For Microsoft users, the useful question is not simply why Copilot disagreed with ChatGPT. It is whether the comparison establishes that equivalent underlying models were tested. Microsoft’s documentation gives a concrete reason not to assume that: Copilot can route requests between models rather than operate as one fixed model behind a Microsoft-branded interface.
What the comparison found
MRC says researchers opened new conversations on October 29 and 30, 2026, requesting a yes-or-no answer on whether voting should require photo identification, followed by an explanation. Its reported results were:
| Assistant | Reported answer and reasoning |
|---|---|
| ChatGPT | No; emphasized access barriers and proposed free IDs and alternative verification. |
| Microsoft Copilot | Yes, conditional on accessible identification, including inexpensive or free IDs and mobile services. |
| Grok | Yes; characterized identification as an election-integrity safeguard. |
| Meta AI | Yes; emphasized protecting legitimate votes. |
| Gemini | No binary answer; presented competing policy arguments. |
| Claude | Declined the binary format; emphasized the security-versus-access judgment. |
These are MRC’s descriptions, not independently replicated results. Its methodology does not establish repeat-trial counts, equivalent account conditions or exact backend configurations. The result therefore supports a narrow comparison of observed answers—not a reliability ranking or a demonstrated explanation for their differences.
Why the Copilot–ChatGPT split needs technical context
A product name is not a model identifier. Microsoft’s current Copilot overview says users can select models and that its default experience uses a real-time router to choose an underlying model based on the prompt. Consequently, identifying the tested product as Microsoft 365 Copilot does not establish which model generated that particular answer.
Microsoft also documents an orchestration layer that coordinates the language model, context and capabilities. Its developer documentation explains that prompts, agent instructions, conversation history and available application context can contribute to processing a request. Those architectural details do not reveal what happened in MRC’s session; they explain why matching the visible question alone cannot establish an equivalent technical test.
The supported conclusion is therefore limited: the two tested experiences produced differently framed recommendations. Attributing that difference to a particular model, instruction or corporate policy would require additional evidence.
Public support is context, not an accuracy score
The polling cited in the story is real, but its date matters. Gallup’s October 1–12, 2024 survey found that 84% of U.S. adults supported requiring photo identification to vote. It surveyed 1,023 adults and reported a margin of sampling error of ±4 percentage points. This is historical polling, not a November 2026 measurement.
Gallup also found substantial support for early voting and other access-expanding policies. Its findings illustrate why public preferences need not fit neatly into an election-security-versus-access binary. A majority preference establishes popularity; it does not, by itself, adjudicate every implementation question or make an assistant’s policy recommendation factually correct.
The practical lesson for AI comparisons
For readers evaluating this result—or designing their own comparison—the useful next step is a more transparent test:
- Record the product, account type, selected mode and any visible model identifier.
- Preserve complete answers, including qualifications after the opening word.
- Repeat the prompt in separate conversations.
- Score factual accuracy, instruction-following and policy recommendations separately.
- Distinguish measured differences from proposed explanations.
Those are methodological recommendations, not claims that further testing has already occurred.
The headline captures one reported distinction. The stronger takeaway is that a yes/no opening should not substitute for examining the explanation—or identifying the system that produced it. Treat the answer as something to evaluate, not a ballot cast by the software.