7 Common GenAI Myths Leading to Missed Opportunities

Generative AI widespread adoption started roughly two years ago, yet significant misconceptions still prevent businesses from unlocking its true potential.

These myths fall into two main categories: where the technology can be used and how to build with it. This article focuses on myths about where GenAI can be deployed. Interestingly, these misconceptions span both ends of the spectrum—either underestimating or overestimating GenAI's capabilities. The consequences of underestimating and overestimating GenAI differ significantly, so it is best to address them separately

Underestimating What GenAI Can Do

Many businesses don't use GenAI for its most valuable applications. Why? Businesses often prioritize lower-value use cases because they incorrectly assume the technology isn't ready for more complex use cases that are often far more valuable.

• Myth 1: It’s Only Good for Chatbots and Content Generation
  Early applications shaped this view: generating images (DALL-E), text/chats (ChatGPT), and code (GitHub Copilot).
  Reality: They can do so much more 
  Today's frontier models (GPT-4, Claude 3, Gemini 1.5) are built not only for generating content but also for reasoning, automating complex workflows, and autonomously taking action.
  

• Myth 2: Hallucinations Are Still Rampant
  Early AI models often produced inaccurate outputs because they were trained to predict the most likely next token, thus producing plausible output, not factual output.
  Reality: They are rare in well-engineered AI systems
  While models have improved, significant advancements have occurred in the AI systems surrounding GenAI models. Retrieval-augmented generation (RAG), which combines external knowledge retrieval (ideally vetted for factual accuracy) with AI generation to ground responses in factual data, and guardrails have reduced this risk significantly. Pinecone had a neat experiment earlier this year demonstrating an increase in faithfulness (close proxy for accuracy or lack of hallucination) with RAG. In my experience, with a bit of optimization one can get faithfulness to be well above 99%. Lastly, multi-agent approaches, where one agent does the task and another reviews it, have also helped.

• Myth 3: You Need Tons of Data and Custom Training
  This belief comes from traditional machine learning, where massive datasets and custom models were needed for every specific task.
  Reality: For most tasks, off-the-shelf models are usually good enough
  Modern models are general-purpose and require far less data to deliver value. Custom data can still be integrated, but methods like RAG eliminate the need for massive re-training. In rare instances where fine-tuning is needed, as few as 100 data points can suffice, and model providers make this easy now for text and images.
  

• Myth 4: It’s Very Expensive to Run
  Early models were costly, as indicated by articles about how AI costs must be very carefully managed and high-profile VC firms warning about how AI might mess up software margins.
  Reality: Prices for frontier models have dropped by at least 85% in a year
  One could argue they have dropped by 99%  because GPT-4o-mini is on par with the original GPT-4. It's not yet "too cheap to meter", but cost should be less of a concern than it used to be.
  

  

Overestimating GenAI’s Capabilities

Overestimating GenAI can lead to unrealistic expectations and disappointment. Occasionally, these disappointments may result in organizations discounting GenAI altogether as immature.

• Myth 5: GenAI Is a Superset of Traditional AI/ML
  Many folks fall into the trap of seeing incredible tasks that Gen AI models can do and assume Gen AI can do all the work of Traditional AI and then some.
  Reality: GenAI can do some tasks of traditional AI, but the capability overlap is smaller than many people assume. The table below is a rough approximation of how to think where each is good and where they overlap.
  
  

• Myth 6: GenAI Is Great at Making Sense of Large Data
  All main frontier models support ingestion of lots of data (over 100K+ tokens or 100+ book pages). The misconception is that GenAI can reason over entire large datasets and automatically discern the most recent, relevant and truthful information.
  Reality: GenAI can’t automatically separate good from the bad
  GenAI can handle large data sets in terms of volume, but it lacks the ability to inherently assess the reliability or quality of data. Without proper data validation, conflicting or incomplete information can lead to misleading outputs. For example, if it has two documents stating different prices and those documents aren’t dated, how can an AI model possibly know which one is accurate? A human couldn’t do it accurately either.
  
  

• Myth 7: AI Tech Is All You Need
  There is a common belief that building the AI system is the only hard part, in part because of the magical way in which they can do things, especially autonomous AI agent systems.
  Reality: GenAI, like all tech, requires careful integration into the organization
  The technology requires careful integration on many layers. As discussed in an earlier post, it will often require process redesign. If it is expected to act as a co-pilot or partially autonomous agent, change management will be critical. If it can’t yet handle all the use cases, then some process decisions on how the jagged boundary will be handled are also required.
  

Conclusion:

Many of the myths surrounding Generative AI were true at one point or contain a seed of truth. However, to be an effective leader, it's essential to ensure that these myths are not leading organizations to suboptimal decisions.

The field of GenAI is evolving rapidly, and what holds true today may not be accurate in a year. It’s critical to stay up to date to avoid the trap of settling for what it could do yesterday and instead use it for what it could do tomorrow.