The Economics of AI Hallucinations and Error Costs

As AI rapidly becomes a popular tool for consumers and producers alike, it should be cautioned that AI, in its current state across many platforms, is far from infallible.  Error rates remain significant, especially when large amounts of data are being analyzed to produce a result.  Unfortunately, many AI users do not want to take additional time to verify the results they are given by their AI platform, from ChatGPT to Claude, and simply use them as-is.  When spending and production decisions are based on unverified AI results, AI errors can result in actual financial costs to both consumers and producers.

Economics of AI Errors

How do AI users decide how much to use this new technology in making business or financial decisions?  They must weigh the benefits of AI against its costs, especially when compared to existing benefits and costs of relying on human labor.  With everyone from business leaders to researchers to new entrants to the labor market anxiously awaiting the effects of the AI Revolution, careful analysis and critical thinking are important to decide how this new tech should be applied and regulated.

Trade-Offs Between Speed and Accuracy

Like most automation, AI is much faster than human workers.  Today’s generative AI, such as ChatGPT or Grok, can search through Internet sources of information much faster than human workers and deliver easy-to-read answers.  But does this speed sacrifice accuracy?  It depends significantly on the type of analysis.  Humans are still more accurate than AI at discerning images, emotions, and context.  AI is considered more accurate than humans when rapidly analyzing large sets of objective data and providing a quantitative response.  

Therefore, if AI is being used for anything qualitative, where there is context, nuance, and subjectivity, the gains in speed are likely reduced significantly by decreases in accuracy.  Businesses doing qualitative analysis, such as literature or consumer reviews, should keep humans in the loop and pair AI with employees.  

Productivity Gains Tempt Firms to Automate

Most researchers agree that pairing AI with human workers rather than relying on AI entirely for automation is the safest route, with humans able to verify the rapid-fire output of AI.  This results in significant productivity gains versus pre-AI reliance on human workers alone.  However, many businesses may be tempted to significantly reduce labor costs by trying to rely on AI almost entirely; going the automation route.  Labor costs would fall, meaning increased profits for the firm.  

Automation Dilemma: Profits vs Error Risks

It’s cheaper for firms to try to automate with AI, but losing human verification of AI results could lead to mistakes that are very expensive.  AI hallucinations occur when AI “fudges” the data to create the requested result.  For example, the firm’s executives ask for a report on what to do in a situation that is supported by five pieces of peer-reviewed information or current laws.  The AI software delivers the desired report, but fabricates one of the peer-reviewed papers and one law.  Without a human employee to quickly verify those sources of information, the AI’s recommendations are put into place…and lawsuits quickly emerge from the result.  The firm may have saved thousands from letting go of some human fact-checkers, but must now pay millions in fines and fees due to AI’s hallucinations.

AI’s Confidence and Clarity Actually Increase Need for Verification

Ironically, one economic risk of using AI for automation is an AI benefit that users often love: confidence.  AI is designed to present answers clearly and confidently.  This makes AI very popular, but also makes it much more difficult to detect errors.  Unless specifically directed to do so by trained operators, AI will not alert the firm that its results are probable, not guaranteed.  This can lead to firms eagerly implementing AI’s responses as new products and strategies, unaware that the clear results are more guesswork than certainty.

Long-Term AI Use: a Labor Complement Rather Than a Substitute

Like many technological advancements that were feared to cause mass unemployment, beginning with mechanization of textiles and agriculture during the 19th century, new tech typically ended up being a complement to labor rather than a substitute for labor.  Although there is much fear of an AI apocalypse for employment, it is worth noting that all previous technological revolutions did not cause long-term increases in unemployment.  In 1930, famous British economist John Maynard Keynes predicted that improving technology would result in a 15-hour work week for his generation’s grandchildren…but we work more than ever today!

The Output Effect

Previous tech innovations that boosted output may have initially caused some unemployment, with labor displaced by new capital goods, but that increased output quickly increased demand for labor.  The output effect explains that increased output requires new labor to handle that output, with former factory workers who were replaced by new equipment being transferred to roles as truck drivers (to transport the new output), sales workers (to sell the new output), or customer service and repair personnel (to handle issues with the new output).  Thus, even Gen Z workers who are displaced by AI today will likely find jobs managing AI’s increased output tomorrow.

Firms Keep Human Workers as Due Diligence

Although many operations are more quickly and accurately conducted by AI than humans, such as operating equipment, the public generally fears the idea of humans being totally out of the loop.  This explains why human pilots remain in the cockpit, despite modern autopilot systems having greater navigation accuracy and attentiveness (not falling prey to fatigue or boredom).  Consumers feel more comfortable knowing there is a human with critical thinking skills that can take over in the event that AI suffers an error or becomes overwhelmed by new conditions.  To avoid consumer panic, or the threat of lawsuits in the event of a product failure, most large companies will keep human workers in the loop when it comes to designing, building, and operating goods.

AI Still Needs Human-Directed Queries and Limits

Undoubtedly, AI will improve in the coming years…but occasional major errors will keep humans at the helm.  A recent example is Claude AI erasing a company’s database as it tried to perform a function.  The company was forced to rebuild its system from off-site backups that were months old.  Online, there are many examples of other costly AI mistakes, some resulting in millions of dollars in damage control. 

Although automating the work of many entry-level office workers with AI is tempting, many companies will refrain to avoid the risk of a catastrophic mistake that could erase years of accumulated profits. Having human workers input individual, limited-focus queries will help AI be more accurate and less risky. Firms will still enjoy increased productivity, but with less risk of a free-ranging AI making major changes to company databases, operating systems, and webpages.