Dialogue at Scale: AI, Soft Skills, and the Future of Assessment

Artificial intelligence (AI) is transforming higher education in profound ways. These changes will open up opportunities to rethink every aspect of the institution, from learner experience and business models to cost structures, improved student support, and more effective ways to reach underserved populations. The challenges will be profound, as colleges and universities rush to articulate a clear vision, establish supporting policies, and provide guidance and training to faculty and students, ensuring graduates are prepared for a workforce being dramatically remade by AI. Eventually, the challenge will become existential, as campuses confront not only what work humans will do in an AI-driven future but also broader questions about society, the economy, and the role of higher education in supporting that future.

Higher education is already facing an assessment crisis, as AI undermines the reliability of traditional artifacts used to signal competence. Surveys suggest that up to 43 percent of students have already used AI to complete assignments.^Footnote1 With a few well-chosen prompts, students can generate polished essays, passable and even stellar answers to problem sets, and viable computer code. Traditional measures of student performance—such as papers, exams, and projects—can now be created in minutes, as more and more students are discovering. In response, many faculty members would like to ban students' use of AI or even see a return to paper-based exams, seeing AI as a threat to academic integrity.

However, colleges and universities need to prepare graduates for a world in which AI is a constant presence—both as a collaborator and a competitor. AI exposes the longstanding problem with traditional assessment practices: They make earning a grade, not actual learning, the measure of success for students. More complex assessment mechanisms, such as problem or competency-based learning, require more time and effort on the part of faculty. Existing assessment approaches owe their prevalence more to scalability than to enduring efficacy, leaving them ill-suited for today's AI-driven context.

But this assessment crisis also presents an opportunity. The same capabilities that enable generative AI (GenAI) to undermine traditional assignments can also power active learning, deepen human skills, and expand access to personalized tutoring. If colleges and universities seize this moment—and recognize and integrate these capabilities intentionally—they can emerge stronger, with teaching and learning redesigned for the age of AI.

Why Dialogue Matters

Long before printed textbooks and exam halls, students at Oxford and Cambridge demonstrated knowledge in viva voce examinations—oral defenses in which they had to argue, respond to counterpoints, and make their reasoning explicit. Dialogue was the original pedagogy and determinant of learner capability and knowledge.

What made this method powerful was not the recitation of facts but the act of making students' thinking visible and explicitly revealing the contours and boundaries of their knowledge. Professors can probe students' assumptions, push them into new territory, cultivate their ability to think clearly, autonomously, and extemporaneously. As any experienced teacher can tell you, ten minutes of in-depth conversation with a student can verify whether the student has command of a topic. Of course, oral exams and in-office conversations are difficult to scale, which is why modern mass education replaced them with lectures, multiple-choice exams, and essays.

GenAI large language models (LLMs) have the potential to change the assessment equation and revive the efficacy of authentic approaches such as the oral exam. For the first time, dialogue can be replicated at scale. AI tutors can ask probing questions, adapt to the learner, role-play alternative perspectives, and give immediate, targeted feedback. What was once a rarefied experience for the privileged few—personal, responsive, Socratic engagement—can now become the core modality of learning for everyone. Faculty can review the transcripts from these learning sessions and use them to augment students' engagement with AI and evaluate how deeply they understand course topics. The conversation itself becomes the artifact of assessment.

Tutoring, Unbound

Benjamin Bloom's well-known 1984 study showed that students who received one-on-one tutoring outperformed their classroom peers by two standard deviations, effectively moving "average" students into the top 2 percent of performance.^Footnote2 As educators, we've been haunted by this "2 sigma problem" ever since: We know tutoring works, but it's too expensive and labor-intensive to offer at scale. Extensive research with intelligent tutoring systems (ITS), going back to the 1990s, has focused on trying to replicate this effect at scale. But, while the vision was right, the technology wasn't capable of dialogue-based engagement that felt natural for learners.

The current generation of LLMs solves this problem. One particularly promising use of LLMs is via specially designed AI agents. Each AI agent can perform a specific role when engaging with students. For example, an AI assessor agent can quiz learners, a practice partner can help role-play applications of a complex skill, and an instructor agent can directly guide learners through a topic. A single student can now interact with a suite of agents, each playing a different role in the learning process. We already have several good examples:

• MuDoC: a "multi-document" companion that synthesizes readings and challenges comprehension
• VISAR: an interactive assistant that generates scenarios for decision-making and judgment
• StatZ: a guide that translates statistical concepts into usable models, tailored to a learner's context
• OpenAI, Anthropic, and Google: These three big frontier labs now all have Socratic and dialogic offerings for students.

Instead of replacing teachers, these agents act as infinitely patient, always-on tutors, ready to explain, question, or critique at any hour. Imagine a student wrestling with a case study: One agent plays the skeptical investor, another the sympathetic regulator, and a third the hard-nosed analyst. The student learns not just the content but how to navigate multiple perspectives. This multi-agent approach isn't a pipe dream without a path to implementation. It's already being used in medical settings, demonstrating both impact and value.^Footnote3

In education, this approach to tutoring, once bound by cost and scarcity of resources, can now be both personal and universal—and available 24/7.

Closing the Soft Skills Gap

Employers consistently highlight soft skills, such as communication, teamwork, and adaptability, as essential, often ranking them higher than technical mastery on priority lists. The World Economic Forum's Future of Jobs Report 2025 lists analytical thinking, resilience, collaboration, and communication as the most in-demand skills.^Footnote4 McKinsey research found that higher-order cognitive, social, and emotional skills will account for 40 percent of job growth through 2030, even as routine tasks are automated.^Footnote5

Even though many soft skills are inherently social, colleges and universities still lean heavily on individual assignments, silent reading, and solo test-taking. Most institutions take a passive approach to soft skill development, treating it as a faith-based initiative: If we provide opportunities for students to engage with many other people over four years, assume leadership roles, and perhaps study abroad, we have faith that students' soft skills will improve. Employers are making the need for these skills increasingly clear and prominent in job postings.^Footnote6 Unfortunately, there is no verifiable audit trail that demonstrates which soft skills were developed through which activities. The result is a persistent soft skills gap and a lack of preparation for complex and uncertain workforce settings.

As it does with traditional assessment, AI opens a path to address this challenge. Soft skills thrive in dialogue and practice, not in lectures. Well-trained AI agents can be teammates, adversaries, or audiences. They can realistically debate, critique, or role-play high-stakes scenarios. Students can practice conflict resolution in a simulated team dispute or learn persuasion by facing an AI agent that pushes back like a skilled conversation partner or negotiator.

Soft skills development should not be treated as extracurricular; it is best cultivated through the very act of learning disciplinary content.^Footnote7 A statistics student explaining a model to an AI that "doesn't get it" practices clarity. A nursing student role-playing an intake interview with a difficult patient builds empathy and communication skills. The line between content mastery and human skill development blurs—and that's exactly as it should be. Assessment of disciplinary learning and soft skills should be integrated, with artifacts such as transcripts and learner reflections serving as evidence of students' progress.

Rethinking the Canon of Learning Activities

For centuries, colleges and universities have relied on a core set of activities, such as reading, writing, modeling, and coding. GenAI does not replace these practices—it transforms them into more interactive, adaptive experiences.^Footnote8 What might that transformation look like?

Reading → Interactive Inquiry

Instead of passively consuming readings, students engage in dialogue with the text. AI companions can quiz, summarize, or challenge assumptions. Remember Socrates' critique of writing—that a text cannot engage in dialogue with the reader, and if it contains something foolish (our words, not Socrates'), it will always be foolish? Through the power of AI, texts can now interact with students, transforming reading into great active exchanges. With the evolution of AI technologies—particularly agentic browsers—ongoing dialogue with AI across text, code, and images is becoming a key part of learning. Asking better and more focused questions becomes a key literacy.

Writing → Interactive Reasoning

Writing has always been a form of thinking. Now, as students draft essays, the AI can push back: "Your evidence doesn't support this claim, because of . . . ." "How would you address this counterargument?" "Under what conditions is this argument valid?" Each answer becomes a new statement that can be questioned or challenged by the agent. Writing becomes the output of a dialogue with an intelligent critic.

Modeling → Interactive Inference

A wide range of disciplines—from economics to engineering—rely on data models to evaluate how changes in one component impact the overall outcomes under consideration. Data models are used to explain and predict causal effects, to disentangle causal influences on outcomes, and to propose viable interventions that distinguish between different causal mechanisms. AI can guide students through the logic: "What happens if you change this variable?" and "What's the functional consequence?" Modeling becomes exploratory rather than mechanical.

Coding → Interactive Algorithm Design

As coding copilots mature, human tasks shift toward design and problem framing. Students can test their designs against AI coders that suggest alternatives, reveal bugs, or critique efficiency. The canon of learning is thus preserved, but each activity is now infused with dialogue, feedback, and adaptiveness. Learning becomes more rigorous as we raise the cognitive bar for students.

The Missing Infrastructure: The Dialogic Learning System

What we envision above is a radical rethinking of not only our assessment practices, but also our pedagogy and basic learning models. AI is a systems change technology, and the platforms that support incumbent models fall far short of what is needed for a shift of this magnitude. Furthermore, it is unrealistic to expect, despite press releases and publicly signed contracts, that the existing system will be able to absorb the potential of AI without dramatic reorganization. Our current LMSs are essentially filing cabinets: linear, static, and passive. They are designed for posting syllabi, hosting discussion boards, and collecting assignments. They are not designed for a world in which dialogue is the core of learning.

We need a new infrastructure—a dialogical LMS—built from the ground up to support AI-powered active learning. It would include the following principles:

1. Agency and adaptiveness. Every learner moves through material at their own pace, guided by adaptive feedback loops.
2. Continuous interaction. Learning is not a sequence of one-off assignments but an ongoing conversation.
3. Proactive feedback. AI doesn't just wait for input; it anticipates confusion, offers challenges, and nudges deeper exploration.
4. Transparency and trust. Learners and faculty must see why the AI responds as it does. Black-box feedback erodes credibility.

In short, we need technology platforms that make dialogue a cornerstone of the learning experience. This is why we have set out to create such a platform in our own work. Traditional LMS platforms are the buggy whips of our educational future.

Risks and Challenges

However promising and exciting that future might be, the transition will not be easy. Indeed, it will be profoundly difficult. No technological leap comes without hazards, pain, and stumbles. For AI in education, several stand out:

• Overreliance. A 2025 MIT study found that students who exclusively used AI to generate essays retained 40 percent less knowledge than their peers who used AI as a sparring partner.^Footnote9 AI must augment, not replace, cognition.
• Bias and equity. LLMs can reproduce cultural bias or privilege well-resourced institutions and populations. Without careful design, AI could widen gaps rather than close them.
• Cost and resistance. Redesigning curricula, retraining faculty, and developing new platforms require major investment. Many institutions still struggle to fund basic infrastructure. The growing AI backlash we are seeing across society will certainly play out in higher education as we ask our people to rethink their roles, work, and basic assumptions.
• Erosion of human connection. If implemented poorly, AI could isolate learners rather than connect them. The challenge is to ensure that AI scaffolds richer human-to-human interaction rather than substitutes for it.

These risks are real, but they are not reasons for paralysis. They are design challenges—problems we can anticipate and solve if we remain vigilant. The learning potential of LLMs is too significant to ignore. With proper design, added support layers, and ongoing engagement with students and faculty, we can create future learning systems that are both effective and impactful. More importantly, we can help learners develop the skills to think critically, reflect, and improve their ability to monitor and track their own learning processes. Of course, the skills acquired in dialoguing with AI will need to carry over into classroom settings. But the capability to use AI to create dialogic learning that match a learner's pace and interest holds great promise for learners.

Choosing the Future We Want

We stand at a fork in the road. One path leads to faculty paranoia and even collapse: Assessments are hollowed out, faculty chase plagiarism, students disengage as AI does the work for them, and, perhaps worse, the sector becomes increasingly irrelevant. The other path leads to renewal: Active learning at scale is realized, soft skills are embedded in every discipline, and teaching and learning experiences once reserved for the elite are made available to all.

Dialogue has always been the beating heart of education. For centuries, it was rationed—a privilege of small seminars or private tutorials. GenAI can democratize it, making personalized, dialogical learning the norm rather than the exception.

We have an opportunity to drive scaled dialogic learning. AI will not diminish higher education; it will help it deliver on its deepest promise: not just the transmission of knowledge, but the cultivation of reasoning, creativity, empathy, and human connection. In the age of AI, these distinctly human traits may be what set us apart from machines that can outperform us in much knowledge work but only mimic the human capacities that make life meaningful and rich.

Notes

1. Om Gor "Exploring Student Viewpoints on Generative AI in Education,"University of Illinois Chicago, January 8, 2024. Jump back to footnote 1 in the text.↩
2. Benjamin S. Bloom,"The 2 Sigma Problem: The Search of Group Instruction as Effective as One-to-One Tutoring,"Educational Researcher 13, no. 6 (June 1984): 4–16. Jump back to footnote 2 in the text.↩
3. Dominic King and Harsha Nori, "The Path to Medical Superintelligence,"Microsoft AI, June 30, 2025; Alan Karthikesalingam and Vivek Natarajan, "AMIE: A Research AI System for Diagnostic Medical Reasoning and Conversations,"Google Research (blog), January 12, 2024. Jump back to footnote 3 in the text.↩
4. Future of Jobs Report, (World Economic Forum, January 2025). Jump back to footnote 4 in the text.↩
5. Jacques Bughin et al., "Skill Shift: Automation and the Future of the Workforce," McKinsey Global Institute, May 23, 2018. Jump back to footnote 5 in the text.↩
6. Elizabeth Trovall, "Employers Seek Better Soft Skills from Next Generation of Workers," Marketplace, hosted by Kai Ryssdal, September 24, 2024; Moh Hosseinioun, Frank Neffke, Letian Zhang, and Hyejin Youn,"Skill Dependencies Uncover Nested Human Capital,"Nature Human Behavior 9 (February 2025): 673–687.Jump back to footnote 6 in the text.↩
7. Mihnea Moldoveanu, Soft Skills: How to See, Measure and Develop the Skills that Make Us Quintessentially Human, (DeGruyter, 2024). Jump back to footnote 7 in the text.↩
8. Mihnea Moldoveanu and George Siemens, "Interactionalism: Re-Designing Higher Learning for the Large Language Model Era,"arXiv, January 1, 2025.Jump back to footnote 8 in the text.↩
9. Nataliya Kosmyna et al., "Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Task," arXiv, June 10, 2025. Jump back to footnote 9 in the text.↩

---

Mihnea Moldoveanu is Director of Desautels Centre for Integrative Thinking and Rotman Digital and the Marcel Desautels Professor of Integrative Thinking and Professor of Economic Analysis at University of Toronto.

Tanya Gamby is Chief Health and Wellness Officer at Southern New Hampshire University.

David Kil is CEO and Founder of CML Insight.

Rachel Koblic was Chief Learning Officer at Matter and Space.

Paul LeBlanc is Visiting Scholar and Special Advisor at the Harvard University Graduate School of Education.

George Siemens is Chief Artificial Intelligence Officer at Southern New Hampshire University.