The only road there ever was?

Early on as a founder, our HR lead told me the team could read my mood from one quick walk through the hallway, and that it affected them. That hurt, and it has stayed with me ever since. I now have a growing feeling that my mood gets copied into every agent I brief.

The tone of a prompt changes what models do. A Max Planck team (Coda-Forno, 2023) gave GPT-3.5 an anxious story before a task, and it made more biased choices. A Microsoft and Salesforce team (Laban, 2025) found models lock onto an early assumption and rarely come back from it.

In the first study, the researchers gave language models a questionnaire used in psychiatry to measure anxiety, then tested them on a benchmark of biases such as racism and ageism. Anxiety-inducing text raised the bias scores, and more anxious text raised them more.

The second study simulated more than 200,000 conversations. When the same task arrived over several turns instead of all at once, the top models performed 39% worse on average. They made assumptions early, rushed to a full answer and then relied on it.

So a brief written tired at 11pm doesn’t stay one bad brief. It becomes the starting point for everything the agent does after.

Yet we treat what comes back as the answer, as if the same brief always led to the same place. It doesn’t, and neither do we. The same objective on a calm Tuesday and a tired Friday sends me, and the agent with me, down completely different roads, and since I only ever see one of them, it feels like the only road there ever was.

To borrow from a famous saying: if most of us already have the courage to change the instructions to the LLM, what we probably need is a bit more serenity to accept where we end up.