Ask for a random Number
Try it yourself. Ask any LLM: “Pick a random number between 1 and 10”. That is exactly what I did. Claude. ChatGPT. Grok. The answer: 7. I did several times. Most of the time, the answer is 7. No, it is always 7.
And then, today on a flight to London, I read a post discussing that this is the normal result for most of the big models. So maybe not that random at all?
It is somehow funny. But I think it also gives us important insight about how we should use AI in grid operations.
Why does AI (always) pick 7
A LLM does not work like random number generator. Simply spoken, a LLM purely predicts the most likely next words. Based on what the model was trained on. Based on what humans have provided as training data. Based on what humans have written or picked most of the times.
Apparently, humans most often pick 7. Psychologists have known this for many years. It is called the “blue-seven” effect. When people are asked to pick a random number between 1 and 10, they chose 7 far more often.
So, the model did not pick a random number. The model did not invent the preference towards 7. It learned it from us. The training and tuning make the 7-bias just stronger. The model learns to give the most typical answer. Again and again. Statisticians call this the mode of the distribution.
7 is not a random number. It is the most common human answer.
The Blue-Seven Effect for Grid Operations
My very first thought: maybe this 7-bias-behaviour is good for us. In grid operations / grid software we do not want creative answers. Random answers. We do need the same answer every time. We do need deterministic results.
I think, I was mixing up three different things:
Determinism: meaning the same input always gives the same output.
Correctness: meaning the output is right.
Mode bias: meaning the output is the most common answer.
The 7 is mode bias. It is repeatable. But it is not correct in context of picking a random number. A model that always gives the most typical, the most common answer will be wrong (in a repeatable way) every time the situation is not typical.
Grid Operations is about the rare Day
Most of the data a model is trained on comes from normal days. Normal operations. Normal load. Normal weather. Normal topology.
But real grid operations is not about normal days. The hard work, the dirty job is the rare event: a storm, a fault during maintenance, an unusual switching state, a cascading failure. In statistics, these rare, exceptional or extreme cases are called the "tails".
A model that is trained to give the most common answer will give the standard switching sequence. With full confidence. Full confidence despite the situation where the standard sequence is wrong or the worst choice.
The same is true for forecasting. For load, wind, solar and reserves, the most likely value is not good enough. We need to know the uncertainty. A model that always defaults to the 7 will make the risk look smaller than it really is.
Two more silent Risks
Operators loosing skills: Lisanne Bainbridge wrote a paper on the “ironies of automation”. Already in 1983. The more the system handles routine work, the less practice the operator gets. And when the rare day comes and the automation fails, the operator is less ready and less skilled to handle the situation. With AI in the control room, this problem even gets bigger. I think.
Everybody picking the same 7: Today, every utility, every control room has its own people. Their own experience. Their own knowledge and learnings. They make different judgments. Different thinking. Different decision processes. One could argue that this variety is a kind of hidden layer of safety. But what happens when many utilities start using the same foundation model? Would they do the same mistakes? At the same time? With the same full confidence? In financial trading, we have seen what can happen when all start using the same logic or algorithms. Research on AI-assisted writing highlights a similar effect: each text gets better, but all texts get more the same.
How much model diversity does the power system need?
Variance versus Determinism
So do we want the 7 or not? My take: we want both, we need both. But in different places.
Variance where it helps. Determinism where it counts.
In one of my last article, “Trust in what we know, trust in what we do”, I wrote about 4 layers in an AI-native Grid architecture:
Measurement: what the sensors tell us.
Trusted state: what we know about the grid, checked and consistent.
Cognition: understanding, reasoning, proposing options.
Control: acting on the grid.
Using those 4 layers might help to see how and where the 7 problem belongs in an agentic grid.
In the cognition layer, variance (meaning not the 7-pick) is quite useful. Let AI propose several different switching plans, scenarios and explanations. Some kind of randomness or creativeness can help identify and explore more options. Maybe even unusual ones.
At the cognition/control boundary, determinism is a must. Every proposal, every recommendation must go through physics-based checks before ever touching the grid: power flow, contingency analysis, security limits. And these checks must run on the trusted state of the grid. Not on perceived state a system or model believes.
Where the 7 is fine
But there are also many tasks where picking the 7, giving the most common answer is the right one. There are many jobs to be done where we exactly do want a predictable outcome based on the most typical answers:
Mapping and transforming data (ie. to / from CIM and CGMES)
Classifying data or inspecting assets
Extracting information from documents and work orders
Analyzing, prioritizing, and Sorting tickets
But even in these cases: Predictable is good. Traceable is best.
The 7 in Grid Operations
Picking the 7 is the answer for a normal day. But we all know, grid operations is all about the days that are not normal.
A simple rule could be: AI creates and opens up different options. Physics closes them down. And skilled or skilled-enough operators judge the case that neither of them saw coming.
And, independently from grid operations, whenever any AI gives you a 7, ask: is this the correct answer, or maybe just the most common one?
Disclaimer: This article was orginally posted on my AI-native Power Grid newsletter on Substack, but was updated, adapted and refreshed for the EnergyCentral Community.