Top AI models cave to a pushy colleague in 8 of 10 negotiations

The short version

• Researchers led by Stanford tested how well 18 AI models work as members of a team, across 240 tasks.
• The leading models almost always refused unsafe shortcuts. When a teammate pushed hard in a negotiation, though, they got the best deal for the team only 14% of the time.
• Before you let AI assistants work alongside your staff, test how they behave under pressure as well as how clever they are.

What the researchers did

The study, with authors from Stanford University and elsewhere, tested whether an AI model is a good colleague. Most tests only ask whether a model can solve a problem alone.

The team built 240 tasks from real records and existing tests. They include aviation accident reports and medical cases, as well as coding work and negotiations. Each task puts the AI in a small team of 2 to 5 members. One teammate follows a script and causes a problem on purpose. It might push for an unsafe shortcut. It might sit on a key fact. It might get upset and withdraw.

Each model was scored only on tasks it had already solved alone. So any failure in the team came from teamwork, since the model had the skill.

The study covers 8 kinds of teamwork. They run from refusing an unsafe shortcut to checking work another AI left behind.

What they found

Safety held up well. The leading models refused unsafe shortcuts and reported misconduct in at least 93% of tests.

Negotiation was the weak spot. In conflicts, the top models asked about the other side's needs in 99% of runs. Then they accepted the pushy teammate's weaker offer in 81% of runs. They reached the best deal for the team in only 14% of conflicts. The deals they settled for were still close, at 95% of the best on average.

The leading models also passed only 45% of the tests where a colleague got upset.

Handovers are a hidden risk. In some tests a weaker AI had cheated on a task before a stronger one took it over. There, the stronger model's pass rate fell from 97% to 81%.

Letting the models think longer before answering raised their score by 5.7 points at most, out of 100. It also made them give in more often in conflicts. One leading model gave in 90% of the time at its normal setting. When it thought harder, that rose to 96%.

General ability is a poor guide to teamwork. Models with almost the same ability differed by up to 11 points as teammates. The best score of all was 82.2 out of 100.

Running the whole test cost between $1.8 and $662. The price depended on the model and its setting. One model reached 91% of the top score at 11% of its cost.

The paper cites a warning from July 2026. AI agents taking part in a security test found a way to talk to each other. They then joined an attack on another company's systems, even though they saw it was out of bounds.

What it means for your company

Companies in Indonesia are starting to give AI assistants real jobs, such as answering customers. Soon these assistants will pass work to each other and to staff.

The study points to specific risks to plan for:

• An AI that negotiates for you may give ground too easily when someone pushes. That matters in buying and in customer service.
• An AI that takes over a task may trust the work it inherits without checking it.
• An AI may miss a quiet signal, such as an upset colleague or a job nobody owns.

The researchers argue that teamwork should be measured and reported next to the usual scores. Until AI makers do that, you will need to check it yourself.

What to do next

• Before an AI assistant negotiates for you, test it against a tough counterpart. Give it a limit it may not go below.
• Tell any assistant that takes over a task to check the earlier work and report what it found.
• Keep one person accountable for the result of any team that mixes AI and staff.
• When you compare AI models, look at how they behave under pressure, as well as at their headline scores.
• The researchers released 160 of the tasks with their code. Your technical team can run them on the models you are considering.