Murphy’s AI Laws. Because If Intelligence Can Fail, It Definitely Will

By Jorge Garcia

 

Credit: MrWashingt0n on Pixabay

If you’ve been in tech long enough, you eventually realize that the universe is held together by two things: duct tape and Murphy’s Law. 

Years ago, I wrote the “Murphy’s BI Laws,” which sadly still hold true (and in my view, remain truer with each release cycle).

Now that we’ve entered the AI-everywhere era, where everything from your fridge to your accounting system thinks it’s a master philosopher, it’s time to face the truth: 

Murphy followed us into AI, and he’s having the time of his life.

So, here are the brand-new, fully field-tested Murphy’s AI Laws, written with love, pain, and several hours lost talking to a chatbot that insisted 2 + 2 “felt like 5.”


1. If an AI system can hallucinate, it will, and at the worst possible moment.

Corollary: The confidence level of the hallucination will rise in direct proportion to the importance of the meeting where it is presented.

 

2. Any AI model you deploy is obsolete the moment you push it to production.

Corollary 1: A new state-of-the-art model will appear at that exact minute.

Corollary 2: It will need triple your current GPU budget.

Corollary 3: Your CIO will ask why you didn’t predict this with AI.

 

3. The probability of your agentic AI failing increases exponentially with the executive level of the person using it.

If the CEO tests it? Prepare the incident-bridge call.


4. The biggest AI failure will never show up during testing.

It will only appear:

  • During your “go live,”
  • Before your coffee,
  • Or five minutes after you told leadership, “We’re confident in the model.”

5. If an AI bug exists, it will only be discovered by an end user, preferably one who’s been skeptical about AI since Windows 95.


6. The likelihood of AI misunderstanding a simple instruction approaches 100% when the instruction is given using natural language.

Corollary: The more carefully you phrase the prompt, the more chaotic the output.


7. No matter how many GPUs you have, you’ll never have enough.

Corollary: Your CFO will start Googling “cost-cutting strategies” within weeks.


8. The probability of AI delay rises in direct proportion to the number of teams claiming to ‘own the model.’

Corollary: AI governance meetings multiply faster than LLM tokens.


9. If users have any opportunity to mislabel training data, they will.

Corollary: And they’ll insist the AI is the one that’s wrong.


10. Your main subject matter expert for model training will always be the person with the least time available.

Bonus: Their calendar will be mysteriously blocked for the next six months.


11. The probability that a key ML engineer leaves during your AI project rises in direct proportion to how much undocumented code they wrote.


12. If there’s a way for AI to misinterpret context, it absolutely will.

Ask it for a summary; it writes poetry. Ask for poetry, and it outputs Kafka.


13. The more “explainable” the AI claims to be, the less anyone will understand the explanation.

Corollary: The explanation will still somehow blame your data.


14. The chance that your model drifts increases with the importance of the metrics it supports.

You wanted accuracy? Enjoy the drift.


15. The more people promise “AI will solve this,” the more likely it is that AI is the reason you’re in the meeting.


Final Thoughts: Why We Laugh Instead of Cry

Behind the humor is something real: AI projects, like BI before them, live at the intersection of messy data, messy humans, and messy expectations.

And no matter how advanced our models get, one universal truth remains: complex systems fail in complex ways.

But like with all great tech revolutions, we learn, iterate, fix, enhance… and occasionally shout at a chatbot that refuses to understand what “No, not that!” means.

Murphy may still be with us, but at least in AI, now we can ask a model to generate a meme about it.


Dearly,

Jorge Garcia

Comments

Post a Comment

Popular posts from this blog

Machine Learning and Cognitive Systems, Part 2: Big Data Analytics

By Jorge Garcia
Image
In the first part of this series, I described a bit of what machine learning is and its potential to become a mainstream technology in the industry of enterprise software, and serve as the basis for many other advances in the incorporation of other technologies related to artificial intelligence and cognitive computing. I also mentioned briefly how machine language is becoming increasingly important for many companies in the business intelligence and analytics industry. In this post I will discuss further the importance that machine learning already has and can have in the analytics ecosystem, especially from a Big Data perspective. Machine learning in the context of BI and Big Data analytics Just as in the lab, and other areas, one of the reasons why machine learning became extremely important and useful in enterprise software is its potential to deal not just with huge amounts of data and extract knowledge from it—which can somehow be addressed with disciplines such as data ...
Read more

Teradata Open its Data Lake Management Strategy with Kylo: Literally

By
Image
Still distilling good results from the acquisition of former consultancy company Think Big Analytics , Teradata , a powerhouse in the data management market took one step further to expand its data management stack and to make an interesting contribution to the open source community. Fully developed by the team at Think Big Analytics, in March of 2017 the company launched Kylo –a full data lake management solution– but with an interesting twist: as a contribution to the open source community. Offered as an open source project under the Apache 2.0 license Kylo is, according to Teradata, a new enterprise-ready data lake management platform that enables self-service data ingestion and preparation, as well the necessary functionality for managing metadata, governance and security. One appealing aspect of Kylo is it was developed over an eight year period, as the result of number of internal projects with Fortune 1000 customers which has enabled Teradata to incorporate several be...
Read more

SAP Data Hub and the Rise of a New Generation of Analytics Solutions

By
Image
“Companies are looking for a unified and open approach to help them accelerate and expand the flow of data across their data landscapes for all users. SAP Data Hub bridges the gap between Big Data and enterprise data, enabling companies to build applications that extract value from data across the organization, no matter if it lies in the cloud or on premise, in a data lake or the enterprise data warehouse, or in an SAP or non-SAP system.” This is part of what Bernd Leukert, SAP’s member of the executive board for products & innovation mentioned during SAP’s Big Data Event held at the SAP Hudson Yards office in New York City as part of the new SAP Data Hub announcement and one that, in my view, marked the beginning of a small yet important trend within analytics consisting on the launch or renewed and integrated software platforms for analytics, BI and data science. This movement, marked by other important announcements including Teradata’s New Analytics Platform as well ...
Read more