Advait Sarkar

Andrew D. Gordon, Carina Negreanu, José Cambronero et al. · microsoft-research

Users are increasingly being warned to check AI-generated content for correctness. Still, as LLMs (and other generative models) generate more complex output, such as summaries, tables, or code, it becomes harder for the user to audit or evaluate the output for quality or correctness. Hence, we are seeing the emergence of tool-assisted experiences to help the user double-check a piece of AI-generated content. We refer to these as co-audit tools. Co-audit tools complement prompt engineering techniques: one helps the user construct the input prompt, while the other helps them check the output response. As a specific example, this paper describes recent research on co-audit tools for spreadsheet computations powered by generative models. We explain why co-audit experiences are essential for any application of generative AI where quality is important and errors are consequential (as is common in spreadsheet computations). We propose a preliminary list of principles for co-audit, and outline research challenges.

Majeed Kazemitabaar, Jack Williams, Ian Drosos et al. · microsoft-research, utoronto

LLM-powered tools like ChatGPT Data Analysis, have the potential to help users tackle the challenging task of data analysis programming, which requires expertise in data processing, programming, and statistics. However, our formative study (n=15) uncovered serious challenges in verifying AI-generated results and steering the AI (i.e., guiding the AI system to produce the desired output). We developed two contrasting approaches to address these challenges. The first (Stepwise) decomposes the problem into step-by-step subgoals with pairs of editable assumptions and code until task completion, while the second (Phasewise) decomposes the entire problem into three editable, logical phases: structured input/output assumptions, execution plan, and code. A controlled, within-subjects experiment (n=18) compared these systems against a conversational baseline. Users reported significantly greater control with the Stepwise and Phasewise systems, and found intervention, correction, and verification easier, compared to the baseline. The results suggest design guidelines and trade-offs for AI-assisted data analysis tools.

Bhuvanashree Murugadoss, Christian Poelitz, Ian Drosos et al. · microsoft-research

LLMs-as-a-judge is a recently popularized method which replaces human judgements in task evaluation (Zheng et al. 2024) with automatic evaluation using LLMs. Due to widespread use of RLHF (Reinforcement Learning from Human Feedback), state-of-the-art LLMs like GPT4 and Llama3 are expected to have strong alignment with human preferences when prompted for a quality judgement, such as the coherence of a text. While this seems beneficial, it is not clear whether the assessments by an LLM-as-a-judge constitute only an evaluation based on the instructions in the prompts, or reflect its preference for high-quality data similar to its fine-tune data. To investigate how much influence prompting the LLMs-as-a-judge has on the alignment of AI judgements to human judgements, we analyze prompts with increasing levels of instructions about the target quality of an evaluation, for several LLMs-as-a-judge. Further, we compare to a prompt-free method using model perplexity as a quality measure instead. We aggregate a taxonomy of quality criteria commonly used across state-of-the-art evaluations with LLMs and provide this as a rigorous benchmark of models as judges. Overall, we show that the LLMs-as-a-judge benefit only little from highly detailed instructions in prompts and that perplexity can sometimes align better with human judgements than prompting, especially on textual quality.

Advait Sarkar, Andrew D. Gordon, Carina Negreanu et al.

Large language models, such as OpenAI's codex and Deepmind's AlphaCode, can generate code to solve a variety of problems expressed in natural language. This technology has already been commercialised in at least one widely-used programming editor extension: GitHub Copilot. In this paper, we explore how programming with large language models (LLM-assisted programming) is similar to, and differs from, prior conceptualisations of programmer assistance. We draw upon publicly available experience reports of LLM-assisted programming, as well as prior usability and design studies. We find that while LLM-assisted programming shares some properties of compilation, pair programming, and programming via search and reuse, there are fundamental differences both in the technical possibilities as well as the practical experience. Thus, LLM-assisted programming ought to be viewed as a new way of programming with its own distinct properties and challenges. Finally, we draw upon observations from a user study in which non-expert end user programmers use LLM-assisted tools for solving data tasks in spreadsheets. We discuss the issues that might arise, and open research challenges, in applying large language models to end-user programming, particularly with users who have little or no programming expertise.

Advait Sarkar

Artificial Intelligence (AI), and in particular generative models, are transformative tools for knowledge work. They problematise notions of creativity, originality, plagiarism, the attribution of credit, and copyright ownership. Critics of generative models emphasise the reliance on large amounts of training data, and view the output of these models as no more than randomised plagiarism, remix, or collage of the source data. On these grounds, many have argued for stronger regulations on the deployment, use, and attribution of the output of these models. However, these issues are not new or unique to artificial intelligence. In this position paper, using examples from literary criticism, the history of art, and copyright law, I show how creativity and originality resist definition as a notatable or information-theoretic property of an object, and instead can be seen as the property of a process, an author, or a viewer. Further alternative views hold that all creative work is essentially reuse (mostly without attribution), or that randomness itself can be creative. I suggest that creativity is ultimately defined by communities of creators and receivers, and the deemed sources of creativity in a workflow often depend on which parts of the workflow can be automated. Using examples from recent studies of AI in creative knowledge work, I suggest that AI shifts knowledge work from material production to critical integration. This position paper aims to begin a conversation around a more nuanced approach to the problems of creativity and credit assignment for generative models, one which more fully recognises the importance of the creative and curatorial voice of the users of these models and moves away from simpler notational or information-theoretic views.

Advait Sarkar

Explaining the behaviour of intelligent systems will get increasingly and perhaps intractably challenging as models grow in size and complexity. We may not be able to expect an explanation for every prediction made by a brain-scale model, nor can we expect explanations to remain objective or apolitical. Our functionalist understanding of these models is of less advantage than we might assume. Models precede explanations, and can be useful even when both model and explanation are incorrect. Explainability may never win the race against complexity, but this is less problematic than it seems.

Advait Sarkar

The research field of end-user programming has largely been concerned with helping non-experts learn to code sufficiently well in order to achieve their tasks. Generative AI stands to obviate this entirely by allowing users to generate code from naturalistic language prompts. In this essay, we explore the extent to which "traditional" programming languages remain relevant for non-expert end-user programmers in a world with generative AI. We posit the "generative shift hypothesis": that generative AI will create qualitative and quantitative expansions in the traditional scope of end-user programming. We outline some reasons that traditional programming languages may still be relevant and useful for end-user programmers. We speculate whether each of these reasons might be fundamental and enduring, or whether they may disappear with further improvements and innovations in generative AI. Finally, we articulate a set of implications for end-user programming research, including the possibility of needing to revisit many well-established core concepts, such as Ko's learning barriers and Blackwell's attention investment model.

Qing Nancy Xia, Advait Sarkar, Duncan P. Brumby et al.

Informal Knowledge Sharing (KS) is vital for end-user programmers to gain expertise. To better understand how personal (self-efficacy), social (reputational gains, trust between colleagues), and software-related (codification effort) variables influence spreadsheet KS intention, we conducted a multiple regressions analysis based on survey data from spreadsheet users (n=100) in administrative and finance roles. We found that high levels of spreadsheet self-efficacy and a perception that sharing would result in reputational gains predicted higher KS intention, but individuals who found knowledge codification effortful showed lower KS intention. We also observed that regardless of occupation, users tended to report a lower sense of self-efficacy in their general spreadsheet proficiency, despite also reporting high self-efficacy in spreadsheet use for job-related contexts. Our findings suggest that acknowledging and designing for these social and personal variables can help avoid situations where experienced individuals refrain unnecessarily from sharing, with implications for spreadsheet design.

Advait Sarkar

Software adoption has traditionally been understood through instrumental lenses, such as usability, cost, security, and interoperability. We argue that a new, ideological dimension is reshaping adoption decisions: one we term digital patriotism, the individual counterpart to the state ideology of digital sovereignty. Through two studies, we trace this phenomenon. First, a directed content analysis of decisions made by European government agencies to switch away from de facto technology standards reveals a shift around 2020: early switches cited costs and vendor lock-in, while later switches invoke sovereignty, geopolitical risk, and investment in local industry. Second, a qualitative analysis of over 700 online comments (over 51,000 words) surfaces how consumers and businesses articulate motivations for seeking European software alternatives. We find that digital patriotism entails a willingness to accept functional compromise in service of ideological goals. Our work extends software adoption theory by drawing attention to value rationality alongside instrumental rationality, and contributes an empirical account of how geopolitics is reshaping technology choice in the workplace.

Advait Sarkar

Filter Babel is a thought experiment about a near future in which everything we read, watch, and even whom we "meet" is privately generated for each of us. If we each recede into a world of purely private experience, we may each develop a Wittgensteinian private language that remains intelligible to others only because an AI translator sits in the middle. This intermediation challenges the integrity of common ground and therefore of communication. On the other hand, private experience is an essential engine of identity and selfhood: as Lanier warns, one must be somebody before one can share oneself. This paper opens a discussion of the challenges and opportunities that Filter Babel might present to human communication and identity, and what constructive directions for research in AI-mediated communication might ensue.

Qing Nancy Xia, Advait Sarkar, Duncan Brumby et al.

Spreadsheet collaboration provides valuable opportunities for learning and expertise sharing between colleagues. Sharing expertise is essential for the retention of important technical skillsets within organisations, but previous studies suggest that spreadsheet experts often fail to disseminate their knowledge to others. We suggest that social norms and beliefs surrounding the value of spreadsheet use significantly influence user engagement in sharing behaviours. To explore this, we conducted 31 semi-structured interviews with professional spreadsheet users from two separate samples. We found that spreadsheet providers face challenges in adapting highly personalised strategies to often subjective standards and evaluating the appropriate social timing of sharing. In addition, conflicted self-evaluations of one's spreadsheet expertise, dismissive normative beliefs about the value of this knowledge, and concerns about the potential disruptions associated with collaboration can further deter sharing. We suggest these observations reflect the challenges of long-term learning in feature-rich software designed primarily with initial learnability in mind. We therefore provide implications for design to navigate this tension. Overall, our findings demonstrate how the complex interaction between technology design and social dynamics can shape collaborative learning behaviours in the context of feature-rich software.

Shraddha Barke, Christian Poelitz, Carina Suzana Negreanu et al.

Large language models (LLMs) are rapidly replacing help forums like StackOverflow, and are especially helpful for non-professional programmers and end users. These users are often interested in data-centric tasks, such as spreadsheet manipulation and data wrangling, which are hard to solve if the intent is only communicated using a natural-language description, without including the data. But how do we decide how much data and which data to include in the prompt? This paper makes two contributions towards answering this question. First, we create a dataset of real-world NL-to-code tasks manipulating tabular data, mined from StackOverflow posts. Second, we introduce a cluster-then-select prompting technique, which adds the most representative rows from the input data to the LLM prompt. Our experiments show that LLM performance is indeed sensitive to the amount of data passed in the prompt, and that for tasks with a lot of syntactic variation in the input table, our cluster-then-select technique outperforms a random selection baseline.

Ian Drosos, Jack Williams, Advait Sarkar et al. · microsoft-research

Effective prompting of generative AI is challenging for many users, particularly in expressing context for comprehension tasks such as explaining spreadsheet formulas, Python code, and text passages. Prompt middleware aims to address this barrier by assisting in prompt construction, but barriers remain for users in expressing adequate control so that they can receive AI-responses that match their preferences. We conduct a formative survey (n=38) investigating user needs for control over AI-generated explanations in comprehension tasks, which uncovers a trade-off between standardized but predictable support for prompting, and adaptive but unpredictable support tailored to the user and task. To explore this trade-off, we implement two prompt middleware approaches: Dynamic Prompt Refinement Control (Dynamic PRC) and Static Prompt Refinement Control (Static PRC). The Dynamic PRC approach generates context-specific UI elements that provide prompt refinements based on the user's prompt and user needs from the AI, while the Static PRC approach offers a preset list of generally applicable refinements. We evaluate these two approaches with a controlled user study (n=16) to assess the impact of these approaches on user control of AI responses for crafting better explanations. Results show a preference for the Dynamic PRC approach as it afforded more control, lowered barriers to providing context, and encouraged exploration and reflection of the tasks, but that reasoning about the effects of different generated controls on the final output remains challenging. Drawing on participant feedback, we discuss design implications for future Dynamic PRC systems that enhance user control of AI responses. Our findings suggest that dynamic prompt middleware can improve the user experience of generative AI workflows by affording greater control and guide users to a better AI response.

Advait Sarkar, Christian Poelitz, Viktor Kewenig

Generative AI tools often answer questions using source documents, e.g., through retrieval augmented generation. Current groundedness and hallucination evaluations largely frame the relationship between an answer and its sources as binary (the answer is either supported or unsupported). However, this obscures both the syntactic moves (e.g., direct quotation vs. paraphrase) and the interpretive moves (e.g., induction vs. deduction) performed when models reformulate evidence into an answer. This limits both benchmarking and user-facing provenance interfaces. We propose the development of a reader-centred taxonomy of grounding as a set of support relations between generated statements and source documents. We explain how this might be synthesised from prior research in linguistics and philosophy of language, and evaluated through a benchmark and human annotation protocol. Such a framework would enable interfaces that communicate not just whether a claim is grounded, but how.

Ahana Ghosh, Advait Sarkar, Siân Lindley et al.

Generative AI (GenAI) tools improve productivity in knowledge workflows such as writing, but also risk overreliance and reduced critical thinking. Cognitive forcing functions (CFFs) mitigate these risks by requiring active engagement with AI output. As GenAI workflows grow more complex, systems increasingly present execution plans for user review. However, these plans are themselves AI-generated and prone to overreliance, and the effectiveness of applying CFFs to AI plans remains underexplored. We conduct a controlled experiment in which participants completed AI-assisted writing tasks while reviewing AI-generated plans under four CFF conditions: Assumption (argument analysis), WhatIf (hypothesis testing), Both, and a no-CFF control. A follow-up think-aloud and interview study qualitatively compared these conditions. Results show that the Assumption CFF most effectively reduced overreliance without increasing cognitive load, while participants perceived the WhatIf CFF as most helpful. These findings highlight the value of plan-focused CFFs for supporting critical reflection in GenAI-assisted knowledge work.

Qing, Xia, Marios Constantinides et al.

Generative AI (GenAI) tools are rapidly transforming knowledge work, making AI literacy a critical priority for organizations. However, research on AI literacy lacks empirical insight into how knowledge workers' beliefs around GenAI literacy are shaped by the social dynamics of the workplace, and how workers learn to apply GenAI tools in these environments. To address this gap, we conducted in-depth interviews with 19 knowledge workers across multiple sectors to examine how they develop GenAI competencies in real-world professional contexts. We found that, while knowledge sharing from colleagues supported learning, the ability to remove cues indicating GenAI use was perceived as validation of domain expertise. These behaviours ultimately reduced opportunities for learning via knowledge sharing and undermined transparency. To advance workplace AI literacy, we argue for fostering open dialogue, increasing visibility of user-generated knowledge, and greater emphasis on the benefits of collaborative learning for navigating rapid technological developments.

Lev Tankelevitch, Elena L. Glassman, Jessica He et al. · microsoft-research

Generative AI (GenAI) radically expands the scope and capability of automation for work, education, and everyday tasks, a transformation posing both risks and opportunities for human cognition. How will human cognition change, and what opportunities are there for GenAI to augment it? Which theories, metrics, and other tools are needed to address these questions? The CHI 2025 workshop on Tools for Thought aimed to bridge an emerging science of how the use of GenAI affects human thought, from metacognition to critical thinking, memory, and creativity, with an emerging design practice for building GenAI tools that both protect and augment human thought. Fifty-six researchers, designers, and thinkers from across disciplines as well as industry and academia, along with 34 papers and portfolios, seeded a day of discussion, ideation, and community-building. We synthesize this material here to begin mapping the space of research and design opportunities and to catalyze a multidisciplinary community around this pressing area of research.

Judith Borghouts, Andrew D. Gordon, Advait Sarkar et al.

Spreadsheet users regularly deal with uncertainty in their data, for example due to errors and estimates. While an insight into data uncertainty can help in making better informed decisions, prior research suggests that people often use informal heuristics to reason with probabilities, which leads to incorrect conclusions. Moreover, people often ignore or simplify uncertainty. To understand how people currently encounter and deal with uncertainty in spreadsheets, we conducted an interview study with 11 spreadsheet users from a range of domains. We found that how people deal with uncertainty is influenced by the role the spreadsheet plays in people's work and the user's aims. Spreadsheets are used as a database, template, calculation tool, notepad and exploration tool. In doing so, participants' aims were to compute and compare different scenarios, understand something about the nature of the uncertainty in their situation, and translate the complexity of data uncertainty into simplified presentations to other people, usually decision-makers. Spreadsheets currently provide limited tools to support these aims, and participants had various workarounds.