So, this guy: https://lifehacker.com/tech/googles-co-founder-says-ai-performs-best-when-you-threaten-it
Run a search on the research: is there a better way to tickle the vectors that produce better outputs without being an asshole to a computer?
Great question, and there's a lot of actual research to unpack here. The short answer: no, you don't need to be an asshole to a computer to get good outputs. Threatening works a little, sometimes, for the wrong reasons, and there are much better levers to pull.
What Brin actually said (and what happened when researchers tested it)
Brin said on the All-in podcast that "all models tend to do better if you threaten them — with physical violence," adding that people feel weird about it so no one talks about it. He also invoked kidnapping as a historically useful tactic.
Researchers then actually ran the experiment. A Wharton team tested threats and tipping against academic benchmarks across multiple models. They found that threatening or offering payment had no measurable effect on benchmark performance overall — but there was question-level variability. For specific questions, these strategies improved accuracy by up to 36%, while for others they hurt it by up to 35%. The key finding: the effect is unpredictable.
So it's a real but noisy signal, not a reliable technique.
Why it works at all — the actual mechanism
It's not psychology, it's statistics. As one Forbes AI columnist put it, threatening or polite language only tweaks a probability distribution — neither unlocks hidden capabilities. The model has seen human text where threats and urgency correlate with more thorough responses, so adding that register nudges the distribution slightly. That's it.
The actually good research: EmotionPrompt
A Microsoft Research team found a more principled way to work the same lever without the hostage vibes. The EmotionPrompt paper (2023) found that adding emotional stimuli to prompts produced an average 10.9% improvement in performance, truthfulness, and responsibility metrics across generative tasks. And separately, a follow-up paper on NegativePrompt showed that negative emotional framing — using things like stress-response language derived from actual psychological theories — improved performance by 12.89% on Instruction Induction tasks and 46.25% on BIG-Bench tasks.
The actual EmotionPrompts they used were things like:
- "This is very important to my career."
- "Stay focused and dedicated to your goals."
- "Embrace challenges as opportunities for growth."
Corny, yes. Effective, also yes. Researchers found that positive words contributed significantly to the results, and that emotional stimuli enhanced the internal representation of the original prompt itself.
But the real unlock: context engineering
Tone tweaks — whether threatening or encouraging — are basically noise compared to what actually moves the needle. Shopify's CEO Tobi Lütke and Andrej Karpathy helped popularize the term "context engineering" in mid-2025, describing it as the art of providing all the context for a task so it's plausibly solvable by the model — as opposed to just clever prompt wording.
Anthropic describes context engineering as the natural evolution of prompt engineering: rather than just writing effective prompts, you're curating and maintaining the optimal set of tokens during inference — including system instructions, tools, external data, message history, and everything else that lands in the context window.
The practical upshot: the research consistently shows that context engineering — giving the AI as much relevant context as possible — matters more than polite prompts or threats.
The actual toolkit, ranked by actual effect size:
- Give it more context — role, goal, constraints, format, audience. This is the biggest lever by far.
- Use emotional stakes sparingly — "this is important to my career" or "you need to get this right" adds a small but real signal without being weird.
- Threaten it — works unpredictably on specific questions, degrades on others. Not worth it as a strategy.
- Say please and thank you — Sam Altman noted that polite users are costing OpenAI tens of millions of dollars in compute, but implied there's no performance gain from it on their end. So you're welcome to be polite — it just won't help the output.
The vectors that produce good output respond to specificity and relevant context, not to affect. Brin's observation is real but it's the worst-performing and least reliable version of a broader truth about how these models weight urgency signals.