Enterprises have fallen into AI adoption, where they deploy chatbots, content generators, and decision support tools. According to McKinsey, 78% of companies use AI in at least one business function.
The madness of implementation is understandable – everyone considers potential value. But in this rush, many organizations overlook the fact that all neural network-based technologies, including all LLMs and generative AI systems used today, and for the foreseeable future, have a big flaw: they are unpredictable and ultimately uncontrollable.
As some people learn, there may be real impact. At a Chevrolet dealer that has deployed the chatbot to its website, customers convinced Chatgpt-Power’s bot to sell the $58,195 Chevy Tahoe for just $1. Another client prompted the same chatbot to write Python scripts for complex fluid dynamic equations, which was delighted. After these events spread, dealers quickly weakened the robot.
Last year, Air Canada, which argued that its chatbot’s chatbot lost its chatbot under bereavement discount, “is an independent legal entity responsible for its own actions.”
This unpredictability stems from the basic building of LLM. They are so large and complex that they cannot understand how they get a specific answer or predict what they produce until the output is produced. Most organizations deal with this reliability issue without fully recognizing it.
A common sense solution is to manually check AI results, which can work but can greatly limit the potential of the technology. When AI is downgraded to a personal assistant – drafting text, holding meeting minutes, summarizing documents and helping with coding – can lead to moderate productivity gains. Not enough to completely change the economy.
When we stop using it to assist existing work, the real benefits of AI will arrive, but reconnecting the entire process, system and company to use AI without human involvement at every step. Consider loan processing: If banks provide AI assistants to loan officials to summarize applications, they may work 20-30% faster. However, deploying AI to handle the entire decision-making process (with appropriate safeguards) could cut by more than 90% and eliminate almost all processing time. This is the difference between incremental improvement and conversion.
Ways to implement reliable AI
Utilizing the full potential of AI without succumbing to its unpredictability requires a refined fusion of technical approaches and strategic thinking. While several current approaches provide partial solutions, each approach has great limitations.
Some organizations try to mitigate reliability issues with system push – cleverly shifting AI behavior in the desired direction, so it responds to certain inputs in a specific way. Anthromorphic researchers demonstrated the fragility of this approach by identifying the “Golden Gate Bridge feature” in Claude’s neural network and artificially amplifying the method, which led to Claude’s development of an identity crisis. When asked about its body form, Claude claimed that it did not admit that it did not yes The Golden Gate Bridge itself. This experiment reveals how easy it is to change the core functionality of the model, and each fine-tuning represents an aspect of the trade-off that may improve performance while lowering others.
Another way is to let AI monitor other AIs. Although this layering approach may encounter some errors, it introduces additional complexity and still lacks comprehensive reliability. A hard-coded guardrail is a more direct intervention, such as preventing responses containing certain keywords or patterns, such as precursor components of a weapon. Despite effective response to known problems, these guardrails cannot anticipate the new and problematic outputs that emerge in these complex systems.
A more efficient approach is to build an AI-centric process that works automatically, with human supervision strategically positioning to capture reliability issues before causing real-world problems. You don’t want AI to directly approve or reject a loan application, but AI can conduct a preliminary assessment of human operators for review. This can work, but it relies on human alert to capture AI errors and disrupt the potential efficiency improvements in using AI.
Built for the future
These partial solutions point to a more comprehensive approach. Organizations that fundamentally rethink how their work is done instead of simply enhancing existing processes with AI assistance will gain the greatest advantage. But AI should never be the last step in a high-risk process or decision, so what is the best way to move forward?
First, AI builds a repeatable process that will provide consistent results reliably and transparently. Second, humans review the process to make sure they understand how it works and that the input is appropriate. Finally, the process runs autonomously (without AI) and the results are regularly reviewed.
Consider the insurance industry. Conventional approaches may add AI assistants to help claim processors work more efficiently. A more revolutionary approach would use AI to develop new tools—such as analyzing corrupted photos or enhancing computer vision for fraud detection models to identify suspicious patterns—and then combine these tools into an automated system managed by clear, understandable rules. Humans will design and monitor these systems rather than dealing with personal claims.
This approach maintains human supervision at the most critical moment: the design and verification of the system itself. It allows for increased exponential efficiency while eliminating the risk that AI unpredictability can lead to harmful results in various situations.
For example, AI may determine potential indicators of loan repayment capabilities in transaction data. Human experts can then evaluate the fairness of these indicators and establish clear understandable models to confirm their predictive capabilities.
This approach to interpreting AI will create a more obvious gap between the tissue that uses AI on the surface and the organization that transforms it around it. The latter will increasingly lead in its industry, able to deliver products and services at price points that its competitors cannot match.
Unlike Black-Box AI, interpretable AI systems ensure that humans maintain meaningful supervision of the application of technology, creating a future where AI enhances human potential rather than simply replacing human labor.
