- 14 Jun 2024
- 4 Minutes to read
- DarkLight
How Henry Ford Invented AI Governance
- Updated on 14 Jun 2024
- 4 Minutes to read
- DarkLight
Thank you to Matt Konwiser for sharing his insight in our knowledge base.
Conscientious AI Design | CTO | Educator | Columnist
![matt.png](http://cdn.document360.io/3e5067ce-3530-4015-94e8-f530f6e78b17/Images/Documentation/matt.png)
The cost of the Ford Model T was high, work in the plant was not optimized, and consistency and quality may have been impacted.
So, Henry Ford did something about it.
In December of 1913, he launched the moving assembly line. This compartmentalized the labor, forced a level of efficiency and process to keep the line going, and created specialized tasks for each person.
As with any innovation, that didn't necessarily sit well with existing workers - some had a difficult time keeping up with the line. Others hated the mundane repetitive work because instead of build a car, they became simply the door station or the fender station. People started leaving.
To counter this, Ford increased wages and shortened the shift length by an hour, then adding a third shift. Workers could work less, make substantially more, and Ford could now operate his plant 24 hours a day.
In addition, this new efficiency dropped the price of his cars by almost 75% within a decade of starting the line.
I'm paraphrasing quite a bit of the story, but if you're interested in the whole deal, please enjoy this few-minute read: https://corporate.ford.com/articles/history/moving-assembly-line.html
So what now does this have to do with AI?
AI is not an IT asset. It isn't hardware, nor is it software. You don't always have to pay a license fee to use it, and depending upon how you build it, it may never fully depreciate.
AI is effectively a new class of artificial worker.
But in order to construct these workers which can have a mind of their own and easily end up off scope or out of control, there needs to be a process.
That process must include methodologies to select models, train or tune them, monitor the inferencing, and also ensure that prompting and prompt outputs are appropriate. There also needs to be a way to audit the overall AI projects so it's possible to validate they are operating within state, federal, and industry specification and regulations.
The process must also be consistent for all models constructed within an organization. The greater the variance, the more likely you'll see drift, errata, and misuse, not to mention the potential wasting of resources and time.
I was pondering the way many businesses have been trying to use AI for anything, or worse, picking up an AI-powered lance and tilting at their company's windmills to find new ways to solve problems that may not need to be solved with AI.
![image(324).png](http://cdn.document360.io/3e5067ce-3530-4015-94e8-f530f6e78b17/Images/Documentation/image(324).png)
Taking into consideration the increasing maturity of how businesses are trying to manage AI, appointing lofty titles such as "Chief AI Officer" and "CEO of AI" to get this sprawl under control, it seems that perhaps a simpler solution from "simpler times" might be the answer.
If you have something new that is somewhat easy to build but poses great risk if built improperly and could be a massive expense and be poor quality if done wrong, standardization is critical.
Enter the 1913 Henry Ford Model T moving assembly line.
As with a car, humans use the finished AI product. There may be no specific end of life date. There can be myriad aftermarket add-ons. Even if mass produced, there is room for some factory-floor variations without sacrificing the process, the quality, or the speed of production.
Henry Ford, without realizing it, created the construct for AI governance pipelines.
Most people might associate the term "governance" with regulations and audits. Understandable given the word sounds eerily like government. The natural root however simply means "to steer"; and AI needs a lot of steering.
Governance for AI initiatives should be all about establishing a process and using that process to efficiently craft, deploy and oversee the models in use. It ensures, to the best of an organization's ability, that the models color within the lines and are well protected from abuse.
With the right assembly line, it's possible for a business to mass produce smaller models that are all fitted with the same overall guardrails required to ensure trustworthy deployments, then each owner can add their own unique flair to tailor their model to achieve the desired outcome.
Instead of governance being a compliance checkbox, a burden and an afterthought, it becomes central to large scale deployments of AI models for any number of potential applications, and not just generative models. An AI pipeline can be applied to predictive ML and generative AI with equal effectiveness.
It doesn't need to be more complicated than that.
![image(325).png](http://cdn.document360.io/3e5067ce-3530-4015-94e8-f530f6e78b17/Images/Documentation/image(325).png)
Yes, the devil is in the details. Much planning has to go into the design of the pipeline and the rules for pushing a model through it, as well as use and training. However, the chief concerns that most people express about AI: drift, tampering, regurgitation, misuse, hallucination, and more are all effectively managed with nothing more than this relatively simple, and over 100 year old strategy.
Sometimes, what's old is new again.
Remember, with the level of innovation and big thinking that has happened since the Industrial Revolution, there are plenty of proven strategies which can be leveraged when breakthroughs are discovered. It doesn't always require something new to manage something new yet it still needs to be properly managed.
Governance before deployment, always.