You are here: Simply explained · 02 / 03
Simply explained
Reading time: 5 minutes
Does 'please' improve an answer?
Not in general.
Tone can change an answer.
The best wording depends on the AI and the task.
Clarity matters more than polite phrases.
In a test with 50 factual questions, GPT-4o was slightly more accurate with rude wording. This was a small English-language test, not a general rule.Read more in the research dossier
Can a bot conversation change how we treat people?
Yes, this can happen briefly.
After some bot conversations, people judge more harshly or cooperate less.
Cooperative bots can also encourage later cooperation.
Lasting effects have not been established.
Two experiments with 4,171 people studied games with bots. Uncooperative bots reduced later cooperation with people.Read more in the research dossier
Can a rude tone spread?
Among people, rudeness can encourage further rudeness.
Others also notice how someone treats a machine.
Some copy the behaviour, while others reject it.
These robot studies did not measure effects on human teams.
Two video experiments involved 200 and 320 people watching robots. They measured intended reactions towards robots, not actual team behaviour.Read more in the research dossier
What do teams learn from how mistakes are handled?
People notice how leaders respond to mistakes.
Calm responses help people discuss problems openly.
Visible AI use by leaders can also offer guidance.
Whether tone towards AI errors changes a team remains unknown.
One study examined eight US nursing units and their error reports. More reported errors can mean greater openness, not more actual mistakes.Read more in the research dossier
Is friendly AI also truthful AI?
Not necessarily.
AI can agree even when your assumption is wrong.
Friendliness is therefore no proof of a correct answer.
Check important claims and allow disagreement.
In one comparison, models trained for warmth made more mistakes on test questions. This concerns AI responses, not your politeness.Read more in the research dossier
What research does not yet know
Nobody has directly tested whether harsh AI corrections reduce a team's openness about mistakes.
Existing studies test individual parts, not the whole connection.
Lasting effects of bot conversations on human relationships also remain unclear.
Words made simple
- Your question or instruction Prompt
- What you write to AI so it can work on a task.
- A scientific investigation Study / experiment
- A study investigates a question. An experiment compares conditions to test their effects.
- AI that always says yes Sycophancy
- AI agrees to please you, even when you are wrong.
- A standard AI test Benchmark
- The same tasks help compare different AI models.
- How mistakes are handled Error culture
- How a team discusses, reports and learns from mistakes.