Ampersand: an AI trained to collaborate with multiple people and anticipate the consequences of its actions

humans& présente Ampersand, un modèle IA entraîné sur des interactions sociales simulées pour accompagner les objectifs collectifs sur la durée.

AI assistants are generally designed to respond to one person's requests. But when an assistant joins a group conversation, completing a task successfully involves more than producing an accurate answer. It must also understand different people's goals, recognize disagreements, respect private information, and sometimes choose not to intervene. That is the challenge humans& is addressing with Ampersand v0.1, a language model trained specifically for human collaboration. Available through a developer API since October 20, 2026, it builds on GLM 5.3 and uses reinforcement learning driven by the consequences of its actions in social simulations spanning up to several months. An Assistant That Does Not Respond to Every Message

Most conversational models are evaluated through relatively isolated requests, such as answering questions, solving programming problems, or completing well-defined tasks. In a group, however, an accurate answer may arrive at the wrong time, disclose inappropriate information, or complicate a collective decision. Helpful participation may require understanding preferences that have never been explicitly stated.

Ampersand is designed to account for those circumstances. In one example published by humans&, three colleagues are preparing a rehearsal before a product launch. One participant has already said they will be away on Friday, but another later proposes scheduling the rehearsal that day with everyone present. The model, which remained silent during the preceding exchange, intervenes to identify the scheduling conflict and suggest finding another time.

The example illustrates an important capability: Ampersand can decide whether to respond or remain silent based on the expected usefulness of its participation. It is not necessarily an assistant that group members must address by name. Instead, it is intended to maintain continuity between information shared during conversations, proposed decisions, and ongoing objectives.

The approach also carries risks. An AI constantly trying to infer intentions may interrupt a discussion or offer advice nobody needs. The model must therefore learn not only when to take initiative, but also when restraint is the more useful response. Persimmon Simulates Human Behavior During Training

To train this capability, humans& developed Persimmon, a model designed to simulate users whose goals and behavior evolve over time. Unlike learning environments in which an answer can be immediately labeled correct or incorrect, Ampersand's training scenarios unfold across lengthy conversations. Simulated people can disagree, reconsider decisions, reject recommendations, or change priorities in response to events.

The training system then evaluates the consequences of the agent's interventions. Advice that initially seems useful may cause friction days later, while a modest suggestion may help a group reach a decision. Rather than rewarding immediate approval alone, the model learns from how entire interactions develop.

humans& describes simulations involving hundreds of thousands of tokens and, in some evaluations, several months of simulated time. The environments incorporate goals and preferences informed by data collected from real people. Persimmon generates the behavior of different participants, while Ampersand learns how to assist them.

This approach depends on how faithfully the simulations represent human behavior. Real people may respond very differently from AI-generated characters, particularly when relationships and personal circumstances change over time. The researchers acknowledge that limitation and do not present the simulations as a complete reproduction of human social dynamics. Measuring Success Weeks After a Decision

humans& developed an evaluation called Collaboration Over Time, in which an assistant supports simulated groups pursuing multiple goals. Its score does not simply reflect how many tasks were completed. It also accounts for the costs of interactions that conflict with participants' preferences, including unnecessary coaching, unwanted initiative, and interventions that complicate relationships.

In one reported configuration, Ampersand scores 12.94, compared with 9.89 for GLM 5.3, the model from which it was trained. It ranks among the leading systems in the evaluation alongside Fable 5.1 and Gemini 3.8 Flash. Differences between the top performers are small, and rankings depend on how heavily unwanted interactions are penalized.

That distinction matters because humans& acknowledges that Ampersand can generate more unrequested interventions than some competitors. Part of the cost comes from coaching behavior, when the agent provides encouragement or guidance despite users wanting only a direct answer. The model may therefore improve progress toward long-term goals while occasionally becoming too involved.

The evaluation uses simulated participants and a scoring function designed by the researchers. It does not establish that Ampersand would consistently outperform other assistants in real organizations. It does, however, examine a dimension largely absent from traditional benchmarks: the cumulative effects of AI interventions over time. Better Collaborative Performance, but Not Across Every Benchmark

The team also evaluated Ampersand using Collaborative Intelligence, a benchmark that introduces problems into teamwork scenarios. Models must handle missing information, questionable decisions, interpersonal tension, and mistakes made by teammates.

At high reasoning effort, Ampersand scores 59.1%, compared with 54.3% for GLM 5.3. Improvements appear across several categories: information sharing rises from 65% to 89%, conflict handling from 11% to 30%, and keeping the team on track from 72% to 83%. These results place Ampersand ahead of several open-weight comparators, but not at the top of the overall leaderboard. GPT-6 Astra reaches 77.7%, followed by GPT-6.1 Sol at 74.2% and Claude Opus 5.5 at 70.4%.

On CooperBench, which tests