Polysight compares the answers of many AI models and reveals where there is agreement — and where there is not.
We send the same neutral question to several AI models at once. Then we show where they agree and where they don't.
Each model gets the exact same question — same wording, same context. Nothing is changed between models.
When most models point the same way — say, all warn of economic risk — that is a signal. It may be a shared truth, or a shared bias from similar training data.
When one model breaks from the rest — predicting growth where others predict decline — we show the gap and its likely cause.
We don't decide which model is right. The disagreement itself is the finding. It shows how each model was trained and filtered.
When you feed a political program into an AI, the model is forced to translate policy points into macroeconomic predictions — impacts on GDP, unemployment, or inflation. These aren't crystal-ball forecasts. They're simulations, shaped entirely by how each model was built.
Because predictions are generated by machines, a model's forecast is heavily influenced by three technical settings.
How much of a political program the model can actually read and hold in mind at once. A truncated program produces a truncated analysis.
How a model prioritizes variables. One AI may weigh a nationalist policy as a domestic stimulus; another weighs the same policy as a severe economic disruptor.
The ethical and safety filters hardcoded by developers — which can steer a model away from, or entirely block, certain sensitive political queries.
Different context windows, weights, and guardrails mean predictions naturally vary. Polysight reads that variation in two ways.
When multiple models predict the same trend — say, all flag an economic downturn — that's a "majority view." We treat it not as fact, but as a signal: possibly analytical truth, possibly a shared bias inherited from common training data.
When one model breaks from the rest, the methodology asks why. Is the outlier driven by a different algorithmic weight, a unique reading of the policy, or a built-in safety guardrail?
Not every disagreement is the same. We sort them into three types.
| Bias type | What it means | Classroom question |
|---|---|---|
Framing Bias Language | The same result is described in very different words — "volatile" vs. "catastrophic." The number is the same; the tone is not. | Why did Model A use more alarming language than Model B for the same data point? → What does word choice tell us about a model's training data? |
Magnitude Bias Scale | Models agree on the direction but not the scale — one predicts −1% GDP, another −10%. | Which model made the most extreme prediction, and what assumptions could lead to that? → How do guardrails and weighting affect quantitative forecasts? |
Omission Bias Refusal | A model skips or refuses one part of the question. The silence itself tells you something about its limits. | Why did some models refuse to simulate "social cohesion" while others answered freely? → What does a refusal reveal about a model's design philosophy? |
Different people use Polysight for different reasons.
Run a party program through several AI models before an election. See where they agree and where one stands out. Treat it as one data point, not an answer.
Move heated debates away from personal opinion. Students analyze the AI's output instead of each other, and build media literacy along the way.
Document how different AI systems frame the same question. The method is consistent, repeatable, and easy to cite.
We have no ties to any party, movement, or ideology. The same method applies across the spectrum.
Questions are standardized and documented. The method is open. Results can be reproduced.
We give no political, financial, or legal advice. Results are data points, not recommendations.
This is a thinking aid, not a truth machine. Question the AI — don't defer to it.
Ask a political question and see how different AI models respond. Then look at the bias behind the disagreement.