the deterministically non-deterministic character of these systems, revealing how temperature settings, sampling methods, and hidden parameters shape the distribution of possible outputs. The research AI could also generate intermediate objects such as variation matrices, semantic clustering tables, or divergence/difference metrics that document the boundaries of output stability. Another approach might be to use an AI sprint to systematically compare different versions of particular frontier AIs (e.g. GPT-4 versus GPT-4.1, or Claude Opus versus Claude Sonnet), documenting how architectural changes, fine-tuning, or safety guardrails alter outputs. Such comparative analysis might reveal how commercial pressures, regulatory concerns, or user feedback reshape AI behaviour across version iterations, providing empirical grounding for claims about AI governance and development.
Comments
Post a Comment