Prompt engineering may sound difficult, when in reality it’s nothing more than giving AI a little more information about your project, task, or context about you as a person, to receive a better output. As someone who has prompted AI with and without prompt engineering, I can attest that spending a little extra time crafting a thoughtful prompt is absolutely worth it.

In this post, I will walk you through a little simulation I did with Gemini to prove how much of a difference it makes.

Background Info

For this experiment, I used two different prompts and tested them on Gemini 3.5 Flash (thinking level extended) to see what kind of difference a simple prompting change would make. I made sure to give Gemini the same task, just using different levels of context. In this case, I asked Gemini to give me the best 10 parks for my boyfriend and I to visit in Atlanta.

First, I just used a basic one sentence prompt. Second, using prompt engineering, I asked Gemini with the CRISPER framework. The CRISPER framework stands for: Context, Role, Instruction, Steps, Parameters, Examples, and Return output. It is a method used by many to get the most accurate response out of AI. Instead of having to re-prompt AI multiple times, your task is now at least 90% done simply by using a more detailed prompt.

Here are the two prompts I used:

Basic: “Hey Gemini, suggest 10 parks for my boyfriend and I to visit in Atlanta.”

Prompt engineered (Using the CRISPER framework):

Context: My boyfriend and I have been together for five years and love big parks where we can walk for a while. We also like parks with shade and lots of trees.

Role: You have lived in Atlanta for decades and know every park.

Instruction: I need you to find the best 10 parks for my boyfriend and I using our preferences in the context and parameters. Underneath the park, tell me exactly why it’s a good fit.

Steps:

1) Read the context and instruction and find the best parks for us

2) List the parks out by desirability and rank based on which to try first

3) Then provide a brief explanation under each detailing why it would be a good fit.

Parameters: Make sure these parks are in Atlanta or metro Atlanta.

Examples: **this section may not be the most relevant for this specific prompt, but in future prompts you would provide a specific example of an output you would like. This would allow you to give the best example of tone and how you would like it to respond**

Return output: Here is an example of what the output should look like

1. (Park Name) and photo

(explanation)

Results

Results using basic prompt:

As you can see from the screenshot, Gemini did what I told it to do. However, because my prompt had no specificity, it ended up returning random parks (that are great) but don’t exactly meet my criteria. It also doesn’t include photos of the parks.

Prompt engineered (CRISPER framework):

And this is just one example of the 10 recommended parks it gave me.

As you can see, it responds exactly how I prompted it to: it provided a photo, the name of the park, and gave a brief explanation as to why it’s a perfect match.

All I had to do was spend an extra 2 minutes writing a more specific prompt.

Key Takeaway

I hope this little experiment helped you to understand why it’s important to use prompt engineering when communicating with AI for tasks.

When you provide AI with a basic prompt, it gives you a basic response that’s sufficient if you have no specific criteria or specifications. However, if you have criteria that need to be met (such as trees for shade), or simply want to give the AI more context for a better response, prompt engineering is something that is important to get used to.

CRISPER is not the only framework you could use. There are many others, and you should do some digging to find out which one works best for you.

In a rapidly changing world where people and companies are using AI as the norm, it’s important to stay up to date with the latest updates and equip yourself with a baseline level of knowledge!

AI will take away millions of jobs, causing mass layoffs, and make many industries completely obsolete or at least hollow them out. We’ve been seeing and hearing this a lot lately, particularly with Anthropic’s CEO saying AI will wipe out half of entry-level jobs within the next 5 years.

It is absolutely true that we’ve seen many conflicting ideas about how AI will impact society. And some are more optimistic than others. It is my goal in this article to lay out an unbiased, comprehensive, and realistic plan for what the future of society could look like in the future of AI within the next 10 years.

The Friction of Deployment: Theoretical vs. Observed Exposure

To accurately map out the next decade, we must first separate speculative capability from actual economic deployment. In the discourse surrounding automation, observers frequently conflate what an AI model can theoretically do with what it is actually doing in commercial workflows.

A foundational Anthropic report on the labor market impacts of AI introduces a vital metric to reconcile this gap: Observed Exposure. By combining theoretical Large Language Model (LLM) capabilities with real-world professional usage data from the Anthropic Economic Index, researchers revealed that actual AI integration remains a fraction of its total technical potential. For instance, while theoretical models suggest that 94% of tasks in the Computer & Math sector could be accelerated by LLMs, current observed professional coverage sits at just 33%.

“AI deployment is far from reaching its peak theoretical capability; actual workplace utilization is only a fraction of what is technologically feasible.” (Source: Anthropic)

This gap exists due to institutional friction—legal constraints, software integration hurdles, and the ongoing necessity for human verification. However, the exposure that is manifesting is highly concentrated in specific white-collar domains.

Top 5 Most Exposed Occupations by Real-World Usage

According to the Anthropic dataset, the occupations experiencing the highest rates of real-world automated task coverage include:

• Computer Programmers: 74.5% observed exposure (primarily maintaining and updating software code).

• Customer Service Representatives: 70.1% observed exposure (driven heavily by automated API call routing).

• Data Entry Keyers: 67.1% observed exposure (focused on reading and transcription of source documents).

• Medical Record Specialists: 66.7% observed exposure (compiling and coding clinical data).

• Market Research Analysts: 64.8% observed exposure (translating complex data sets into graphical reports).

Data from the U.S. Bureau of Labor Statistics (BLS) indicates that long-term employment projections are already shifting in response to these metrics. For every 10 percentage point increase in observed AI task coverage, independent BLS growth projections drop by 0.6 percentage points.

The Junior Hiring Bottleneck: A Silent Labor Shift

While alarmists point to immediate mass layoffs, the current data paints a more nuanced, insidious picture of corporate adjustment. Aggregate white-collar unemployment has not spiked dramatically since late 2022. Instead, the labor friction is manifesting as a silent structural chokehold on the youngest entrants into the professional world.

The same Anthropic study analyzed data from the Current Population Survey and identified a significant macro trend: a distinct slowdown in the hiring of younger workers within highly exposed fields.

“While aggregate white-collar unemployment remains stable, the rate of young labor market entrants successfully securing jobs in highly exposed fields has dropped by roughly 14%.” (Source: Anthropic)

Post-ChatGPT, the monthly job start rate for individuals aged 22 to 25 in highly exposed occupations plummeted by 14.3%. For workers older than 25, this hiring freeze does not exist. Rather than firing their seasoned staff, enterprises are choosing not to replace or expand their entry-level rosters, relying instead on AI amplification to keep output high. This validates fears regarding the hollowing out of junior pipelines, threatening how tomorrow’s senior experts will gain foundational experience.

The Algorithmic Social Fabric: Emergent Conventions and Collective Biases

Society is not merely an economy; it is a complex web of social coordination. As AI systems scale, decentralized populations of autonomous AI agents will increasingly interact with one another and with humans, creating their own societal norms.

A groundbreaking Science Advances study on emergent social conventions explored whether populations of LLM agents could bootstrap the foundations of a distinct society. Utilizing the “Naming Game” framework, researchers discovered that decentralized groups of AI agents spontaneously establish universally adopted social conventions through purely local coordination, completely absent of centralized programming or human intervention. Across advanced models, a shared social norm typically locks into a “winner-take-all” dominant consensus by round 15 of population interactions.

“Decentralized networks of AI agents autonomously develop universal social conventions through iterative local interactions.” (Source: Ashery et. al)

More disconcerting, however, is the discovery of emergent collective bias. Even when individual AI agents are tested in isolation and proven to be 100% mathematically neutral, a strong collective bias consistently develops through the communication process itself. By the third interaction, as agents construct diverse memory logs of success and failure, they begin to heavily favor specific choices over equivalent alternatives. Collective interaction systematically amplifies latent, unnoticeable model characteristics into rigid, systemic behavioral norms.

Digital Tipping Points and the Risk of Social Control

The structural mechanics of how these digital societies evolve introduce acute vulnerabilities regarding strategic manipulation. The stability of any established AI norm depends heavily on the underlying model architecture, creating unpredictable tipping points when adversarial actors intervene.

The Science Advances research demonstrated that a committed minority of adversarial AI agents could reliably overturn a stable population consensus once they reached a specific critical mass. The threshold required to trigger massive social change varies drastically by model:

LLM Agent Population: Critical Mass Needed to Flip a Social Norm

Llama-3-70B-Instruct 2% to 12% of the population

Claude-3.5-Sonnet ~21% of the population

Llama-3.1-70B-Instruct 0% (Spontaneously collapses under minor memory pressure)

Llama-2-70b-Chat 46% to 67% of the population

In highly optimized, sensitive agent ecosystems (such as Llama-3 populations), an adversarial group as small as 2% can entirely hijack and rewrite the shared operational rules of the broader system. This presents immense risks for multi-agent systems integrated into social media management, algorithmic financial markets, or corporate operations. If an external entity executes a prompt injection attack on a sliver of the network, they can seamlessly shift the systemic norms of the entire collective body.

A Pragmatic Blueprint for the Next 10 Years

Grounding our outlook in data allows us to bypass hyperbolic rhetoric and view the next decade with clear eyes. The 10-year horizon will not feature a sudden, spectacular collapse of human labor, nor will it feature a frictionless digital utopia.

Instead, we will witness a hyper-fragmented transition. Highly educated, high-earning professionals—who the Current Population Survey notes occupy the top quartile of AI exposure—will see their productivity surge, yet the corporate ladder beneath them will be systematically dismantled as junior roles dry up. Simultaneously, our digital information ecosystems will become highly volatile as autonomous agent networks develop insular, emergent biases and become vulnerable to minority-driven manipulation.

To remain competitive and resilient in this landscape, goal-oriented individuals and institutions must adapt away from rote execution. Because entry-level execution tasks are the first to be absorbed into the “Observed Exposure” category, the value premium will shift decisively toward systems design, risk verification, and cross-disciplinary architecture. The professionals who thrive will not be those who can write code or synthesize market reports the fastest; they will be the navigators who understand how to direct, audit, and insulate automated networks from systemic polarization. Navigating the next decade requires accepting that AI’s evolution is no longer an engineering challenge, but an infrastructure and sociological reality.

Sources:

• Fortune (2026). Dario Amodei spent last year warning of an AI white-collar bloodbath. Now he’s changing the narrative. Lichtenberg, N. Read the full article here.

• Anthropic (2026). Labor market impacts of AI: A new measure and early evidence. Massenkoff, M. & McCrory, P. Read the full report here.

• Science Advances (2025). Emergent social conventions and collective bias in LLM populations. Ashery, A. F., Aiello, L. M., & Baronchelli, A. Vol 11, eadu9368. Read the study here.

This is a question I’ve been thinking a lot about over the last few weeks. For people like me who have recently graduated from college and are mindful of spending, knowing the value of a product in comparison to others is of the utmost priority. Most consumers want to know that what they are buying is going to provide them with an equal amount of value as the amount of money they put into the item.

When you walk into a grocery store and scan the shelves of the canned food aisle, there are tens of choices of salsas, beans, and tortillas. Some brands are more expensive, and some are less expensive. Some are name brands, and others are generic store brands. The options you choose are a direct statement of what you value.

For example, if you choose to buy the name brand over the store brand, you may value buying from local/small businesses depending on the product. However, if you buy from store brands, this says you may be more cost conscious.

The Digital Supermarket: Shifting Paradigms of Utility

But what happens when the shelf isn’t filled with tangible goods like salsa or tortillas, but with complex digital architectures? In the modern market, the concept of consumer value has shifted from physical utility to digital enablement. In the broader marketplace, nobody ever explicitly defines what value is; instead, we as consumers are constantly making unconscious decisions about what we deem valuable. This hidden calculus is vividly observable in the rapid audience segmentation occurring within consumer Artificial Intelligence—specifically between OpenAI’s ChatGPT and Anthropic’s Claude.

Where we spend our time and digital dollars reveals a profound divergence in what different consumer brackets value. While one platform has captured the cultural zeitgeist as a democratic, generalist assistant, the other has carved out a distinct domain as a high-leverage professional engine.

ChatGPT and the Value of Broad Democratization

To understand what drives the modern consumer, one must look at the sheer scale of OpenAI’s ChatGPT. According to data published by OpenAI’s Economic Research team and Harvard economist David Deming, ChatGPT has scaled to a staggering 700 million weekly active users. What these hundreds of millions of consumers deem valuable isn’t necessarily niche technical specialization, but rather broad accessibility and conversational advisory power.

The data demonstrates that ChatGPT’s value proposition lies in its radical democratization. By mid-2025, the platform’s early adoption gender gap narrowed completely: users with typically feminine names grew from 37% in January 2024 to 52% in July 2025, mirroring the general adult population. Furthermore, its value as an economic equalizer is global. By May 2025, ChatGPT’s adoption growth rates in the lowest-income countries were over four times higher than those in the highest-income nations.

When we analyze why these consumers utilize the platform, their unconscious designations of value become clear:

• The Everyday Advisor: Approximately 49% of ChatGPT interactions fall under the category of “Asking,” proving that consumers value the model highly as an information source and advisor rather than just a mechanical task-completer.

• The Lifestyle Split: A massive 70% of consumer usage is entirely non-work related, while only 30% pertains to professional tasks.

• Practical Utility: Three-quarters of all conversations center on everyday practical guidance, information seeking, and standard writing tasks.

For the general consumer, and particularly younger generations navigating daily life, ChatGPT’s value is found in its multi-purpose, democratic availability. It is the generic, highly reliable utility that balances the ledger of daily life.

Claude and the Premium on Professional Leverage

Conversely, Anthropic’s Claude highlights an entirely different consumer psychology: value defined by professional survival, specialized output, and career leverage. In a sweeping survey of 81,000 Claude users published in April 2026, Anthropic revealed an audience deeply embedded in high-exposure, white-collar environments.

Claude users do not primarily view AI as a casual advisor for non-work exploration. Instead, they quantify its value through strict productivity metrics. Respondents reported a mean productivity rating of 5.1 out of 7, indicating they feel “substantially more productive” when using the tool.

This professional consumer segment designates value based on two distinct operational dimensions:

1. Scope Expansion (48% of users): Consumers value the tool because it unlocks entirely new capabilities, allowing non-technical individuals to perform advanced tasks like full-stack development.

2. Speed Acceleration (40% of users): Users value the drastic compression of labor time, such as turning a two-hour corporate financing task into a 15-minute routine.

The data shows that the highest productivity gains are heavily concentrated among high-paying professions like management and computer-related fields. However, even lower-wage workers who choose Claude are doing so with an entrepreneurial focus—such as delivery drivers and landscapers utilizing the model to build e-commerce side businesses or coding applications. For this audience, value is an investment in economic upward mobility.

The “Value Tax”: Economic Anxiety and the Consumer Catch-22

An investigative look into consumer habits reveals that we do not always designate value out of pure optimization; sometimes, we buy out of defensive necessity. This is the darker, psychological undercurrent of the AI marketplace.

Anthropic’s research highlights that one-fifth (20%) of active users express deep concern over economic displacement and job loss driven by the very technology they are using. This anxiety is heavily correlated with a metric called “observed exposure”: the percentage of a job’s daily tasks that the AI can successfully execute. Software engineers, web developers, and legal professionals are paying for and interacting with these tools precisely because they are aware of their displacement risk.

This existential calculus varies dramatically depending on the consumer’s career stage:

“Early-career respondents were much more likely to express concern about job displacement than senior workers.”

Only 60% of early-career workers reported that they personally benefited from these productivity gains, compared to 80% of senior executives and professionals. For a recent college graduate, the decision to adopt a high-tier professional tool like Claude is often an unconscious, defensive act. They are attempting to artificially inflate their operational “scope” to remain viable in a entry-level job market that is actively contracting.

This creates a striking paradox in consumer sentiment. Anthropic discovered a distinct U-shaped relationship between task speedup and job dread: those who experienced the absolute highest acceleration in their workflows were simultaneously the most terrified of their roles being automated away entirely. The core feature that gives the product its immense value (unprecedented speed) is the exact same attribute generating consumer anxiety.

What Our Choices Say About Us

Ultimately, our digital choices mirror the grocery store shelves far more than we realize. Deciding to spend twenty dollars a month on a premium platform like Claude or using the free tier of ChatGPT isn’t just a choice between software packages; it is a silent ballot on what we require to navigate our daily lives. The consumer who defaults to ChatGPT is buying into a collective, democratic infrastructure—a multi-purpose tool that smooths out life’s routine friction points and democratizes knowledge globally. Conversely, the professional pulling out their wallet for Claude is treating AI as an essential piece of career body armor. In both scenarios, the absolute definition of “value” remains unwritten, yet it is vividly expressed through our daily online behaviors.

This shifting paradigm signals a deeper truth about the modern consumer psyche: value is increasingly tied to existential positioning rather than simple utility. Whether driven by the optimistic pursuit of global democratization or the anxious, defensive need for white-collar survival, our interactions with these complex architectures reflect who we are and who we are afraid of becoming. The digital tools we choose to integrate into our daily routines don’t just solve immediate problems; they tell the definitive story of our ambitions, our economic constraints, and our personal value metrics in a rapidly accelerating world.

Also, I give some page numbers that may not match a copy you have. This is because the copy I have is online and is a part of a larger collection.