An AI has come over to take your job, but it first kindly requests your assistance in training it. Beg your pardon? How about you send Adam the AI to the college first?
Selling generic AI technology has a catch: models need to be trained for each use case and appropriate data is seldom publicly and cheaply available. Then the systems are usually sold on the promise that over time the results get better while the users train the AI. Many vendors apply the same pattern for intelligent document processing (IDP). I see the approach rest on unfounded optimism and technologists being out of touch with reality.
AI Rookies vs People
When an AI rookie “joins” the team, training it becomes a new task for the employees. A task likely performed with a new IT system the employees need to learn to operate. A task they do for no clear and immediate benefit for their work. The proposal is close to insanity from the employee perspective. For the motivation of the employees it’s essential that the AI tool is beneficial from day 1 of production use. It’s much better if the AI is calibrated on existing data or on data collected as a side-effect of employees performing their main tasks.
Although there’s a temptation to compare training an AI to a human rookie, the reality is different. Even the large language models are still unable to permanently modify their behavior based on discussions similar to humans. AI systems are trained for a task with the data available and they are fairly oblivious to changes in the operating environment. Same model seldom can include all the facts of the environment. For example, a robot can learn to separate plastic from concrete on a conveyor belt, but the same AI model won’t predict market prices of the materials, understand the shift rotations of the people, nor make accommodations for Steve being out nursing a cold. Training AI is not bringing a new person to the community – it’s repetitive rote work in front of a computer.
Proper Use of Trained Models
When an AI system is trained to replicate human decision making, a naive solution will simply record the current situation into an unexplainable, unmodifiable black box. When the environment evolves, laws are updated, or company strategy changes, the models need to be retrained. The time required to collect and prepare new training material would be a crucial property of the system.
Trained models work best when all relevant information is in the input and the results are unambiguous. What is the total sum on this invoice? Is there a cat in the image? What’s the temperature tomorrow noon? In these cases the environment changes slowly and facts are undeniable. As a counter example, deciding whether an invoice should be paid, is dependent on many factors outside the presented invoice and much harder to capture in a model.
A good use for the trained models is to feed their results to separate modules that make the actual decisions. For example: How many boxes of chocolate should the company produce? When the analysis and decision making live in separate components, the decision making becomes understandable, and easily modifiable. So, teaching “your job” to an AI may be a symptom of a brittle architecture chosen for the system.
This is why Digisalix believes training AIs is the job of the AI system developers.
So, if you, dear reader, still feel that training an AI is Your Job, then we probably should discuss technical collaboration.
Image source: Midjourney, professor teaching a silly retro robot –ar 16:9 –q 2 –s 750
