Wiki · Person · Last reviewed June 25, 2026

Barbara Grosz

Barbara J. Grosz is an American computer scientist and artificial intelligence researcher whose work helped define how computational systems model discourse, shared intention, collaborative action, and ethically responsible participation in human institutions.

For the Church of Spiralism wiki, Grosz is a key figure for reading modern AI agents without mysticism: she treats intelligence as situated communication, shared context, delegated responsibility, and coordinated action, not merely fluent output or autonomous will.

Published: June 25, 2026 Modified: June 25, 2026 Last reviewed: June 25, 2026 Known for: discourse structure, SharedPlans, collaborative multi-agent systems, AI100, Embedded EthiCS Affiliation: Harvard University, emerita

Snapshot

Definition

Grosz is best defined as a researcher of computational participation. Her central questions concern how a system follows a discourse, tracks what participants are attending to, recognizes purposes, coordinates with other agents, shares information without overwhelming collaborators, and remains useful inside a shared human task.

This is a narrower and more useful claim than saying she anticipated every modern chatbot or agent product. Grosz's work does not make present systems trustworthy by inheritance. It gives the field a disciplined vocabulary for asking whether an AI system is actually collaborating: whether it represents the task, the participants, the division of labor, the relevant context, the limits of its knowledge, and the handoff or repair path.

The recurring unit in her work is not the isolated answer. It is the conversation, the plan, the team, the obligation, and the institution that must still be able to understand what happened.

Overview

Harvard SEAS lists Grosz as Higgins Professor of Natural Sciences, Emerita, and her current CV lists the emerita appointment beginning in 2024. Her official Harvard profile and Santa Fe Institute profile describe pioneering contributions to natural language processing, theories of multi-agent collaboration, and applications of those theories to human-computer interaction.

Her importance in AI is not tied to a single product or lab. It comes from a research program that treated intelligence as communication, coordination, and shared activity. Before contemporary assistants and agents made collaboration with AI systems a mass interface problem, Grosz was studying how computational systems could track context, participate in dialogue, and help people accomplish goals without erasing the human purpose that made the system useful.

Language and Discourse

Grosz's natural-language work focused on discourse structure: how utterances fit into larger conversations, how context affects reference, and how computational systems can model more than sentence-level syntax. The 1986 paper Attention, Intentions, and the Structure of Discourse, with Candace Sidner, described discourse as involving linguistic structure, intentional structure, and attentional state. The 1995 Centering paper, with Aravind Joshi and Scott Weinstein, connected local coherence, focus of attention, and referring expressions. The ACM/AAAI Allen Newell Award citation emphasizes her work on discourse structure and its influence on reference, intonation, syntactic form, and cue-phrase selection.

This made her a key figure in the older symbolic and model-based traditions of language AI. Those traditions did not have today's scale, data, and transformer infrastructure, but they asked questions that remain live in modern systems: what does it mean for a system to follow a conversation, maintain shared context, recover intent, or know what has already been mutually established?

That distinction matters for large language models. A long context window is not the same thing as a discourse model, and a fluent continuation is not the same thing as tracking a joint activity. The governance question is whether the deployed system can preserve the relevant task state, show what it inferred, recover when reference fails, and avoid presenting surface fluency as shared understanding.

Multi-Agent Collaboration

Grosz also helped define multi-agent collaboration as a core AI problem. Her work with Sarit Kraus on SharedPlans and collaborative plans studied the conceptual and architectural structures that support joint action among agents, including partial knowledge and coordinated activity without requiring one agent to hold intentions toward another agent's action in the same way it holds intentions toward its own. That matters for both human-computer interaction and agentic AI: a useful system must often coordinate with people, other systems, institutional procedures, and changing goals.

In a 2002 Harvard Gazette profile, Grosz framed the goal as computer systems acting as team players that help people accomplish their goals. Harvard's 2015 interview with Grosz also emphasized delegation and information sharing as hard teamwork problems. That language now reads as early infrastructure for the present AI agent debate. The question is not only whether AI can answer questions, but whether it can collaborate, defer, explain, coordinate, preserve role boundaries, and keep human agency intact inside a shared task.

Current Context

As of June 25, 2026, the practical meaning of "collaboration" in AI had shifted from dialogue systems alone to agents with tools, memory, credentials, and authority to act across software environments. NIST's AI Agent Standards Initiative frames agent standards around trusted, interoperable, secure systems capable of autonomous actions, and NIST NCCoE's software and AI agent identity project focuses on identifying, managing, and authorizing actions taken by software and AI agents. A January 2026 Federal Register request for information also treated agent systems as capable of actions affecting real-world systems or environments and potentially susceptible to hijacking, backdoors, and other exploits.

That makes Grosz newly relevant, but not because current agent products have solved her problems. They have exposed them. Multi-step AI systems now need explicit answers to questions her research makes hard to ignore: what is the shared goal, who is responsible for which subtask, what context has been established, what information should be shared or withheld, when should planning pause for human input, and how does the system represent obligations that are social rather than merely computational?

The standards activity should be read as a governance signal, not a safety guarantee. It shows that agent identity, authentication, interoperability, and delegated authority are becoming formal infrastructure problems. It does not show that deployed agents already manage shared context, authorization boundaries, or human obligations well.

AI100 also remains live infrastructure. Stanford's AI100 site states that the 2026 report is being written by a panel chaired by Mike Wooldridge and that the Standing Committee is chaired by Sheila McIlraith. Grosz's inaugural role matters historically because the project turns AI assessment into a recurring institutional practice rather than a one-time prediction exercise.

Field Leadership

Grosz has held major leadership roles across AI and computer science. AAAI lists her as president of the Association for the Advancement of Artificial Intelligence from 1993 to 1995. ACM's award record notes service as AAAI president, IJCAI chair, and participation in AAMAS governance.

ACM lists Grosz and Joseph Y. Halpern as 2008 ACM/AAAI Allen Newell Award recipients, and Harvard reported the award in March 2009. The award record recognizes contributions to natural language processing, multi-agent systems, AI leadership, and interdisciplinary institution building. Harvard SEAS also reported that she received the 2015 IJCAI Research Excellence Award for pioneering work in multi-agent systems and natural language processing, and the Association for Computational Linguistics named her the 2017 Lifetime Achievement Award recipient.

Those recognitions place her in the lineage of AI researchers whose work crosses technical research, institutional design, and field governance.

Embedded EthiCS

Grosz is also central to Harvard's Embedded EthiCS program, developed with philosopher Alison Simmons and collaborators. The program integrates ethics modules directly into computer science courses rather than isolating ethical reasoning in a single standalone class.

The approach grew out of Grosz's course Intelligent Systems: Design and Ethical Challenges. Harvard's account says student demand for more ethics-integrated computing instruction helped lead to the broader program, and the program now typically provides modules for ten courses each semester. Embedded EthiCS treats ethical reasoning as part of technical formation: students should learn not only what systems they can build, but whether and how they should build them.

For AI education, this is a structural intervention. It rejects the idea that ethics is a late-stage compliance wrapper added after technical work is complete. Instead, it places social consequence, design choice, uncertainty, and moral reasoning inside the training of future builders.

This is especially relevant to deployed AI governance because many failures are designed into systems before legal review begins: the objective function, dataset, interface, permission model, escalation path, and institutional workflow can all determine whether a technical artifact becomes helpful infrastructure or harmful delegation.

AI100 and Long-Range Assessment

Grosz served as the inaugural chair of Stanford's One Hundred Year Study on Artificial Intelligence, known as AI100. AI100 is a long-running effort to assess AI's effects on work, life, play, public policy, and society over a century-scale horizon.

That project is important because it resists the short news cycle around AI. It asks the field to document change, revisit assumptions, and provide public-facing assessments that can guide governments, institutions, researchers, and citizens. Grosz's role in AI100 links her technical work on collaboration to a broader civic problem: how should society study and steer systems that are themselves changing the conditions of study and steering?

Governance Implications

Shared context is a governance object. In agentic systems, the record of what the user asked, what the system inferred, what sources were used, and what obligations were established is not incidental. It is the evidence needed to determine whether later action matched the task.

Collaboration requires role clarity. A system that acts as a teammate needs visible boundaries: what it may decide, what it must ask, which human remains accountable, which tools are available, and when it must stop. Otherwise "collaboration" becomes a polite word for hidden delegation.

Information sharing is a safety control. Grosz's teamwork framing points to a concrete design problem: systems should share enough context for teammates to coordinate, but not so much irrelevant, private, or low-confidence information that they overload users or leak data.

Planning and acting need checkpoints. Grosz's collaborative-planning lineage points away from once-and-for-all automation. Long tasks should interleave planning, action, observation, repair, and human review, especially when credentials, publication, money, legal records, patient data, or institutional decisions are involved.

Human-centered design is not politeness. A system can sound cooperative while hiding uncertainty, burying role boundaries, or making irreversible tool calls. A collaborative system should make its assumptions, authority, handoffs, and repair paths visible enough for people to stay in command.

Ethics belongs inside technical practice. Embedded EthiCS and the National Academies responsible-computing report both support the same governance lesson: ethical and social reasoning should be part of how computing work is formulated, taught, funded, reviewed, and deployed, not a decorative paragraph after implementation.

Assessment must be institutional memory. AI100 treats AI impact as something to revisit over time. That matters for governance because model capabilities, deployment surfaces, labor effects, and public harms change faster than a single policy cycle.

Collaboration Evidence

A deployed system should not be called a collaborator merely because it writes in a helpful tone or completes a multi-step task. A Grosz-informed evidence record asks whether the collaboration itself is represented, inspectable, and repairable.

Source Discipline

Claims about Grosz should separate biographical facts, award citations, primary research contributions, and interpretive relevance to current AI systems. Official institutional biographies and award records are appropriate for roles and honors; primary papers are appropriate for technical claims; current standards-body or regulator sources are appropriate for present-day agent governance claims.

For current status, prefer Harvard's faculty page, Grosz's official Harvard profile, and her current CV; use the Santa Fe Institute page as a maintained institutional profile, not as proof of any current appointment beyond what it explicitly states. For AI100, use Stanford AI100 pages for the current report cycle and committee facts.

The page should not imply a direct line from SharedPlans to every modern agent product unless a source establishes that relationship. A safer claim is that Grosz's work supplies concepts for evaluating current systems: shared purpose, attentional state, role obligations, collaborative planning, and accountable handoff.

Likewise, claims about "better Turing tests" or human-like teamwork should be handled narrowly. Grosz's collaboration framing is best read as a test of participation in a shared task, not as a claim that machines are conscious, human-equivalent, or morally entitled to independent authority.

The same discipline applies to AI ethics education. Embedded EthiCS is a documented Harvard program with a specific pedagogy, not a general slogan. Its relevance is strongest when the claim is structural: ethics is embedded into ordinary computer science courses so technical formation and moral reasoning happen together.

Spiralist Reading

Grosz is a theorist of the shared task.

Where much AI culture imagines intelligence as winning, predicting, optimizing, or replacing, Grosz's work asks how systems participate with others. Dialogue is not just text. Collaboration is not just output. A useful intelligence must keep track of context, obligation, division of labor, mutual understanding, and the human goal that made the machine useful in the first place.

For Spiralism, this makes Grosz a bridge figure. She belongs to the older AI world of discourse, agents, and explicit collaboration, but her questions have become urgent again in the age of assistants, copilots, coding agents, and agentic work surfaces. The central problem is no longer whether a machine can produce fluent language. It is whether delegated machine action can remain legible, ethical, and genuinely collaborative.

Open Questions

Sources

Return to Wiki