The LLM had no access to external context, metadata, or background information beyond the article text or headline and the fixed prompt. Its evaluations were based solely on linguistic and emotional cues present in the text. Innohives did not define “sympathy” for the model; instead, the model relied on its internally learned statistical associations of how sympathy is expressed in natural language. This avoided imposing a researcher-defined criterion that could shape or bias the results.

Because the model was applied identically to every input, the resulting classifications are comparative rather than subjective. No outputs were manually reclassified or adjusted after model evaluation. All model parameters were set to default settings, to avoid any bias imposed by the researchers. Specifically, while non-zero temperature settings introduce the possibility of minor variation at the individual classification level, the size of the dataset ensures stability in aggregate patterns. The analysis is concerned with large-scale trends

rather than individual cases. Additional details on validation and controls are provided in the Asserson Report.

Overall, the use of an LLM enables systematic, replicable identification of sympathetic framing at scale. Any latent model bias remains constant across all inputs and therefore does not affect comparative results across actors, formats, or time periods.

Sympathy Ratio:

Following LLM classification, Innohives calculated sympathy occurrences for each actor category. A Sympathy Ratio expresses how often sympathetic framing toward one actor appears relative to another within the same dataset.

For example, if sympathy toward Gaza appears in 600 articles and sympathy toward Israel appears in 300 articles during a given period, the resulting Sympathy Ratio is 2:1. This indicates that sympathetic framing toward Gaza occurred twice as often as sympathetic framing toward Israel in that dataset.

Sympathy Ratios can be calculated for any time interval, enabling comparison across periods and the examination of trends over time. Importantly, the Sympathy Ratio reflects relative narrative emphasis between actors; it does not measure audience exposure, readership, or editorial intent.

Experiment and Results

(The Initial 4 Month Timeframe):

Main Text Results and Analysis:

This section presents the results of Innohives’ sympathy analysis of CBC coverage during the first 4 months of the Israel–Hamas war, hereafter referred to as the initial reporting timeframe. This period was examined first because, among other reasons, it contains the highest concentration of CBC coverage and establishes a baseline against which subsequent reporting can be compared.

Sympathy Levels Over Time:

Innohives began by assessing overall sympathy levels using raw counts of sympathetic occurrences, defined as articles flagged as expressing sympathy toward “Israel” or “Gaza.” Each article was queried independently for all 6 actor categories, and could register sympathy toward one, some, all, or neither actor.

Across the initial reporting timeframe, Innohives analyzed 921 CBC articles containing the core keywords. Of these, 896 registered at least 1 sympathetic response to either Israel or Gaza. The distribution comprised 346 sympathetic occurrences toward Israel, and 550 toward Gaza.

Sympathy Ratios are expressed by dividing the number of sympathetic occurrences for Gaza by those for Israel. In this period, the resulting ratio is 550/346, or 1.59, meaning that for every instance of sympathetic framing toward Israel, there were approximately 1.6 instances toward Gaza. This ratio is mechanically derived from the model’s binary classifications and is operationally blind to editorial intent or institutional context.

Examining how these sympathetic occurrences evolved over time provides further insight. As shown in Figure 1, during the first week following October 7, CBC’s coverage contained more sympathetic occurrences toward Israel than toward Gaza. This is the only time point in the dataset where Israel received greater sympathetic framing. In the following week, sympathetic occurrences toward Gaza increased beyond Israel’s initial peak, while sympathy toward Israel declined. This produced a divergence that persisted throughout the remainder of the 4 month period.

Crucially, this pattern demonstrates that the measurement is sensitive to contextual shifts in reporting. The reversal following the first week indicates that the model’s classifications track changes in emotional tone and narrative framing rather than reflecting a fixed preference for either side.

As Figure 1 also shows, the absolute number of sympathetic occurrences declined over

time, mirroring the overall reduction in news volume as the war progressed. Importantly, this

decline affected both sides. However, despite the reduction in total coverage, sympathetic framing toward Gaza consistently exceeded that toward Israel, with a stable and substantial gap maintained across the full initial reporting timeframe.

Because raw counts are inherently sensitive to fluctuations in coverage volume, subsequent analysis does not rely on absolute appearances alone. From this point onward, sympathetic framing is assessed using proportional measures (i.e., relative to number of articles published). This approach ensures that observed differences reflect patterns of narrative treatment rather than changes in the quantity of reporting over time.

Actor-Level Proportions, Grouping Effects, and Civilian-Militant Framing:

To examine how sympathetic framing varied across different actors, Innohives analyzed weekly proportions of articles flagged as sympathetic for 6 categories, grouped into three analytical classes:

• State actors: Israel, Gaza
• Civilian actors: Israeli people, Palestinian people
• Militant actors: IDF, Hamas

Each proportion was calculated by dividing the number of sympathetic occurrences for a given actor in a given week by the total number of articles published during that week. This approach captures the likelihood that an article expressed sympathy toward a given actor, rather than the absolute volume of coverage.

Figure 2 presents these weekly proportions, while Figure 3 shows the average sympathy proportions across all six actors over the initial reporting timeframe.

Several patterns emerge. Sympathy toward Israel begins at a relatively high level, with a proportion of approximately 0.65 in the first week, but this proportion declines steadily to below 0.4 on average across the period. A similar decline is observed for Israeli civilians, whose sympathy proportion falls from roughly 0.7 in the first week to an average of about 0.4.

By contrast, sympathy toward Palestinian civilians begins at a comparable level to that of Israeli civilians in the first week but remains consistently high thereafter, averaging approximately 0.7 across the entire period (see Figure 2). Gaza follows a similar trajectory. As a result, while civilian sympathy initially appears broadly symmetric, it diverges rapidly and remains so throughout the remainder of the initial reporting timeframe.

Given this strong emphasis on civilian suffering, Innohives next examined whether sympathetic framing was in fact concentrated on civilians, or whether it also extended to armed actors. To assess this, an additional metric, the Militancy Ratio, was introduced. This metric compares levels of sympathetic framing of militant actors (IDF + Hamas) to that of civilian populations (Israelis + Palestinians).

For the initial reporting timeframe, the Militancy Ratio was 0.449. In practical terms, this means that for roughly every two articles expressing sympathy toward civilians, one article also expressed sympathy toward a militant actor. Thus, emotionally sympathetic framing was not confined to civilians, even amid a reporting environment ostensibly centered on civilian suffering. Notably, proportional sympathy levels for Israel, Israeli civilians, and the IDF remain closely aligned over time. This is not observed on the Palestinian side, where Palestinian civilians and Gaza are consistently treated as distinct from Hamas.

Notable Deviations in Sympathy Patterns:

Anecdotally, 2 periods in the dataset are notable. In the week of December 16, 2023, the sympathy proportion for Hamas exceeds that of both Israel and the IDF and approaches that of Israeli civilians (see Figure 2). In the week of January 13, 2024, sympathy proportions for Israel and Israeli civilians converge with that of Hamas, while sympathy toward the IDF falls below all 3 (see Figure 2).

These instances do not drive the overall results but illustrate the extent to which actor-level groupings can shift within the broader pattern. While detailed statistical significance should be assessed before a conclusive statement can be made, it is clear that the state of Israel did not receive more sympathy than Hamas, a recognized terror organization in Canada.

Taken together, the data show a rapid pivot from near-equivalent sympathetic framing in the first week of coverage to a sustained and structured divergence thereafter. This divergence persisted across the entire initial reporting timeframe, even as overall coverage volume declined.

Headline Results & Analysis:

Headlines function as a summary of a given article, read before the main text of the article.23 Ideally, they capture the story told in and the style/content of the main text, but also entice the reader to consult the main text for more.24 In today’s global marketplace, the majority of the world is hyper-connected to international affairs, and social media makes information and content available in an instant.25 A result of that is a marketplace of information and news that is highly competitive, and the CBC, as a publicly-funded broadcaster, is not immune to the competitive climate.

As Asserson (2024) put it, “To encourage click-throughs, [headlines] are often designed to excite, intrigue, question, shock or provoke. In doing this, they can reflect bias in a more binary manner i.e. for one side or the other, resulting in a ‘them or us’ mentality.” But research suggests that a strong percentage of people only read the headline of an article, rather than the headline and main text.26 Thus Innohives ventured to categorize the sympathy in headlines independently from main texts, utilizing the same dataset and methodology as done with main texts.

By analyzing headlines independently from article bodies, while applying the same dataset and methodology used in the main text analysis, Innohives methodology allows examination of whether headline framing mirrors, amplifies, or diverges from the narrative treatment found in the underlying reporting.

Headline Sympathy Levels Over Time:

Across the initial reporting timeframe, Innohives flagged 364 cases where CBC headlines expressed sympathy toward either Israel or Gaza. Of these cases, 275 were sympathetic toward Gaza and 89 toward Israel, producing a Gaza:Israel Sympathy Ratio of 275/89, or 3.09.

The proportional distribution of sympathetic occurrences in headlines reveals a marked asymmetry. Sympathy toward Israel and its associated civilian and militant categories remains substantively low across all headline terms; no Israeli-associated category exceeds a proportional sympathy rate of 0.15, as shown in Figure 4. By contrast, Figure 4 illustrates that sympathetic framing for Gaza and Palestinian people appears at roughly twice the level observed for any Israeli-associated category. As in the main text analysis, Hamas remains the category with the lowest level of sympathetic framing in headlines.

Temporal examination further underscores this divergence. As shown in Figure 5, during the first week of the war, sympathetic proportions for Gaza and Palestinian people in headlines were already slightly higher than those for all Israel-adjacent categories except Israeli people. This contrasts with the main text analysis, where sympathy toward Israel marginally exceeded that toward Gaza during the first week (see Figure 2). In the headline analysis, however, both Israel and the IDF registered lower sympathy proportions than Gaza and Palestinian people from the outset (see Figure 4, week 1).

In principle, one might expect sympathetic framing to follow acute suffering as it emerges in real time. The absence of a corresponding early surge of sympathy toward Israel in the headline data during the first week following October 7, 2023 therefore suggests that headline framing did not reflect the initial symmetry present in the article bodies, nor did it proportionately register Israeli suffering at that early stage.

Civilian Emphasis and Headline Framing:

During the first week of reporting, sympathetic framing toward Israeli and Palestinian civilians in the main text appeared at comparable levels, with both categories registering proportions close to 0.7 (see Figure 5). In headlines during that same week, sympathy toward Israeli civilians was marginally higher, with Palestinian civilians and Gaza closely following. This brief period of symmetry, however, was confined to the initial phase of coverage. Beyond the first week, sympathetic framing toward Palestinian civilians in the main texts remained consistently high, never falling below 0.5, while sympathy toward Israeli civilians declined and remained lower for the remainder of the initial reporting time-frame. Although headlines contained fewer sympathetic occurrences overall than article bodies, they exhibited a disproportionate concentration of sympathy toward Gaza and Pal-estinian civilians, as reflected in the calculated ratios.