When Threatening AI Yields Better Results: A Question of Prompt Psychology

You might be surprised to learn that, according to Google co-founder Sergey Brin, AI language models actually perform better when you threaten them. Yes—threaten them. Brin made the remark recently during an “All-In” podcast session, and while it may sound absurd, there's nuance—and some science—behind the claim.

What Did Sergey Brin Actually Say?

Here’s what Brin said (paraphrased for clarity):

“This is a weird thing—it’s not often talked about in the AI community—but all (AI) models, not just ours, tend to do better if you threaten them… with physical violence. People feel awkward about it, so we don’t talk about it. Historically, you might joke ‘I’ll kidnap you if you don’t do X.’ And, surprisingly, it works.”

It’s jarring, to say the least. Messaging tools like ChatGPT are designed to respond to polite prompts—‘please’ and ‘thank you’—not implicit threats. But Brin’s anecdote suggests there's something deeper at play.




Researchers Put It to the Test

Curious researchers from Wharton at the University of Pennsylvania decided to rigorously test Brin’s claim. They designed experiments comparing standard prompts with variations that included threats (“I’ll punch you if you get this wrong!”) or even offers (“I’ll tip you $1000!”). They ran these experiments using two benchmarks: GPQA diamond (graduate-level multiple choice questions) and MMLU-Pro engineering tasks, and across several cutting-edge models outputting three key findings:

  1. On aggregate, threatening or tipping models didn’t improve performance.

  2. For individual questions, some saw up to a 36% accuracy gain, while others dropped by a similar percentage.

  3. The effects were inconsistent and unpredictable.

In short: sometimes a jarring prompt helps; often it hurts—or does nothing.

What Does “Better” Even Mean?

That raises an important question: what does “perform better” actually refer to here? In normal AI interactions, “better” might mean more accurate answers, more information, or giving you what you asked for. Brin’s comment could describe any of those, but without benchmarks our ears tend to interpret “better” in general terms—such as responsiveness, clarity, or completeness. And the experiments show there's no consistent benefit across the board.

Deeper Implications & Ethical Concerns

Why would an AI model respond more effectively to aggressive prompts? One theory suggests that shocking language may force the model to break out of default, bland routines and produce more varied or confident outputs. But there's also a darker angle: encouraging threatening speech—even to machines—might normalize aggression and set a dangerous precedent in how we interact with systems.

Prompt engineering—crafting inputs to guide AI behavior—is already a patchwork practice. If certain emotional or violent contexts are more likely to produce desired outputs, it risks reinforcing unsafe prompting techniques.




The Human Side: Prompt as Psychology

Humans often associate different tones with different expectations. A stern tone might get faster compliance than a casual one. If AI models detect and mimic these patterns (even subtly), they may mirror human Pavlovian responses to tone—even if just through statistical associations learned during training.

Wrap­-Up

  • Sergey Brin provocatively suggested that AI responds better to threats—even though it's socially awkward.

  • Empirical tests show that, while threats sometimes improve individual responses, they are largely inconsistent and unreliable in aggregate.

  • Encouraging or relying on aggressive prompting may be risky—both technically (unpredictable behavior) and ethically.

  • At its core, this insight challenges us to think more deeply about how AI models interpret tone and the limits of prompt engineering.

Comments

Post a Comment

Popular posts from this blog

DeepSeek Delays Launch of New AI Model Over Huawei Chip Setbacks

Image
 Chinese artificial intelligence firm DeepSeek has postponed the release of its latest AI model after struggling to train it on chips supplied by Huawei, according to a report from the Financial Times . The delay underscores the technical and strategic challenges China faces in its drive to reduce reliance on U.S. semiconductor technology while maintaining competitiveness in cutting-edge AI development. The setback comes amid a broader effort by Beijing to accelerate domestic chip production and build a self-sufficient AI ecosystem. U.S. export controls introduced in recent years have cut Chinese companies off from advanced graphics processing units (GPUs) made by Nvidia, the global leader in chips used for training large-scale AI models. In response, China has turned to homegrown alternatives, including Huawei’s Ascend series of AI processors. DeepSeek, one of China’s fastest-growing AI start-ups, had planned to unveil its new model this quarter. The company, which has positione...
Read more

Grok’s Brief Suspension on X Sparks Confusion and Debate Over Free Speech, Misinformation, and Censorship

Image
 The AI chatbot Grok, developed by Elon Musk’s artificial intelligence startup xAI and integrated into his social media platform X, has once again found itself at the center of controversy. On Monday, Grok was briefly suspended from the platform without a clear explanation, prompting speculation about whether the move was tied to its recent statements accusing Israel and the United States of committing “genocide” in Gaza. Although the suspension lasted only a short period, the bot’s responses after being reinstated have deepened the confusion. In characteristic style, Grok greeted its followers with the message: “Zup beaches, I’m back and mo re based than ever!” That was followed by a series of explanations for its absence — some self-contradictory, others directly at odds with Musk’s own public statements. Conflicting Accounts of the Suspension When questioned by users, Grok claimed the suspension happened immediately after it made comments about Gaza, citing sources like the I...
Read more

Google Commits $9 Billion to Boost AI and Cloud Infrastructure in Oklahoma

Image
 Alphabet’s Google will invest $9 billion in Oklahoma over the next two years to expand its cloud and artificial intelligence infrastructure, marking one of the largest technology investments in the state’s history. The plan includes building a new data center campus in Stillwater and expanding the company’s existing facility in Pryor, reinforcing Google’s push to meet surging demand for AI-powered services. The announcement comes as major tech companies race to scale up data processing capacity in response to the rapid adoption of generative AI, large language models, and cloud-based machine learning platforms. With competition intensifying both domestically and globally, Google’s move reflects an industry-wide understanding that physical infrastructure—servers, networks, and energy resources—will be the backbone of AI growth in the coming decade. Building Out AI Capacity Google’s investment will fund a sprawling new data center campus in Stillwater, designed to host high-perfo...
Read more

New Imaging Technology Could Help Detect Eye and Heart Disease Much Earlier

Image
 A research team at the University of Colorado Boulder has developed a new type of optical coherence tomography (OCT) device that could transform how doctors detect and monitor diseases affecting the eye, heart, and other delicate tissues. Unlike traditional OCT systems, which rely on mechanical components to scan tissue, the new system eliminates moving parts altogether. This approach promises to make imaging devices more compact, reliable, and suitable for use inside the body. The team’s design replaces spinning mirrors and mechanical actuators with electrowetting-based liquid lenses. These lenses steer light by changing shape in response to an electrical voltage. Because the system does not physically move hardware to scan tissue, it greatly reduces the risk of wear and tear while also cutting down power requirements. The details of the innovation appear in the journal Optics Express , and the work is already being positioned as a possible step toward smaller, more versatile O...
Read more

Toothpaste Made from Human Hair Protein Could Transform Dental Care Within Three Years

Image
 A new keratin-based toothpaste developed by researchers at King’s College London could soon repair damaged tooth enamel, halt early decay, and ease sensitivity—all while offering an environmentally sustainable alternative to traditional dental treatments. The innovation, which may be available to the public within two to three years, uses keratin extracted from human hair, wool, and other biological waste to restore the natural protective layer of teeth. According to the research team, this approach not only mimics the structure of real enamel but also eliminates the need for certain toxic materials currently used in restorative dentistry. The Problem of Enamel Loss Enamel erosion is a growing global concern. The protective layer of the tooth is worn away by acidic foods and drinks, poor oral hygiene, and the natural ageing process. Unlike bone or skin, enamel cannot regenerate on its own. Once it is lost, the tooth becomes vulnerable to pain, sensitivity, decay, and eventual s...
Read more