5 AI Weaknesses That Can Burn Your Business

No technology is perfect. They all have limitations.

And when you think of imperfect technology, one of the first things that probably pops into your mind is artificial intelligence.

If you’ve used AI at all, you’ve probably heard of AI hallucinations, which are when a chatbot gives you a false or misleading response. But making up answers isn’t the only weakness.

Let’s examine some of the more popular weaknesses of AI and what you should do to ensure you’re using it responsibly.

Source Reliability

While public chatbots are improving at citing sources, they still sometimes depend on unreliable sources or don’t provide sufficient context for their answers.

For example, it might accurately cite statistics in a news report, but it might miss the year the data is from or other contextual caveats.

It’s quite common for chatbots to even forget what year it is, as it cites outdated sources.

Always double-check where AI is sourcing its answers, and remind it to cite up-to-date, relevant sources if you see too many outdated sources. Make sure they are trustworthy sources like research papers, government websites, academic institutions and established news outlets.

Biases

Humans are inherently biased, and the AI they develop can reflect that. If the data used to train AI is not diverse or completely representative, the answers will reflect those biases.

For example, a Stanford University study found AI has falsely flagged nonnative English writing as AI-generated, which could lead to accusations of cheating.

Additionally, scientists from MIT found biased language models think “flight attendant,” “secretary” and “physician’s assistant” are feminine jobs, while “fisherman,” “lawyer” and “judge” are masculine.

The companies that make each chatbot also prevent their chatbots from generating dangerous and unethical prompts. However, these intended biases can sometimes produce unintended ones.

Not only should you recognize bias from AI answers, but you should also recognize bias in your prompts that could lead AI to answer in a specific way.

To illustrate bias in prompts, here’s an example of a loaded assumption baked into a question.

Biased prompt: “Why is remote work less productive than office work?”

Neutral version: “How does productivity compare between remote and office work?”

The first prompt assumes the conclusion and forces the answer to justify it, while the second asks for the pros and cons of the two.

AI Sensitive to Prompt Quality

Ambiguous prompts will get you nowhere fast with any AI model.

You should provide overly specific prompts to get better results. For example, rather than “write an email to a client about new policies,” try using “write an email to [Client A] about the new bank information policy. The email should note that any request to change bank information should be verified over the phone to avoid phishing attacks. The tone should be approachable but professional.”

Follow the GCSE framework for prompting AI models, which stands for goal, context, source and expectations. These are the four pieces to add to any prompt to better your chances of getting a good response.

AI Overconfidence

Humans can sometimes be overconfident in their abilities, whether it’s in feats of athleticism or knowledge. But a study in the Memory & Cognition journal found that various AI models were more overconfident than humans, even when they performed poorly on a given task.

The study found humans typically adjusted their expectations when asked afterward how they performed a task. AI, on the other hand, doubled down, becoming even more confident regardless of whether they actually did well.

The key is to be skeptical of AI answers. As humans, we are conditioned to pick up on confidence cues from others, such as upright posture or firm eye contact. But AI doesn’t have the same cues. Even when proven wrong, it can sometimes be reluctant to admit mistakes.

AI Struggles with Math

AI isn’t great at math. But it’s also advancing so rapidly that this could be untrue in just a couple of years.

In 2024, a nonprofit research organization released FrontierMath, a tool designed to measure AI’s mathematical reasoning capabilities. It categorized math problems into Tiers 1-4, with 1 being advanced undergraduate problems and 4 being postdoctoral level.

In 2024, AI models couldn’t solve more than 2% of the problems. But in 2026, the best public AI models solved more than 40% of Tier 1-3 problems and more than 30% of Tier 4 problems.

AI-provided math answers should be double-checked with the same level of scrutiny as written answers.

Can AI Help Your Organization?

AI can save your team hundreds of hours per week on busy work, but it isn’t an all-knowing replacement for good old-fashioned research.

Contact us to schedule a consultation, and we’ll discuss how AI can help your employees be more productive while avoiding the pitfalls of misinformation.

Stay updated! Get tips and insights delivered to your inbox weekly by subscribing to our newsletter.