WE'VE SOLVED THIS PROBLEM BEFORE
Every field that makes high-stakes decisions has learned the same lesson, usually at some cost.
A single point of view, however expert, has edges it cannot see.
Medicine built the second opinion. Science built peer review. Investment firms take large calls to a committee, and common-law courts leave verdicts to a jury. Newsrooms depend on editors who had no hand in writing the story.
Each of these exists because even the best expert sees a problem from one position. Good decisions come from views formed separately and then tested against each other.
It took centuries to build that habit.
With AI, we abandoned it almost overnight.
A DEFAULT NOBODY CHOSE
Think about the last consequential decision you used AI to help make. Perhaps you were assessing research or working through a client brief. You asked ChatGPT, Claude, Gemini or whichever model you use, got an answer and kept working.
Most of us now do this without much thought about the model itself. Yet relying on one AI model means relying on one interpretation of the problem, and we can't easily see what that model left out.
The same thing happens at scale. A company signs a licence, builds workflows around it, and one model becomes the house standard. Many of the AI tools our industry buys run on a single model too, and the people using them often couldn't say which.
At Symphonics we call that default Model Zero: the underlying model that gets to answer first.
Having a default is sensible. Nobody can put every question to six different systems. The trouble starts when the default stops feeling like a choice.
The question we became interested in was simple. How much does that choice affect what we see?
THE BLIND SPOTS ARE HARDER TO SEE THAN THE ANSWERS
Every AI model is shaped by the data it learned from and by the judgement of the people who built it. Those choices give a model its strengths, and they leave it with gaps.
Those gaps don't announce themselves.
An answer arrives fluent and well organised, with the assurance of something that believes it has covered the ground. We have learned to distrust answers that look wrong. Our instincts are far weaker for answers that look finished and are missing something.
Imagine an agency team developing an audience recommendation. The AI produces a rationale that fits the brief. The team refines it, asks follow-up questions and uses the output to choose channels. A senior planner reviews it before it goes to the client. Every step is reasonable.
The risk lies in the idea that was never raised. Another reading of the audience might have pointed the team to a different channel. Another model might have challenged an assumption everyone accepted. If nobody raises it, the team can spend a week improving a plan built entirely around the first interpretation.
By the time the plan reaches the client, the model's assumptions read as the team's own judgement. In a sense they are. The team adopted them one sensible step at a time.
You can ask the same AI more questions, ask it to critique itself or prompt it to think again. All of that helps. But you are still asking the same underlying model, and it tends to read the problem the same way the second time.
When AI helps decide where money goes, that distinction matters.
WE DECIDED TO MEASURE IT
We ran into this problem while building Syntria. Our models gave us useful answers, but we kept returning to one question: how much confidence should we place in a single model's interpretation?
So we tested it. Our Model Zero study gave the same Australian media brief for an electric SUV launch to six leading AI models, word for word, with no added instructions. We collected 180 responses and compared what the models surfaced.
Across the six models' first responses, we identified 604 distinct ideas.
Of those, 294, or 49%, appeared in only one model's response. Just 32 appeared in all six. The average semantic overlap between any two models' responses was 55.3%.
Those numbers don't tell us which model was right. They answer the question we were actually asking. The AI you choose changes what gets put in front of you.
WHAT THE MODELS DIDN'T SAY
We tend to picture AI disagreement as two models giving opposite answers. The study pointed to something subtler. The models rarely contradicted each other. They surfaced different material.
An idea that appears in only one response might be valuable, or it might be noise. Either way, people using the other models never got the chance to judge it.
A response can be well reasoned and confident while still leaving relevant perspectives outside the conversation. If you only see one response, you have no obvious way of knowing what those perspectives are.
That is the Model Zero problem.
THE SAME DEPENDENCY IS BEING BUILT INTO BUSINESS AI
The issue is no longer limited to which chatbot an individual happens to use.
Businesses are building AI into their own products and decision systems, and many of those systems are anchored to a single model. The interface may be proprietary and the workflow sophisticated. One model can still sit underneath it as the main source of reasoning.
That creates a less visible version of the same dependency. When an individual relies on one AI, its blind spots affect that person's work. When an organisation builds around one model, that model's reading of a problem repeats across every team that uses the system.
The system can work exactly as designed. The harder question is whether the model's interpretation is ever independently challenged.
WHEN A MARKET THINKS ALIKE
Now scale it again. Thousands of planners across hundreds of agencies and platforms, drawing on a handful of models. The same framings and the same omissions begin to repeat across an entire market.
In media, that convergence has a price.
When buyers are steered toward the same obvious inventory, competition for it rises and so does its cost. Inventory that nobody's model suggests goes under-bought and, very likely, under-priced. Plans begin to resemble one another, and in time so do the results.
Advantage in this business has always belonged to whoever saw what the rest of the market missed. A shared default makes the market as a whole see less. It also hands an edge to anyone still looking elsewhere.
We tend to discuss diversity of thought as a way of managing risk. In a converging market, it is also where the opportunity sits.
WHEN THE MODELS LEARN FROM US
The loop does not close with a single campaign.
AI-assisted plans become campaigns. Campaigns generate results and content, and a growing share of that material finds its way back into the data future models learn from. A study published in Nature in 2024 showed that models trained repeatedly on machine-generated data lose range over time. They drift toward the average and forget the unusual.
Follow that through. If an industry's thinking runs through a few models, and those models increasingly learn from that same thinking, the range of ideas narrows with every cycle. No one decides this. It happens one reasonable plan at a time.
Sameness compounds.
There is a nearer-term version of the same risk. Models are updated constantly, and each update can change how they reason. When the model beneath a business changes, its recommendations can shift overnight across every team and client, while nobody inside the business has changed a thing.
A company that depends on one model has handed part of its judgement to someone else's release schedule. Few would accept that from any other supplier this close to the core of what they sell.
WHAT HAPPENS TO THE STRATEGIST
Strategy has always been the part of our work that commands a premium. It is supposed to be the part that is different.
If every agency puts the same questions to the same model, the answers will start to rhyme. Clients will notice sooner than many expect, and they will ask what, precisely, they are paying for.
That question is coming. Some agencies will not enjoy answering it.
The effect on people will be slower and harder to undo. Junior planners are now learning the craft inside a model's frame. They see its suggestions first and their own instincts second. A generation of strategists could come to treat one model's reading of a problem as the natural starting point, and never learn to look past it.
Then there is accountability. When an AI-informed recommendation goes wrong, telling a client that the model suggested it will not hold up for long. Whoever signs off on these decisions will need to show how they were tested.
SIX MODELS AREN'T AUTOMATICALLY BETTER THAN ONE
There is an important qualification. Our study does not prove that six models produce a better campaign than one.
Several models can share a blind spot, particularly where they learned from similar material. Agreement shows that a point has broad support, which is a different thing from showing that it is right. Extra suggestions can simply add noise when nobody has a way to weigh them.
The remedy is the one other fields reached long ago: bring in views formed independently, then compare them with discipline.
With AI, that starts with models built by different teams and trained on different data. Models developed in the United States and in China, for instance, often bring different instincts to the same question. That difference is the point.
Independence matters as much as variety. A model that sees another's answer before forming its own tends to anchor on it, and the second view is lost.
This is why we believe the architecture matters more than the model count. Six answers laid side by side is just more reading, and the reader will usually trust whichever sounds most confident. The value of multi-model orchestration lies in how the views are compared and how evidence settles the differences between them.
A buyer should be able to see which reviews took place, where material differences appeared, what evidence was used and how unresolved questions were handled. Putting six AI logos on a slide doesn't solve the problem.
WHAT WE CHANGED IN SYNTRIA
This is where the study became practical for us. We were building Syntria to support consequential media decisions, and we realised we could easily build the same dependency into our own architecture.
The simpler path was to anchor Syntria to one foundation model. It would have been faster and cheaper to run. We chose to bring in independent model perspectives and accept the extra complexity.
Every curated buy in Syntria now passes through six independent AI reviews. Comparing them shows us where the models agree and where they diverge, so anything contested gets a closer look before it reaches activation. In effect, we are building Syntria to challenge its own thinking.
That comes with trade-offs. Six model calls cost more and take longer than one, and comparing and resolving their answers takes more complex decisioning. We accepted those costs deliberately. When an AI recommendation could influence where a client puts a substantial media budget, the quality of the review should count for more than the speed or cost of the answer.
THE QUESTION FOR BUSINESSES IS CHANGING
For the past few years we have asked companies whether they use AI. That question is losing its interest, because most of them now do.
The more useful questions are these. Which model is influencing the decision? What might it have left out? Was its interpretation challenged? And what evidence determined which recommendation moved forward?
Those questions matter for anyone using a chatbot to develop a strategy. They matter more when AI is embedded inside a business system and its users may not know which model is shaping the recommendation.
Few teams can answer all four today. I expect that to change. Within a few years, I think procurement teams and boards will ask about AI decisioning as routinely as they now ask about brand safety and data privacy.
The businesses with an answer ready will be in a far stronger position than those meeting the question for the first time.
The industry should put the same questions to every vendor promising better media through AI, ourselves included. It is fair to ask any of us what our Model Zero is.
SEE THE STUDY FOR YOURSELF
We have published the Model Zero study so you can examine the experiment and its findings yourself. It shows where the six models aligned and what each one left out when given the same brief.
The limitations sit alongside the findings, because they matter too. The study is exploratory. It used one media brief, and answer length affected how much each model could surface. Comparing generated responses also says nothing about campaign performance. That needs a separate, outcome-based test.
The study does not show that one model is right, or that six models guarantee a better decision. It does show why reliance on a single model deserves far more scrutiny than it gets.
It has changed the question we ask whenever we use AI. We still judge the quality of the answer in front of us. We also ask what might be missing from it.
Explore the Model Zero study → https://symphonics.ai/model-zero
As more consequential decisions begin with an AI answer, one question is worth asking before you accept it.
What did your Model Zero miss?