WSO2
CONTACT
Blog
 

Why Does AI Have to Be Held to a Higher Bar Than Humans (and Why It Matters)?

AI is already surpassing humans in many fields. Consider these examples:

If AI is so capable, why haven’t we delegated everything to AI? 

The answer is a paradox. Because AI is faster and more scalable than humans, we must hold it to a much higher standard. When deployed autonomously, AI behaves differently from people in three critical ways:

  1. Speed and Invisibility of Errors - We want AI because it is fast, scalable, and less expensive. However, these same traits make errors dangerous.
    • Human mistakes happen slowly. Checks and balances (e.g., peers or supervisors) often catch them before they cause catastrophic damage.
    • AI mistakes happen instantly and often invisibly. When an AI takes over, the decision-making process disappears from our awareness. We often discover the error only after the damage is done. Example: Two Amazon pricing algorithms once got into a bidding war, automatically driving the price of a biology textbook up to $23 million before a human noticed.
  2. Systemic vs. Distributed Bias - Human bias is distributed. If one hiring manager is biased against you, you can still apply to other companies and find a job. However, AI bias could become systemic as successful algorithms get replicated across entire industries. If a flawed model becomes the standard, that bias is deployed everywhere simultaneously. If one AI rejects you, everyone rejects you.
  3. Paradox of Unpredictable Predictability - AI models are very good at predictions, but paradoxically, they show unpredictable failure behaviour. Let's assume that a medical use case has 10 hardness levels. If its failures are predictable when the model is 90% accurate, all its mistakes should be at levels 9-10, and it does not make 1-3-level mistakes. This property is called monotonicity. Most AI models exhibit weak monotonicity and, thus, are much more unpredictable than humans. The paper by Ganguli et al., "Predictability and Surprise in Large Generative Models," is a good discussion about the topic. Unpredictability increases the risk associated with models and makes them hard to explain. 

Because of the three problems above, the risks posed by AI are significantly higher than the risks posed by humans doing the same task. What does this mean for us? The implication is that we need both humans and AI to work together. 

First, AI is already handling low-risk use cases like generating an image.

Second, what about higher-risk use cases? The good news is we do not have to replace humans with AI. When risks are high, asking "when can AI replace humans?" is the wrong question. Instead, we should be asking "how can humans work with AI?"

Actually, AI is already working with us by being part of the workflow. As I write the article, AI helps me in many ways, from grammar suggestions to QA. Human-AI collaborations often take the form of Suggestions for Improvement, Warnings, Informative Feedback, Abundant Information, and Predictive Displays. 

Collaborations are working in more risky domains, too. For example, "Integrating large language models with multimodal virtual reality interfaces to support collaborative human–robot construction work," and "Human-AI collaboration in large language model-assisted brain MRI differential diagnosis: a usability study," discuss two case studies where human and AI collaborations yielded better decisions in complex decision-making. 

An excellent analogy for how to think about human-AI collaboration comes from "From artificial intelligence to hybrid intelligence," by Catholijn Jonker, who compares it to a horse and rider. While the rider has overall control, the horse makes many split-second decisions, and their collaboration uncovers many feats that are not possible for either alone. 

Instead of thinking about AI vs. humans, we need to think about AI and humans. A key design parameter in this process is autonomy. It is a trade-off among costs, benefits, and risks involving ethical, political, legal, and technical considerations. Answers are likely to differ from one use case to another and to evolve as AI models improve. Some of those arguments are now happening in the context of self-driving cars.

However, leaving the responsibility in the hands of humans is not always the best choice either. For example, asking a human to approve an AI decision without giving them enough context and details can protect the integrity of the technological system at the expense of the nearest human operator. In “Who is Held Accountable When Agents Fail,” Rania Khalaf discusses avoiding such "moral crumple zones" by building the necessary platform foundations that can make AI safer. We need "Appropriate Reliance". I do not know the answer, and I believe we need to find the answers on a case by case basis. 

In conclusion, we need to hold AI to a higher bar than humans because AI behaves differently. However, that shouldn’t deter us. Realizing value with AI is not about asking whether to do a task with an AI system or a human - it is about determining the right level of autonomy to provide. Two key questions can help make that determination: a) how accurate is the AI component of the system, and b) what is the risk if it fails?

The better the AI performs and the less risky the application, the higher the autonomy can be. AI and agent platforms are one way to help determine how well the AI components are doing and provide guardrails, fallbacks, and human intervention along the way. In parallel, the models within these systems continue to become more reliable. This challenging and exciting domain requires a sharp focus on essential platform pillars: from establishing monitoring and evaluations to measure quality, to applying LLM guardrails for secure model interactions, and implementing agent identity for tool security and full lifecycle agent management.