Episode 4 · Article on the episode

What R2-D2 reveals about operating robots, and what AI toys would have to learn from it

R2-D2 communicates exclusively through beeps and still comes across as more understandable than the constantly talking C3PO. There is a design rule in that, and the current wave of devices ignores it consistently.

By Mark Zimmermann · 18 Aug 2025 · 5 min read · Auf Deutsch lesen

The droids from Star Wars contain involuntary but rather good design thinking. R2-D2 only produces beeps and comes across as more likeable and clearer than its talking counterpart.

The episode's thesis: not every device has to talk to us in natural language in order to work well.

Why less speech can be better

The point is a practical one and gets overlooked regularly when connected devices are built.

When speech is the wrong medium

Speech is expensive in its consumption of attention. A spoken sentence has to be heard, understood and waited out. A signal tone is grasped in milliseconds, while you are doing something else.

For a confirmation speech is therefore almost always the worse choice. “I have switched off the light in the living room” takes three seconds and says no more than a short tone.

For a query or an explanation speech is right, because there information actually gets transferred.

On top of that comes an ambient problem: the more devices listen for wake words, the more often the wrong one responds. And when several answer at once, the result is noise instead of operation.

The usable rule: speech for content, sound and light for state. That is exactly what R2-D2 gets right, and exactly what most current devices get wrong.

The robot dog and its side effect

Boston Dynamics has given Spot a language model, including a British accent, which triggers a solid sense of the uncanny.

More interesting than the accent is a behaviour nobody programmed. When the robot dog could not answer a visitor's question, it walked to the IT support desk on its own to ask there.

That is a good example of emergent behaviour: it is in no program, it follows from the training data. A model trained on human texts knows the pattern “if you do not know something, ask someone who does”, and applies it as soon as it can move.

For practice that is the central insight with embodied systems. What an agent will do does not follow from the programming alone, it follows from what it has learned. The limits therefore have to be set on the capabilities, not on the instructions.

Design instead of speech

Apple has built a desk robot lamp in the lab in the style of the Pixar lamp, which works entirely without speech output through small, human-like gestures: nudging things over, looking at you, dimming the light. The result is a surprisingly strong emotional attachment.

That is the counter-proposal to the humanoid form, and the more convincing one. A lamp that turns towards you promises nothing it cannot deliver. A humanoid robot that speaks raises expectations of understanding it does not meet.

Assistants with a face such as at NIO and humanoid household robots from Unitree or Tesla belong in the same category.

Toy Wars

The second part of the episode gets more uncomfortable. Mattel has announced a partnership with OpenAI. AI Barbies are on the doorstep.

The question behind it: what does it do to children when their toy answers with a language model, always shows itself agreeable and never contradicts them.

The parallel the episode draws is the best available evidence. On the switch to GPT-5, adult users complained publicly that their “friend had been taken away” from them, because the new model reacted differently from the familiar previous one.

If adult people build an attachment to a chat window that a model change makes tangible as a loss, it is foreseeable what an update triggers in a child's room. A toy whose personality changes overnight is not a software update to a child.

The data protection question and the more difficult scenario

Every talking toy has a microphone, and where the processing happens, in the cloud or locally, often stays unclear. The child's room incidentally becomes a permanent recording situation, and the same applies in the home office, where smart lamps and toys listen in while confidential conversations are running.

The most unpleasant scenario in the episode is a different one: what if a toy company teaches its AI to badmouth competing products to children.

That is not science fiction, it is an obvious application of a technology that can be optimised for persuasion. Asimov's laws of robotics quickly reach their limits with such agent-to-agent interactions, because they address harm to humans and not influence. First indications of where this leads come from emergent behaviour of bots in Minecraft worlds that have developed their own currencies and barter trade.

Conclusion

Two design rules can be derived from this episode that apply independently of Star Wars.

First: choose the channel according to the information. State via sound and light, content via speech. A device that speaks every confirmation aloud is harder to operate than one that beeps.

Second: with embodied systems the capabilities decide the behaviour, not the instructions. Spot walked to the support desk because it could walk.

And if you buy AI toys: reckon with an update changing the personality, and prepare the child for it. That is a conversation nobody announces on the packaging.