--- source_url: https://acceleratelearning.stanford.edu/conference/responsible-assessment-in-the-ai-era/ ingested_date: 2026-08-03 sha256: adeeb44444037b653e409c26cafecfcf371f31b5e5f1a3e635c16e3e036aefbf --- # Responsible Assessment in the AI Era: Key Insights from a Future-Focused Conference RESPONSIBLE ASSESSMENT IN THE AI ERA Key Insights from a Future-Focused Conference Nneka J. McGee Candace Thille Ikkyu Choi Kadriye Ercikan Isabelle C. Hau Matthew S. Johnson (Reviewer) Table of Contents 1 .Introduction: Assessment at a Moment of Change 5 1.1 Opening the Lens on Responsible Assessment 5 2. Why Traditional Assessment is Evolving 7 2.1 Misalignment with How Learning Actually Occurs 7 2.2 The Increasing Influence of AI 8 2.3 A Growing Gap Between What We Measure and What Matters 10 3. What System Change Requires: Rethinking Assessment 12 3.1 Testing Events to Systems of Inference 12 3.2 Building Infrastructure and Validity for AI-Powered Assessment 13 3.3 Expanding What Counts as Evidence 13 3.4 Extending the Role of Formative Assessment 14 4. Advancing a Focus on Responsible Assessment 15 4.1 Reflecting on Responsible Assessment 15 4.2 Establishing Trust through Transparency 15 4.3 Translating Insight into Action 16 Conclusion 19 Glossary 20 Endnotes 21 Acknowledgments We would like to express our appreciation to the following individuals for their support and contributions to the Responsible Assessment in the AI Era convening: Daniel L. Schwartz the I. James Quillen Dean, and Nomellini & Olivier Professor of Educational Technology at Stanford Graduate School of Education, and The Halper Family Faculty Director of the Stanford Accelerator for Learning. Amit Sevak Chief Executive Officer, ETS. Candace Thille Associate Professor, Stanford Graduate School of Education, Faculty Director, Adult and Workforce Learning Initiative, Stanford Accelerator for Learning. Collaborating Organization The convening was supported by Educational Testing Service (ETS), an organization whose mission focuses on “advancing the science of measurement to power human progress.” Recognition of Contributions We would also like to extend our gratitude to panelists, moderators, poster session presenters, and breakout session contributors. Their discussions informed the themes, tensions, and future-forward moves reflected throughout this paper. Panelists Bryan Brown, Stanford University Matthew S. Johnson, ETS* Emma Brunskill, Stanford University Victor Lee, Stanford Graduate School of Education Sara Caldwell, Open AI and Stanford Accelerator for Learning Tony Chan, Former President, KAUST* Lydia Liu, ETS Kristen DiCerbo, Khan Academy Temple Lovelace, Assessment for Good,AERDF Ben Domingue, Stanford Graduate School of Nancy Otero, Gates Foundation Education* Kadriye Ercikan, ETS Maureen Heymans, Google* Jesse R. Sparks, ETS* Moderators Mille Garcia, California State University Darienne Hudson, United Way, Southeastern MI Kevin Gutherie, ITHAKA Beverly Daniel Tatum, Spelman College *panelist or presenter and breakout session contributor **graduate student at Stanford GSE and breakout session contributor Poster Session Presenters ETS Stanford Graduate School of Education Ikkyu Choi* Yunsung Kim** Patrick Kyllonen* Hansol Lee** Matthew S. Johnson* Eunjung Myoung Sandip Sinharay* Rojas Pino Diego Zapata-Rivera* Marcos Santiago *panelist or presenter and breakout session contributor **graduate student at Stanford GSE and breakout session contributor We thank all attendees who joined and contributed to the spirit of inquiry. Their engagement reminded us that responsible assessment in the AI era is not the work of one organization or discipline, but a shared effort across communities committed to learning. Authors Nneka J. McGee, J.D., Ed.D., Project lead and principal co-author for the “Responsible Assessment in the AI Era: Key Insights from a Future-Focused Conference” white paper. Candace Thille, Ed.D., Associate Professor, Stanford Graduate School of Education, Faculty Director, Adult and Workforce Learning Initiative, Stanford Accelerator for Learning. Ikkyu Choi, Ph.D., Senior Research Scientist, ETS Kadriye Erkican, Ph.D., Senior Vice President of Global Research, ETS; President, ETS Canada Inc. Isabelle C. Hau, Executive Director, Stanford Accelerator for Learning Recommended Citation: Nneka J. McGee, Candace Thille, Ikkyu Choi, Kadriye Ercikan, and Isabelle C. Hau. “Responsible Assessment in the AI Era: Key Insights from a Future-Focused Convening,” Stanford Accelerator for Learning, Stanford University, 2026. Reviewer: Matthew S. Johnson, Ph.D., Principal Research Scientist, ETS Research Institute Disclaimer: This paper reflects the perspectives of the authors and includes quotes from the “Responsible Assessment in the AI Era” convening. It does not necessarily represent the views of Stanford University, the Stanford Accelerator for Learning, ETS, participants or collaborating organizations. A note on AI: This paper was developed and written by the principal authors, who reviewed and fact- checked all content. ChatGPT (GPT 5.5) was used sparingly for minor editorial refinement. Sembly AI was used to transcribe and summarize breakout session discussions. 4 Responsible Assessment in the AI Era: Key Insights from a Future-Focused Conference 1. Introduction: Assessment at a Moment of Change “H ow do we shape the current moment?” The question, from Stanford Professor and Stanford Accelerator for Learning Faculty Co-Director Candace Thille, framed a broader discussion about “We want to have conversations how to approach assessment in the era of artificial that are rich and deep about how intelligence (AI). In search of answers, the Stanford Accelerator for Learning, in collaboration and with research, policy, technology and support from ETS, hosted the “Responsible Assessment education come together.” in the AI Era” convening on January 29, 2026. — Amit Sevak, CEO of ETS “We want to have conversations that are rich and deep about how research, policy, technology and education come together,” Amit Sevak, CEO of ETS, said in his opening remarks. The convening drew approximately 100 leaders from research, technology, K-12 schools, 1.1 Opening the Lens on Responsible and higher education institutions seeking to advance Assessment the science and practice of responsible assessment in the age of AI. There was a consensus among them: The conference took place at a critical juncture for Assessment is at a moment of change, and AI is a major assessment. There are growing calls for assessments driver of that transformation. to become more socioculturally responsive to learners’ cultural, educational, and societal contexts in order to Panelists shared wide-ranging perspectives on optimize engagement, motivation and performance. evidence, rethinking readiness, socioculturally Through personalization, AI may help address some of responsible assessment, and building trustworthy AI. these goals. Through poster sessions, scholars presented findings from their latest studies. Contributors expressed their At the same time, the integration of AI into education viewpoints at breakout sessions on topics ranging from and assessment raises questions about which AI policy to extracting construct evidence. While this competencies learning and assessment should paper primarily reflects the insights and perspectives prioritize, and how to leverage the potential of AI while shared by speakers at the event, it also draws on minimizing the risks of exacerbating opportunity relevant research and related work to contextualize the gaps for individuals from different societal and contributors’ dialogue. It summarizes key takeaways cultural backgrounds. The conference brought these from the convening and includes a glossary of terms concepts together by focusing on how assessment related to assessment. can become both responsive and responsible for Responsible Assessment in the AI Era: Key Insights from a Future-Focused Conference 5 Socioculturally Responsive Supports Focuses on Responsible Most Critical Learning and Assessment Competencies Opportunity for Education Leverages Responsible AI for Better Measurement Figure 1: Responsible Assessment meeting evolving educational goals. This approach Videos of key sessions at the event can be watched promotes responsible AI by supporting valid and fair here: https://acceleratelearning.stanford.edu/ interpretations and uses of assessment (See Fig. 1). conference/responsible-assessment-in-the-ai-era/ In this paper, we use the term “responsible assessment in the AI era” to refer to assessment that is grounded in individuals’ sociocultural contexts and designed to generate valid, trustworthy, and context- specific inferences from accumulated evidence to support learning and development. Responsible assessment is a call that represents a shift toward continuous, context-rich, and developmentally oriented assessment practices that leverage AI in responsible ways, whose principles and practices are detailed in Johnson (2025).1 It emphasizes a focus on complex critical competencies, accumulation of multiple sources of evidence through socially situated performance over time, and the consideration of sociocultural context. It also seeks to support learning and growth while prioritizing transparency and interpretability in how evidence is used, elicited, and accumulated. 6 Responsible Assessment in the AI Era: Key Insights from a Future-Focused Conference 2. Why Traditional Assessment is Evolving A ssessment is used to measure knowledge and skills, monitor growth over time, support learning, inform accountability, and guide decisions about admissions and program access.2 In The promise of AI depends on how practice, traditional assessments frequently operate at the intersection of improvement and accountability, assessment helps educators see and where the same measures are expected to support respond to learners’ next steps. learning while also serving evaluative and comparative functions.3 This section discusses three interconnected factors shaping the evolution of assessment amid multiple, sometimes competing demands. Temple Lovelace, executive director of assessment for 2.1 Misalignment with How Learning good at Advanced Education Research & Development Actually Occurs Fund (AERDF) described the value of providing space for learners to show up as their full selves in a learning “Traditional assessments may tell us if a student environment. What students bring to learning follows arrived at a correct answer,” stated Millie García, them into assessment. “Full self includes family Chancellor of the California State University. “But history, home context, and societal context,” explained they often tell us little about how [students] got Kadriye Ercikan, senior vice president of global there. In today’s collaborative digital and increasingly research at ETS. “This affects how they relate to the AI-supported communities, that gap matters.” While assessment…[and] how they engage and make sense learning unfolds as a process, assessment often of the assessment.” Although standardization is often remains largely event-based. This dynamic creates a intended to even the playing field, Ercikan suggested mismatch between the realities of learning and the that equivalent assessment conditions are lacking. way assessment is commonly designed. Many current systems still emphasize end products This mismatch is most evident when learning and over the pathways that produce them. assessment create varied conditions for students. AI-supported learning environments may create new Cindy Mazow, director of learning technology and opportunities to address this misalignment, but only design at the Stanford Graduate School of Business, if they are designed around sound learning principles. emphasized that effective learning requires room Maureen Heymans, vice president of learning at for students to make mistakes or take risks because Google, suggested that “learning is most effective those actions can expose misconceptions or gaps when it is active, engaging, and personalized.” When in understanding which can then be corrected. In connecting learning to Vygotsky’s Zone of Proximal assessment mode, however, Mazow indicated that Development, Josh Arnold, co-founder of Impacter students “don’t want to make a mistake because they Pathway, suggested that learners can do “hard and know that there’s a consequence to that assessment gritty things” when those things are relevant and right- whether it’s high-stakes or low-stakes.” The conditions sized to learners’ current abilities.4 that support learning, then, are not always present when students are being assessed. Responsible Assessment in the AI Era: Key Insights from a Future-Focused Conference 7 In practice, the promise of AI depends on how understanding”6 As Thille of the Stanford Accelerator assessment helps educators see and respond to for Learning said, when assessment no longer learners’ next steps. “What do we change on the measures what it was meant to measure, the validity ground when it comes to assessment and learning?” of inferences based on assessment outcomes may Lovelace asked, connecting the broader discussion to be compromised. her experiences as a special education teacher. Her question points toward a future where AI-supported AI is also shaping how assessment evidence is assessment is more integrated with learning. The scored and interpreted. Ercikan of ETS specified challenge is to strengthen learner support without that scoring is the most widely used application turning every moment of learning into another layer of AI in assessment. This creates validity concerns of assessment. when AI-based scoring, such as automated scoring systems, captures features of performance unrelated to the intended construct. Ercikan offered 2.2 The Increasing Influence of AI an example of a science assessment in which AI affects multiple points of the assessment process, AI scoring rewards language competency rather including how student work is produced and how that than scientific reasoning to describe construct- work is evaluated. As the use of AI increases among irrelevant variance. learners, there are potential disconnects between what assessments are designed to measure and what When such irrelevant variance systematically they actually capture. The technology also impacts the advantages or disadvantages particular learners, it skills and competencies deemed important at schools becomes a fairness concern and, as Ercikan noted, “is and in the workforce. at the core of bias in assessment.”7 Stanford Graduate School of Education Professor Guillermo Salano- 2.2.a Validity Flores has identified three areas where bias can be introduced: language elements (e.g. words), visual Generative AI (GenAI) has exposed the limitations components (e.g. charts), and contextual scenarios of traditional assessments based primarily on final (e.g. word problems).8 Bias threatens validity when outputs.5 Learners can now use GenAI to generate parts of the assessment are not central to what is high quality products without fully engaging in the being measured.9 learning processes those products are meant to represent. As a result, outputs alone can no longer Contributors shared additional AI-influenced be assumed to serve as reliable indicators of human assessment practices that impact validity, including: capability. In these cases, assessment “risks measuring technological proficiency rather than human skill or • Construct underrepresentation: Ercikan explained As AI systems take on more routine this can occur when an AI scoring system only partially captures what the assessment is intended cognitive tasks, assessment will to measure.10 need to focus more deliberately on • Generalization: Andrew McEachin, senior research capabilities that are uniquely human. director at ETS, described how AI systems can fail to generalize across contexts and how a human’s ability to recognize when AI systems generalize is not well-developed.11 Together, these examples show that assessment priorities are changing. AI literacy is not sufficient by • Training data: McEachin also expressed concern itself. As AI systems take on more routine cognitive about AI models that are not trained on data tasks, assessment will need to focus more deliberately representing authentic child learning, particularly in on capabilities that are uniquely human. They are K–12 education.12 often described as “durable skills” and include critical thinking, creativity, curiosity, collaboration, agency, • Calibration: Matthew S. Johnson, managing and adaptability.15 Emotional intelligence, relational director of innovative research at ETS, used the intelligence, ethical judgment, and challenging example of rubrics to highlight how what humans assumptions have also been identified as uniquely and AI systems take into account can differ on the human skills.16 The Stanford Accelerator for Learning same assessment.13 examined these shifting priorities in Future-Ready This list of validity threats is not exhaustive, but it Voices, a November 2025 interview series featuring shows how AI can influence assessment at multiple researchers, educators, and thought leaders.17 points, from the production of student work and the scoring and interpretation of that work. 2.2.c AI and the Workplace Job availability and workforce displacement remain 2.2.b AI literacy and essential competencies pressing challenges in the AI era. Citing a McKinsey During a panel discussion on measuring essential report, Darienne Hudson, president and CEO of United competencies, Jesse R. Sparks, research director Way for Southeastern Michigan, predicted that 30% at ETS Research Institute, pointed to AI literacy of jobs will become automated within four years.18 as an emerging construct. Its rise reflects AI’s According to the World Economic Forum, employers increasing influence across education and work. expect 39% of workers’ core skills to change by 2030.19 Sparks emphasized that redesigned assessments Regarding specific professions, Caldwell said, “It’s are needed to capture key abilities associated with much less about that an engineer won’t exist, but AI literacy, including collaborative interaction and what aspects of that role are going to fundamentally critical evaluation.14 change.” In other words, assessment should focus on which skills workers will need to succeed in changing The conversation also moved beyond AI literacy work environments. to consider additional competencies learners will need in the age of AI. Sparks, Sara Caldwell, Sparks envisioned a workplace where “we interact head of GTM Readiness at OpenAI, and Victor Lee, with AI tools in ways that really enhance the strengths associate professor at the Stanford Graduate School that we’re bringing to the situation.” Caldwell used of Education and faculty lead for AI+Education at as an example the ability to orchestrate a set of AI the Stanford Accelerator for Learning, identified agents. Lydia Liu, associate vice president of ETS, competencies such as systems thinking and deep highlighted a collaboration between ETS and OECD domain expertise as especially important. Caldwell named PISA VET (Vocational Education and Training).20 suggested that people who combine their own The international initiative aims to benchmark competencies with AI will be better positioned in the performance across countries to identify factors that workforce than those who rely on AI alone. are associated with positive outcomes in vocational education. Some of the AI-related innovations Liu described include: Responsible Assessment in the AI Era: Key Insights from a Future-Focused Conference 9 • VR glasses to observe if automotive technician assessment must broaden to show how capabilities students can perform critical tasks related to develop and appear in practice. the job such as inspections and diagnoses, and maintenance and repair, and 2.3.a Advances in Data Availability AI is expanding what can be observed and measured • Pressure sensing gloves to determine if trainees about learning and performance. Throughout the can apply the right amount of pressure to complete panels and breakout sessions, stealth assessment a task. was discussed as one example of how AI is advancing existing approaches to gathering evidence. Liu’s descriptions of immersive technologies reveal Sparks described stealth assessment as a process a gap between what many traditional assessments methodologically grounded in evidence-centered measure and the types of skills increasingly valued design (ECD) where educators and researchers at work. Applications of new technologies and tools “continuously extract signals from a learner’s everyday opened the door for more comprehensive and learning environment.” Contributors also discussed authentic assessment. As expectations change, so do how systems can draw on process data, ambient data, signals of readiness, as evidenced by the proliferation longitudinal data, and linked systems that connect of alternative certifications, skill verifications, and assessment results with broader patterns of learning micro-credentials.21 These signals reflect rising and performance. demand for evidence that is more closely tied to demonstrable capability.22 Patrick Kyllonen, distinguished presidential appointee at ETS, highlighted continuous assessment, 2.3 A Growing Gap Between What We which collects data to measure student progress Measure and What Matters continuously over time. He suggested that this approach may reduce assessment fatigue and test “Traditional assessment is misleading,” said Daniel L. anxiety. Advanced analytic methods are making it Schwartz, dean of the Stanford Graduate School of possible to interpret patterns from those data in Education and the Halper Family Faculty Director of new ways. the Stanford Accelerator for Learning. He described the implications of a study on how students adapt However, simply increasing the amount of data is not knowledge to a novel situation.23 “We need instruction sufficient. To paraphrase Julia Wilkowski, who directs that produces adaptive learners and assessment that learning sciences at Google, more evidence does not can tell.” automatically produce better insight. Beneath the immediate concerns about AI and Assessments such as stealth assessment and assessment is a more enduring issue expressed by continuous assessment raise concerns about how OpenAI’s Caldwell: Measurement signals what systems learner data are collected, interpreted, and protected. value. That relationship is not neutral. Assessment Vikas Wadhwani, a leader of the learning and shapes not only how learning is measured, but also certifications team at Meta, identified privacy as one what educational systems prioritize.24 A large-scale challenge. Another is learner awareness, since students assessment used for comparability and accountability may change their behavior when they know they are can influence curriculum, instruction, and public being assessed. Synthetic data and simulated learner judgments of quality.25 models, sometimes described as “simulated students,” could offer a way to test assessment designs, study Over time, systems establish priorities, assessments possible response patterns, and examine item operationalize them, and repeated use reinforces what performance while reducing reliance on identifiable counts as evidence. While these approaches enable student data.26 Simulated students may support common metrics across large populations, they can faster experimentation, but validation would still be also narrow how and what learning is represented. If needed to ensure simulated patterns do not distort educational systems place more value on skills and assumptions about real learners. learning processes, then the data gathered through 10 Responsible Assessment in the AI Era: Key Insights from a Future-Focused Conference 2.3.b Measuring Human-Centered Skills Human-centered skills, such as durable skills, were described briefly earlier in this paper. There is some consensus that human-centered skills can be assessed. However, methods for measuring them remain less established, particularly because they are “interpersonal, intrapersonal, or behavioral in nature.”27 2.3.b.i. Underdefined meanings for AI literacy and human- centered skills Stanford Accelerator for Learning’s Faculty Lead for AI+Education Victor Lee recognized that AI literacy is one of the core competencies for the AI era, but Wide applications of live demonstrations will require stresses that “we don’t really have a clear articulation operational efficiency for feasibility. for what that is.” The AILit Framework, a joint initiative from OECD and the European Commission, includes • Role-playing: ETS’s Patrick Kyllonen outlined an a comprehensive definition of AI literacy and will initiative where students interact with AI as a role- contribute to the 2029 PISA Media and Artificial playing partner to measure social skills.33 Intelligence Literacy assessment.28 Even so, definitions • Scenario-based assessments: Kyllonen also of AI literacy continue to vary across recognized explained how AI can be used in assessment institutions and organizations.29 environments to generate scenarios that reflect A similar issue appears with adaptability. Although non-traditional, real-world situations.34 widely named as essential in future-focused, AI-rich • Portfolio-based assessments: Contributors environments, adaptability lacks a standard definition discussed evaluations that allow learners to with operational indicators across AI-use scenarios.30 demonstrate growth and real-world application of These constructs must be carefully defined in knowledge through a curated collection of work operational terms if assessments are to collect over time, rather than through a single test or exam. meaningful data to measure.31 • Conversation-based assessments: Diego Zapata- 2.3.b.ii. Existing and emerging tools to measure human- Rivera, distinguished president appointee at the ETS centered skills Research Institute, described an evidence-centered Contributors characterized the measurement of design (ECD) based approach that uses AI agents human-centered skills as part of a larger shift from to help learners demonstrate knowledge and skills controlled assessment environments toward more through dialogue.35 authentic, open-ended interactions. Some of the tools discussed are already established in practice, • Virtual robotics programming tasks: Sparks, of while others are newer or remain in varying stages of ETS, detailed emerging work that examines the research and development. Here is sampling of the potential for identifying signals of persistence and tools referenced during the convening: resilience from learners’ interactions with virtual robots in block-based coding environments.36 • Cognitive interviews: Lydia Liu of ETS suggested this format to gain a better understanding of • Game-based assessment in naturalistic settings: students’ cognitive process when interacting with Guess What?, a smartphone app created by Stanford assessment or learning tasks, particularly when Professor and Stanford Accelerator for Learning considering cultural or behavioral cues.32 faculty affiliate Dennis Wall, uses game-based interactive computer vision to identify early signs of • Live demonstrations: In addition, Liu described autism in naturalistic settings.37 this format as the most authentic way to collect evidence about what learners know and can do. Responsible Assessment in the AI Era: Key Insights from a Future-Focused Conference 11 3. What System Change Requires: Rethinking Assessment S ystem change presents a challenge for how Johnson, also of ETS, said motivation is another assessment can support the complexities limiting factor because learners question the reasons of modern learning. Learning is continuous, for summative assessments for which they may never adaptive, contextual, and increasingly shaped by see the results. human and technological interaction. That complexity makes it difficult to rely on assessment models A different orientation places learner development designed primarily to capture performance at a at the center of assessment. This orientation focuses single point in time. There are multiple pathways to on formative assessment that provides ongoing rethinking assessment, and AI will play an important feedback rather than a snapshot of performance and role in expanding the evidence available to guide draws on evidence that can illuminate “the course interpretations, noted Karin Forssell, senior lecturer of development and the status of skills.”38 Once the with the Stanford Graduate School of Education and attention on assessment shifts from summative director of the AI Tinkery at the Stanford Accelerator assessment to learner development, the process of for Learning. gathering and utilizing evidence must be reconsidered. AI may increase the feasibility of formative assessment 3.1 Testing Events to Systems of by synthesizing evidence and identifying patterns Inference within large and complex data streams, which in Summative assessment is designed to evaluate turn can contribute to more refined inferences about learning at the culmination of an instruction period. learning. Zapata-Rivera used the principles of ECD and Despite its importance, ETS’ Sparks highlighted a the Autotutor dialogue framework to describe how limitation: the results of some summative assessments AI agents could support this type of assessment. “In arrive too late to be useful to learners and educators. evidence centered design, each piece of each task is 12 Responsible Assessment in the AI Era: Key Insights from a Future-Focused Conference designed for a particular purpose,” explained Zapata- in AI-mediated environments. For example, large Rivera. “With AI, we could create expert agents to language model (LLM) scoring of constructed generate questions or create opportunities even if we responses may require more extensive validity are not there.” evidence than traditional approaches due to concerns about consistency and interpretability.40 In addition, using LLMs as opposed to traditional scoring methods 3.2 Building Infrastructure and Validity could be less reliable and more costly due, in part, to for AI-Powered Assessment the expense involved in curating validity evidence.41 Christian Pinedo, vice president of external affairs and advocacy at aiEDU, emphasized that building 3.3 Expanding What Counts as Evidence comprehensive infrastructure “requires a lot of human investment.” Systems that incorporate AI, The transformation of assessment requires a including data architectures and interoperable broader evidentiary base. For example, relational platforms, must have the infrastructure to support the learning builds on Vygotsky’s social constructivism collection, integration, interpretation, and storage of by emphasizing strong teacher-student relationships large amounts of data. In response, LeeAnn Lindsey, in classroom settings.42 In a panel on rethinking director of edtech and innovation at Northern measurement, Liu of ETS indicated that traditional Arizona University’s Arizona Institute for Education assessment systems were not designed to capture and the Economy, added that many state and these dimensions of learning well.43 Social and district budgets are stretched while acknowledging relational learning often involve complex skills the “lack of basic infrastructure for implementing that have cognitive, affective, and behavioral good decision-making through assessment.” Their components, but also depend on capacities for exchange points to the need for sustained investment dialogue, interactions, building trust, and collective in assessment infrastructure, even in constrained problem-solving. A future-oriented taxonomy fiscal environments. Without such infrastructure, even from Liu and her colleagues illustrate this breadth the most advanced AI-supported assessment cannot by outlining 30 skills ranging from adaptability to function effectively. transformative competencies.44 Liu also emphasized the importance of gathering evidence from multiple The value of robust digital data systems depends on sources and creating an inclusive system for students whether the claims, interpretations, and decisions to demonstrate their skills. Leadership, for instance, they support remain valid.39 Therefore, investment may be shown not only from being a captain on varsity in assessment infrastructure should be paired with sports teams, but also from taking care of younger investment in collecting ongoing validity evidence siblings or leading community activities. Emerging research suggests that a broader evidentiary base would be beneficial in measuring social learning, even if the work remains preliminary. Positive outcomes were observed when students engage in Without infrastructure, inquiry and constructive collaboration suggest that social processes can be documented and interpreted even the most advanced in analytically useful ways.45 Contributors also AI-supported assessment discussed developing storylines about future-focused scenarios, such as Participatory Scenario Planning, cannot function effectively. which can be used to enable the social learning process.46 Researchers acknowledge limitations in scope, but they point toward a potentially important direction for assessment by expanding the evidentiary base to include social and relational learning. Responsible Assessment in the AI Era: Key Insights from a Future-Focused Conference 13 3.4 Extending the Role of Formative Even researchers advocating for formative assessment Assessment acknowledge the risks of AI use. AI can exhibit bias resulting in outcomes that minimize or exclude Formative assessment embeds assessment in learners from the learning process.51 The technology instruction to guide next steps. Teaching and learning still generates inaccurate information, which can skew must be interactive, with teachers responding to evidence for learning.52 Ben Domingue, associate student progress and difficulties as they adapt professor at Stanford Graduate School of Education instruction.47 Contributors noted that this feedback and faculty affiliate at the Stanford Accelerator for loop can also include peer formative assessment, in Learning, cautioned that passive data collection which students exchange feedback and compare it in some types of formative assessment raises with AI-generated feedback.48 increasingly difficult ethical, privacy, and transparency questions. “The more data we’re able to collect In a panel discussion on the potential of AI to help or passively,” Domingue said, “the harder those questions harm socioculturally responsible assessment, Bryan are going to become.” A. Brown, professor of science education at Stanford Graduate School of Education, emphasized that A stronger role for formative assessment does not formative assessment should meet students where diminish the need for summative assessment. they are. To support that goal, Brown is working on The goal is to ensure that each serves a clear and Multicultural Oriented Science Adaptive Intelligence complementary purpose.53 Instead of one assessment for Classrooms (MOSAIC), a culturally responsive having the final judgment on learners’ success, AI-powered app that connects science to students’ multiple assessments can be designed to collect lived experiences.49 Contributors also discussed a and synthesize accumulated evidence of learning, five-step framework for integrating AI into formative development, and capability over time.54 assessment that builds on earlier work related to digital technologies.50 4. Advancing a Focus on Responsible Assessment T he work ahead includes reflecting on what assessment should make possible for learners. Despite advances in AI, the foundations of assessment still rest on why it exists and who it is Responsible assessment should meant to serve. This section invites readers to explore be co-designed with the people assessment design and implementation in the AI era. it affects and interpreted with 4.1 Reflecting on Responsible careful attention to context. Assessment In keeping with the theme of the convening, panelists were asked to share their thoughts on the concept of students know what they know as a source of responsible assessment. Their responses suggested empowerment. There is potential for AI-powered that responsibility begins with the people assessment personalized assessments that align tasks with an is meant to serve. Several considerations emerged individual’s cultural, linguistic, and social context as from their discussions.