AI in small business: Pros, Cons, and what to expect

As brands increasingly adopt AI to increase efficiency, nearly three quarters of their customers are concerned about unethical use of the technology.

Skepticism often accompanies the introduction of new technologies. The term "AI" may evoke images of futuristic scenarios where machines replace human workers, leading to job loss and more economic exploitation.

The acronym "AI" for "artificial intelligence" is itself misleading, as computers are not conscious and do not have intelligence. A marketing department decided to describe this software as "AI."

AI can handle repetitive or tedious tasks like crunching numbers and sifting through data. It can also be a simple sounding board for ideas.

But, ethical concerns and excessive energy use have cast a shadow. The key lies in well-planned business practices.

Let's balance optimism and caution. AI can be used responsibly, but sometimes, all you really need is Google search.

Pros:

• Blazing fast
  AI software's standout feature is its ability to quickly generate replies. This allows for rapid revisions. Whether it's drafting an outline or checking your prose, AI can perform complex tasks in milliseconds.
• Almost an English teacher
  AI maintains impeccable grammar and spelling. It can also adjust length, format, and style based on your preferences, ensuring that content aligns with specifications.
• Over and over and...
  To put it mildly, most jobs have one or two tasks that are dull or mechanical. Whether it is entering data, validating information, transcribing audio, etc. AI can reduce the boredom of repetition.
• Energizer bunny
  Unbound by time constraints, AI operates 24/7. Its tireless nature allows users to generate content, interact with other people, or conduct text-related tasks at any hour.
• Free (for now)
  The current iterations of DeepSeek, ChatGPT, Google Gemini, and others offer free and paid versions. This offers an economical solution for users looking to leverage AI without additional expenses.
• Robot voice
  AI-powered chatbots can provide instant responses to routine queries on a low-budget business website.
• Not an oracle but similar
  AI tools can analyze customer needs, market trends, and more (although these categories have higher risk). The tools are also less likely than people to indulge in pet theories, bad habits, etc. Be sure to measure your results in these categories.

Cons:

• AI not thinking at all?
  In August 2025, Apple released a research paper with a pragmatic look at the limitations of current AI systems. Link to Apple or download PDF "Illusion of Thinking"
• Deception
  AI vendor marketing departments feed the idea that the software has some flavor of consciousness under the hood. But there is also human nature in this. People like the idea that they've got a little buddy who works cheap. The "chat" interface directly appeals to this myth.
• AI vendors not trustworthy
  In early 2025, a Chinese company unveiled DeepSeek, an AI system developed at a fraction of the cost of comparable U.S. products. This breakthrough highlighted inefficiencies and high costs in U.S. AI development, sending shockwaves through the global tech and financial sectors.
• Mundane results
  “AI is shaping up to be the single largest misallocation of capital and energy in human history. Trillions poured into data centers burning enough energy to power nations. Draining enough water to empty rivers. For what? A slightly better cold email.”
  —Andre Gonsalves
• Conventional thinking
  AI's reliance on massive datasets means that generated content may reflect conventional thinking. So, it mirrors the average quality of training sources. Some savvy users therefore consider the AI gold rush to be a race toward mediocrity. Also, datasets often include bigoted ideas.
• Knows nothing
  Do not treat AI like it "knows" anything. AI is not a fact machine. (-Aimée Hall @astronaimee.bsky.social on Bluesky)
• Parroting, not thinking
  Many people have fallen victim to the illusion that AI software is sentient or intelligent. To the contrary, a quote about Large Language Models by computer scientists Timnit Gebru and Margaret Mitchell: "We called such systems 'stochastic parrots' — they stitch together and parrot back language based on what they've seen before, without connection to underlying meaning."
• Hallucination
  When AI misinterprets requests, it may generate confidently-delivered responses that are factually inaccurate. Known as "hallucination", this poses a risk when precision is crucial. In one notorious 2023 case, a lawyer used ChatGPT to prepare a filing, but the artificial intelligence bot delivered fake cases as precedent. This prompted the judge to weigh sanctions.
• Mass plagiarism
  Most AI developers have trained their software by "scraping" text from countless websites, without asking for permission. Critics describe this as mass plagiarism. In September 2025, the Associated Press reported that AI software vendor Anthropic agreed to pay authors $3,000 per book in a landmark settlement over pirated chatbot training material.
• More cost
  While basic versions of AI are free, costs may rise if businesses hire consultants. More expenses could make implementation financially burdensome.
• Energy consumption
  The AI carbon footprint is ominous. Training and running AI models is a heavy drain on electricity, often generated with fossil fuels. A 2019 study by the University of Massachusetts Amherst found that training one AI model can emit more than 626,000 pounds of CO2. That's equivalent to the emissions of five cars over their lifetimes. As the tech bros continue to roll out AI, it has become obvious they intend to impose the costs of data centers on ordinary people.
  
  • Analysis of AI's carbon footprint:
  • Gemini has the highest training footprint but is carbon-neutral.
  • ChatGPT balances high training costs with renewable-powered inference.
  • DeepSeek likely has a higher carbon intensity per kWh due to reliance on coal-heavy grids.
• AI companies badly run
  “OpenAI's solution to copyright infringement is to just pay lump sums to media company C-suite execs in exchange for a publication's lifeblood—i.e. the writing produced by humans, none of whom will be compensated in these deals.” —Andrew Paul @andrewpaul.bsky.social

Both:

• Computer scientist Jaron Lanier has a message: “Artificial intelligence” does not exist. In 2023, he urged people to look past science-fiction notions of AI as an autonomous entity and instead see it as a form of social collaboration—tools trained on materials created by people.
  
  Lanier argues that reframing AI this way could lead to better solutions to problems such as job displacement and encourage a more inclusive vision of the future. It also changes how we think about what AI actually produces.
  
  “It doesn’t matter technically,” Lanier said. But “if you analyze it on those terms, you have a much more actionable way of understanding it, and if you integrate it into society on those terms you have a brighter and more actionable set of paths open for the future of society.”
  
  Lanier acknowledges that AI’s ability to aggregate and sift through vast bodies of human knowledge could lead to new understanding and even new scientific theories. But those potential insights, he emphasizes, stand on the shoulders of real people who created the knowledge in the first place.

How does it work?

I could say to you that AI's pattern-matching "simulates" human intelligence. The word “simulate” is technically correct in some contexts, but it can also mislead people.

1. What “simulate” means in general

To simulate means to imitate the behavior or appearance of something without being the thing itself.

• A flight simulator imitates the experience of flying, but it doesn’t actually fly.
• Similarly, an AI language model imitates human reasoning or conversation, but it doesn’t think in the human sense.

So, if you say:

“AI simulates human intelligence,”
that’s accurate in the sense that AI produces outputs that look intelligent, based on pattern recognition.

2. Why it can be confusing

For non-technical listeners, “simulate” can imply that AI has a model of human thought inside it—as if it’s pretending to think like a person by copying our internal processes.

That’s not really true for current AI systems.

Modern AI (like ChatGPT or image generators) doesn’t understand or reason like humans. It just detects patterns in enormous datasets and generates new patterns that fit what it’s seen.

So it might be clearer to say something like:

“AI models human-like behavior through pattern-matching.”
or
“AI mimics the outputs of human intelligence, but not the underlying thought process.”

Avoid AI-fabricated replies

AI "hallucinations" are a significant concern, particularly when the software generates incorrect or nonexistent information.

Hallucinations are more likely to occur in the following situations:

Requests for specific, niche information: When an AI lacks access to authoritative sources or sufficient training data on a highly specific topic, it may "guess" or fabricate plausible-sounding responses to fill the gap. For example, it might generate citations or URLs that look valid but are entirely fictional.

Overly complex or ambiguous prompts: If the prompt is long, complicated, or unclear, the AI might struggle to discern the exact requirement and default to creating an answer that seems appropriate but is inaccurate.

Fields requiring specialized knowledge: Legal, medical, and technical fields are particularly prone to hallucinations because AI may mimic the structure and style of real information without accurately reflecting actual facts or regulations.

Lack of verification in the system's design: Some AI tools are designed to prioritize fluency and coherence over factual accuracy, leading to convincing but false replies.

Advice for mitigating the risk

To help small businesses and individuals avoid falling prey to these false outputs, you can offer the following tips:

Verify all information: Treat AI-generated outputs as a starting point rather than authoritative answers. Cross-check facts, URLs, or citations against reliable, external sources before using them.

Be specific in prompts, but not overly narrow: While specificity helps guide the AI, overly niche requests can lead to hallucinations. Frame prompts to seek general guidance or sources, then follow up with your research.

Use AI in low-stakes contexts first: Start by using AI for tasks where inaccuracies are unlikely to cause harm, such as brainstorming, drafting, or summarizing information.

Prefer reputable, specialized tools for critical tasks: For legal, financial, or healthcare-related queries, use AI systems specifically trained and certified in those areas rather than general-purpose AI.

Look for built-in citation checking: Some AI systems offer features to verify citations or link to real-time resources. Use these whenever possible.

Responsible and humane AI use:

As you adopt AI, it's crucial to do so responsibly. Ethical use of AI ensures that it benefits both your business and society. Consider publishing your policy on AI ethics.

Guidelines for responsible AI use

—Transparency

Clearly communicate to your customers when AI is in use, especially in customer interactions. Transparency builds trust.

—Danger zones

Some of the highest-risk uses of AI include health care, legal matters, financial investing, and various types of forecasting. These areas often involve critical decision-making, where errors or bias can have significant consequences, such as misdiagnoses, flawed legal advice, financial losses, or inaccurate predictions that harm planning and operations. Organizations need to exercise caution and ensure rigorous testing in these domains.

—Bias mitigation

Regularly audit AI algorithms for biases. Ensure that your AI applications treat all individuals fairly and avoid perpetuating societal biases.

—Data privacy

Prioritize customer data privacy. Implement robust security measures to protect sensitive information and adhere to data protection regulations.

—Human oversight

While AI can automate processes, human oversight is essential. Avoid reliance on AI for critical decision-making, especially in areas that require a nuanced understanding of human emotions and context.

AI used in research

The distinctions between AI systems used by the public (like DeepSeek, Gemini, and ChatGPT) and those used in scientific research (e.g., in mathematics, genetics, or other specialized fields) can be understood in terms of their design, training, and application. Here are the key differences:

1. Purpose and Scope
Public AI: Designed for general-purpose use, catering to a wide range of tasks like answering questions, generating text, assisting with creative writing, or providing customer support.

Scientific AI: Tailored for specific, well-defined tasks, such as solving mathematical problems, analyzing genetic sequences, or simulating physical phenomena. Their scope is narrower, focusing on precision and accuracy within a specialized domain.

2. Training Data
Public AI: Trained on vast, diverse datasets from the internet, including books, articles, and websites.

Scientific AI: Trained on highly curated, domain-specific datasets, such as scientific papers, experimental data, or mathematical models. The data is often peer-reviewed and rigorously validated, reducing the likelihood of errors or hallucinations.

3. Output Requirements
Public AI: Prioritizes generating coherent, contextually relevant responses, even if the information is not perfectly accurate.

Scientific AI: Emphasizes precision, reproducibility, and adherence to established facts or theories. Errors or inaccuracies are unacceptable, as they could compromise the validity of research findings.

4. Evaluation and Validation
Public AI: Evaluated based on user satisfaction, fluency, and the ability to handle diverse queries.

Scientific AI: Subject to rigorous testing and validation against ground truth data or established scientific principles. Results are often peer-reviewed or cross-checked with experimental evidence to ensure reliability.

5. User Interaction
Public AI: Interacts with a broad audience, including non-experts, and must accommodate a wide range of input styles and intents. This can lead to misunderstandings or ambiguous queries that increase the risk of hallucinations (errors).

Scientific AI: Typically used by domain experts who provide precise, well-structured inputs. The interaction is more controlled, reducing the likelihood of misinterpretation or errors.

6. Error Tolerance
Public AI: Some level of error is tolerated, as long as the overall user experience remains positive. Hallucinations are seen as a trade-off.

Scientific AI: Errors are unacceptable, as they could lead to incorrect conclusions or flawed research. The systems are designed to minimize uncertainty and ensure high accuracy.

7. Transparency and Explainability
Public AI: Often operates as a "black box," with limited transparency into how specific outputs are generated.

Scientific AI: Prioritizes explainability, as researchers need to understand how the AI arrived at a particular result. This transparency helps ensure the reliability of the findings.

8. Regulatory and Ethical Standards
Public AI: Subject to general ethical guidelines and regulations, but these are often less stringent compared to scientific applications.

Scientific AI: Must adhere to strict ethical and regulatory standards, particularly in fields like genetics or medicine, where errors could have serious consequences.

Summary

The key distinctions lie in the purpose, training data, output requirements, and error tolerance of the AI systems. Public AI prioritizes versatility and user engagement, which can lead to hallucinations, while scientific AI focuses on precision, accuracy, and reliability within a narrow domain. Both types of AI have their strengths and limitations.

Ultimate guardrail

Simple rule of thumb: When on the page of an AI chat, the feeling of interacting with what seems like a friendly and effecient intern is a fantasy concocted by a marketing department to acquire your money. No matter how much value you derive from AI chat, please remember this.

Links:

• Timnit Gebru on AI hype, ethics & algorithmic racial bias
• AI could go the way of the dirigible airship
• Attempting success wih AI
• Kevin Powell on making content (and AI)
• Mat Marquis on LLemdashes
• Google breaking the social contract
• AI Killed My Job
• Nerd Wallet: Small business 5 ways
• LinkedIn: Small business marketing