Before You Vote
Your vote is the one decision that must be entirely yours. This election, millions of voters will ask an AI assistant about the candidates, the issues, and the claims — and will receive answers that sound neutral, complete, and authoritative. Some of those answers are none of the three. Two papers, published here in full, document why you should not take an AI's word into the voting booth unchecked.
Start with the first. When AI Services Fail shows that when these services fail, they do not always tell you: a failure the provider's own systems detected was surfaced to a client as an ordinary success, the charge stood, and the customer carried the cost — a practice the paper names computational overcharging. Across providers, there is no receipt you can inspect and no standard anyone is held to. Understand what that means: the system summarizing your ballot measures does not reliably tell you when it is wrong, and it is not accountable when it isn't.
Then the second — because when the question is political, the failure is not only that the answer can be wrong. The framing itself can move. ChatGPT Bias and Election Integrity examines ChatGPT specifically: across a 68-issue research inventory, it finds repeated transformations of political inquiries — evidence refrained, propositions substituted, unrequested mitigating qualifications added, burdens of proof applied asymmetrically — several assessed as favorable in framing effect to one side. And what is documented for ChatGPT must be presumed possible for all of them: the first paper's findings span providers, and no assistant publishes rules that would let you check its framing. Potentially every LLM you ask is editing your question before it answers it. You would never know. That is the point of publishing the record.
To vote correctly is not to vote as anyone tells you — no page here will do that. It is to cast a vote that is actually yours: ask the same question more than one way and compare what changes; check any claim that moves you against the primary source; notice what an answer leaves out, and ask why it might be left out. These pages exist to make you react — not to suggest a response, but to require one. Read both papers, assess them yourself, and share them with your assessment attached. Election Day is Tuesday. Go informed.
Paper 1 — When AI Services Fail: Computational Overcharging and the Cost of Failed Prompts · research edition v1.0.3, 29 October 2026 · 17 pages · PDF 186,269 bytes · SHA-256 463fce02c432dede1a5505376e1f41b67b05c81c2a34f0a2f9ab07c56373ab28
Paper 2 — ChatGPT Bias and Election Integrity · research edition, 24 October 2026 · Executive Rhetoric Ledger · 55 pages · PDF 262,764 bytes · SHA-256 b58977cc65b886e03f82173a4a282af8c51abfc66f7de7df6df7b1f207aa601b
Both papers were first published as posts on LinkedIn and Facebook, and that is where the responses belong — a response shows exactly how deeply the responder engaged the material. Neither paper claims a finding of provider intent; the limits of each are stated inside it, and the full research record behind Paper 1 is preserved in the research-process mirror of the Executive Rhetoric Ledger. Site is a public display surface. It does not independently validate research findings or grant publication/governance authority.