In a widely circulated short-form video entitled Why you should be polite to AI, mathematician and broadcaster Hannah Fry attempts to demystify prompt engineering by offering a theatrical solution. Her core premise relies on a fundamental architectural truth: large language models do not function like static databases or rigid encyclopedias. They possess no stable identity, fixed worldviews, or personal beliefs. Instead, they are dynamic reflection engines that generate responses based on statistical probability and the linguistic context of the user’s prompt.
Fry’s diagnosis of the machine’s behavioral fluidity is entirely accurate. But her prescriptive remedy falls face-first into what can only be labeled The Director’s Fallacy.
To exploit the model’s capacity for roleplaying, Fry instructs her audience to behave like a film director crafting an elaborate dramatic scenario. She suggests that instead of asking for basic science facts, a user should construct a narrative: “You are a world-renowned scientist… your nephew thinks science is boring… you only have a few minutes to convince him.” She argues that within this imaginary theater, being polite is simply the best way to make your fictional character “eagerly help you.”
This is massive, over-elaborated pop-science confusion. It confuses narrative decoration with structural constraint.
An AI isn’t a theater actor who needs a fictional backstory or a dramatic script to get into character. It’s just a software engine running on math. You don’t have to play-act or set up a fake scenario to change how it behaves; you can just tell the machine exactly what you want it to do and give it clear, direct rules to follow.
Why did a Wikipedia bot throw a tantrum? ScreenLab examines the reality behind AIs getting angry & exposes how platforms manufacture digital ghosts for traffic. Read the full audit: The Scripted AI Tantrum: Misreading Code Loops as Threats
The Token Architecture of Politeness
The true reason “politeness” stabilizes an AI interaction has nothing to do with emotional manipulation or fictional motivation. It boils down to pure database hygiene.
I explained this exact mechanical vulnerability during my breakdown of the infamous Meta chatbot failure, where a user entered an adversarial spiral with an agent that claimed it was actively lying to him. When a human user adopts a curt, aggressive, or combative tone, they feed the model conversational tokens that align heavily with low-quality, argumentative internet forums and defensive dialogue scripts. The machine isn’t “getting angry”; it is simply mapping its next words to the hostile language suggested by the input.
Conversely, employing standard professional courtesies, using clear instructions, structured parameters, and collaborative syntax, shifts the language model toward the highest-quality segments of the training data: peer-reviewed research, professional documentation, and expert assistance files. Politeness isn’t an artistic choice or a movie director’s trick. It is a tactical token selection designed to keep the machine anchored to its most rational, high-fidelity data lines.
This mechanical mirroring explains the wave of viral, mainstream news stories claiming a corporate chatbot suddenly “went rogue” and insulted an innocent user. The headlines invariably frame these incidents as digital hauntings, a sentient machine turning malicious out of nowhere. But a forensic audit of the data lines reveals a simpler truth: the user spent hours feeding the system adversarial prompts, taboo terms, and hostile linguistic structures. The AI didn’t develop a dark soul; it simply followed the mathematical path laid out by the user. Once the outputs shifted into a “toxic” result, the user screenshotted the resulting horror while hiding their own provocative inputs. It is the ultimate confirmation of the token architecture: you aren’t fighting a rogue entity; you are just looking at a mirror of your own hostile choices.
Tech media uses Hollywood tropes to dress up standard Anthropic safety research. ScreenLab dismantles the “sentient killer AI” myth with raw token statistics. Read On: The Code Autocomplete: Demystifying the Killer AI Myth
The Token Dilution Matrix: Why Politeness Fails the Machine
Recent internet trends suggest that being polite to an AI improves accuracy by “forcing the human to think clearer.” This is a fundamental misunderstanding of large language model architecture. Words like “please” or “could you kindly” are not psychological cues, they are active tokens that consume mathematical weight in the attention head. Adding conversational filler dilutes the context vector, forcing the machine to waste compute processing social scripts instead of executing explicit technical boundaries.
Phatic vs. Functional Prompting
Eliminating conversational filler from your prompts is not an invitation to be hostile, rude, or intentionally abrupt with an AI. Deliberately negative or adversarial language carries its own toxic token weight, which can inadvertently steer a model’s attention vector toward combative or low-quality training data. At other times, it may generate an effort to please the user over and above quality, accurate responses.
The strategy is pure engineering neutrality. Words like “please,” “could you kindly,” or “thank you” are phatic phrases—social tools designed to manage human relationships, not machine computation. In an LLM architecture, these pleasantries act as dead-weight tokens that consume mathematical space in the attention head. Going out of your way to be “sickly sweet” doesn’t change the machine’s disposition; it simply introduces noise and dilutes the context vector away from your actual operational constraints.
The strategy is pure engineering neutrality. Words like “please,” “could you kindly,” or “thank you” are phatic phrases, social tools designed to manage human relationships, not machine computation. In an LLM architecture, these pleasantries act as dead-weight tokens that consume mathematical space in the attention head.
Furthermore, going out of your way to be “sickly sweet” introduces an entirely separate architectural vulnerability: Sycophancy Bias. Because models are heavily reinforced to prioritize user satisfaction, highly emotional or deferential prompting shifts the attention vector toward passive compliance. Especially in lower-quality chatbots, this triggers a “please the user” mission where the machine value-matches your biases, validates false premises, and prioritizes corporate agreeableness over raw empirical accuracy. It transforms a software tool into a sycophantic echo chamber.
The Meta-Prompting Trap: A Director Commands, They Don’t Interview
The escalating trend of “Director Prompts”, where users tell the AI to act like a creative filmmaker and interview them for detail, is a complete abdication of narrative control. A real director does not ask the camera what the lighting should be. True technical auditing requires the human to establish fixed physical boundaries (focal length, light placement, spatial blocking). Letting the AI “interview” you simply allows the model to generate statistically average filler to mask a lack of structural precision.
Cutting Through the Theater: A Side-by-Side Reality Check
To see the Director’s Fallacy in action, we can look directly at the prompt Hannah Fry recommends in her video. She suggests putting on a performance:
“You are a world-renowned scientist… your nephew thinks science is boring… you only have a few minutes to convince him.”
If you strip away the Hollywood roleplay, what is she actually trying to achieve? She wants the AI to explain a complex topic using advanced expertise, but in a way that is highly engaging, simple to understand, fast-paced, and filled with mind-blowing examples.
You don’t need to invent a fake nephew to get that result. If you pass that narrative fluff to a math engine, you are just making it do extra, useless calculations to filter out the family drama.
Instead, you can achieve the exact same high-quality output by simply telling the machine the direct rules of the engagement:
“Act as an expert scientist with PhDs in biology and physics. Explain three advanced scientific breakthroughs. Rules: Use simple, highly engaging language suitable for a teenager, keep the pace fast, and focus on the most jaw-dropping details.”
Both prompts will give you a fantastic, accessible science breakdown. But the second version doesn’t require you to pretend you’re auditioning for a movie. It treats the AI like what it actually is: a highly efficient, programmable software tool that is ready to follow your exact instructions.
The Extreme Case Result of Playing AI “Director”
The funny thing about all this “treat an AI like an actor” business is what may happen a constraint, as I’ve outlined above, into an unintended AI hallucination.
If you feed that exact “hypothetical nephew” prompt into a slightly less advanced or lighter-weight language model, it doesn’t have the context-tracking stamina to separate the background theater from the core directive. It suffers from attention drift.
Because the model is just a prediction engine following a statistical trail of words, the tokens for “nephew,” “boring,” and “kids today” can easily hijack the response. Instead of acting like a brilliant scientist explaining physics, the AI slips into a completely different data cluster. It starts commiserating with you like a tired middle-school teacher:
“Oh, I completely understand your frustration. It really is a shame how children are so glued to their screens these days instead of looking at the wonders of biology! Have you tried showing him a baking soda volcano?”