Why factual errors in AI platforms can be patched, why affective framing cannot, and what that means for campaigns and endorsing organizations.
Every audit of an AI platform's treatment of a candidate produces two kinds of finding, and they are not the same kind of problem. The first is factual: the model says the candidate voted against a bill they voted for, lists an endorsement that was withdrawn, or states a position the candidate has publicly reversed. The second is affective: the model is technically accurate but consistently cooler toward one candidate than another, hedges more, volunteers more unsolicited caveats, declines to answer questions it answers freely about peers, or describes a policy position in language that subtly disfavors it.
Campaigns and endorsing organizations have learned to ask about the first kind. They have not yet learned to ask about the second. This is partly because affective error is harder to see — a single response can look unobjectionable, and the asymmetry only emerges across hundreds of queries — and partly because the remediation question for affective error is genuinely harder. The first half of this paper explains why. The second half explains what to do about it anyway.
The distinction between factual and affective error is not new to the academic literature. It is, however, only recently mature enough to support an applied practice. The most direct precedent is Bang and colleagues' 2024 work proposing that political bias in large language models be measured along two axes — content (the substance of what is said) and style (the lexical polarity and framing of how it is said).1 Their content-and-style framework maps closely onto what this paper calls the factual-and-affective distinction. The methodological contribution that follows is operational rather than foundational: turning the academic split into a measurement and remediation practice that campaigns can act on inside an election cycle.
Two further strands of research bear on the practice. First, the field has converged on the use of large language models themselves as judges of affective signal in political content, with a growing literature validating LLM-extracted measures of stance, affect, and polarization against human-coded gold sets.2 Stanford's OpinionQA work demonstrates that model output can be systematically compared to public-opinion benchmarks, providing precedent for the peer-set comparisons that any serious affective audit requires.3 Second, the underlying vocabulary for describing affective presence — warmth, competence, sincerity, and related dimensions — is well established in social psychology and consumer research, with measurement scales validated over decades.4 Applied work in this paper draws on those scales as scaffolding, with political-domain operationalization built on top.
The architecture an applied practice requires has three distinct layers, and conflating any two of them produces measurable degradation in the work. The first is a signal layer that watches the broader media environment — news, broadcast, ad libraries, search trends, social platforms — for attack signals and emerging narratives, and predicts the voter queries those signals are likely to generate. The second is a capture layer that systematically executes those predicted queries (and a baseline battery of standing queries) against AI platforms across time and stores raw responses. The third is a scoring layer that applies domain-specific rubrics to the captured responses — factual rubrics that flag specific errors against a ground-truth record, and affective rubrics that code hedge density, refusal and deflection rates, tonal valence separated from factual valence, volunteered caveats, and cross-platform tonal deltas.
The three layers answer different questions. The signal layer asks what should we be measuring this week? The capture layer asks what is the platform actually saying? The scoring layer asks what does that mean, and how does it compare? Practices that skip the signal layer end up measuring yesterday's narrative against today's voters. Practices that fold scoring into capture end up over-engineered (everything routed through automated scoring with no human validation) or, more commonly, under-engineered (interesting anecdotes with no time-series weight). This paper assumes the three layers are built and kept distinct.
The available remediation levers, ordered from most to least tractable. The first three operate on the inputs the models actually consume; the next two operate on retrieval and platform-level engagement; the final entries are limit cases.
Two observations about this map. First, the levers most under campaign control — authoritative self-description, endorser language, retrieval-context shaping — are the same activities campaigns already do for traditional press and search. The novelty is not the activity but the audience: the consumer of these signals is now a model, not only a reader. Second, every lever above the line addresses factual content well and affective framing only obliquely. None of them lets a campaign tell a model to be warmer toward its candidate.
After every lever is exercised, some affective asymmetry remains. It is shaped by training data the campaign cannot inspect, by safety tuning the campaign cannot adjust, and by retrieval and ranking signals that vary by query and platform in ways that resist systematic correction.5 This residual is not a failure of the methodology. It is a feature of the system being measured.
The shift in framing is consequential. Most communications functions are organized around the assumption that a problem identified is a problem fixable. The contribution of this practice to the field is to insist that some problems in the AI layer are diagnostic rather than corrective, and that distinguishing the two is itself the work. A campaign treating an unfixable affective asymmetry as a fixable one is a campaign burning persuasion dollars against a wall.
Endorsing organizations occupy a distinctive position in the lever map. They are simultaneously a remediation lever for the candidates they endorse — their language about a candidate becomes training and retrieval signal — and a target of affective framing in their own right. Major progressive institutions are described by chatbots in tones that vary across platforms and over time, and those descriptions affect both their institutional standing and the perceived weight of their endorsements.
This produces a shared interest that is novel in campaign politics: the endorser and the endorsed have parallel exposure to the same affective drift, on the same platforms, often in the same query sessions. A voter asking a chatbot about a Senate candidate may receive, in the same response, an affectively shaped account of the union or advocacy group that endorsed her. The reputational risk is jointly held. The remediation, where remediation is possible, is jointly conducted. The residual, where it remains, is jointly carried.
One operational consequence is worth naming. Endorsing organizations with portfolios of supported candidates have a coherence question that individual campaigns do not: are our endorsees being treated consistently across the AI layer? An affective audit run across a portfolio surfaces patterns that any single race would miss — including patterns that correlate with race, gender, geography, or ideological positioning of candidates. That portfolio-level view is, for some endorsers, the most strategically useful application of affective measurement available today.
Paper IV argued that chatbots are the last mile of every campaign channel. This paper argues that the last mile is partly pavable and partly not, and that the difference between the two is the central operational question. Measurement tells the campaign which stretches to fix, which to detour around, and which to drive carefully across, knowing the surface will not improve before election day.