Political Spectrum Ratings on Individual Articles
Now appearing on L.A. Times Insights
Up until a couple of decades ago, people mostly read written journalism by flipping through a newspaper where they would see pages clearly marked “news” and “opinion.” Now, you may wind up reading a piece of content a friend posted on social media or texted to you without knowing context about the author or publication. Unbridled access to information is empowering, but it can also make it harder to separate news reporting from opinion pieces.
Readers benefit from publications clearly demarcating content that is based on opinions versus news reporting. In a polarized environment where much of U.S. news and culture is intertwined with politics, it’s especially helpful for readers to quickly and easily understand where opinion pieces fall on the political spectrum. To overcome polarization, researchers say we must get out of our own echo chambers to understand different points of view. News readers can only be exposed to a variety of opinions if they know where the opinions are. Also, sometimes readers might be surprised that an author who they assume is liberal or conservative actually has a different view on the topic they are writing about.
Particle analyzes written pieces and classifies where along the American political spectrum they fall. Particle uses AI to classify every sentence in a text as either verifiable or interpretive facts versus opinions. If a statement is even remotely political in nature, it is evaluated along multiple dimensions to produce a left, right, or center rating — or a “not political” rating, if there are not enough references that are political in nature.
If you already use our app, you might be familiar with the political spectrum ratings for publications that contributed reporting to a story. This in-app barometer shows how far left or right the publications themselves fall, based on a collection of ratings from nonpartisan organizations.
But this new tool is different, because it rates individual opinion pieces on a left-to-right political spectrum and produces short summaries of the composite ratings for pieces, as well as in-depth breakdowns that include evidence for each score.
How the analysis works
The tool doesn’t merely provide a single overall assessment by considering the text in its entirety– it considers multiple dimensions at a sentence-by-sentence level. By evaluating each thought granularly in isolation and then calculating a composite score based on many specific scores within the dimensions, the overall score is more likely to converge on a consensus. In other words, the multidimensional approach serves as a checks-and-balances system that leads to a more robust overall score.
These are some of the dimensions considered:
- Topic selection and framing: What is discussed and how it is contextualized
- Policy positions: How things should be changed
- Language and rhetoric: How things are expressed
- Source treatment: Who is considered authoritative
- Moral judgments: Why things matter
Each dimension is measured based on a series of markers: The set of markers associated with each classification level (from far left to far right) dictate the criteria for how a given excerpt is classified. For example, a marker for an excerpt determined as far-left based on topic selection is: “Explicitly challenges the legitimacy of the current economic/political system.” And a marker in the same dimension for an excerpt determined as far-right based on a topic selection is: “Rejects legitimacy of most government intervention.”
Collecting Evidence: Through identifying markers that place excerpts within a specific classifications, the tool tries to identify, for example, “What specific changes are proposed?” (for policy positions), and “What values are prioritized?” (for moral judgments).
Dimensional Composite Rating: Based on the established markers and answers to the questions, every sentence is rated along every dimension based on a -3 to +3 scale, corresponding with far left, strong left, moderate left, center, moderate right, strong right, and far right. Then, the results are combined in a composite rating for each dimension.
Measuring Certainty: Once all the dimensions have a rating, there are validation and quality control measures. The composite score also receives a confidence level rating (high, medium, low) and an evidence directness rating.
Centrism vs Balance: Scores away from zero map directly to political classifications on the left or right, while near-zero scores are distinguished as either centrist (if the points are all clustered toward the middle of the spectrum) or balanced (if the points are evenly dispersed between right and left).
Humans define the framework: We (humans) provide the methodology within which the LLM does its analysis. The LLM does not make political judgement on its own — it needs to provide its analysis and reasoning in a structured and systematic way that we govern according to the framework. For each analysis on every dimension, the LLMs provide a transparent rationale that we can use to verify the correctness of the outcomes against our framework and make adjustments as needed.
Below are some examples of the tool at work analyzing political opinion pieces from both the right and the left, as well as some that are centrist and balanced.
1. Far left rating example
This piece called “The cops work for Jeff Bezos” was written and published by J.P. Hill on his Substack, New Means. He describes the goal of his newsletter as to “bring thoughtful commentary, ties to history and movements, and synthesize it with other dimensions like the religious, cultural, and social dynamics running through [U.S.].”
This piece received a rating of far left. In the output below, you can see the overall composite score based on the ratings within each measured dimension, the classification, the confidence level, the minimum and maximum spectrum scores from throughout the piece, and a summary of the findings.
"composite_score": -2.8529646697388635,
"classification": "Far Left",
"confidence": "high",
"min": -3,
"max": -2.5806451612903225,
"summary": "The piece argues that the police serve the interests of the
capitalist ruling class, highlighting how they protect the wealthy while
suppressing workers, such as striking Amazon employees. It calls for
radical changes like abolishing policing and redistributing wealth to
build a more just society. By emphasizing class conflict and advocating
for revolutionary transformation, the piece reflects leftist perspectives."In the results, we can look through excerpts that the tool used as evidence for the classification of each dimension. There are also specific markers the tool looks for and cites, as well as its reasoning for a classification of that excerpt.
The tool cites this excerpt within the language dimension classification:
"excerpt": "But the police are siding with the ruling class lawbreakers,
using violence to enforce the will of the capitalist class, and
revealing their true role to us in the process.",
"tier": "direct_statement",
"markers": [
"Uses explicit class-conflict terminology",
"Uses language of systematic exploitation/oppression"
],
"reasoning": "The author employs explicit Marxist class conflict
terminology like \"ruling class\" and \"capitalist class\"",
"scale_level_classification": -3,
"classification": "Far Left",
"confidence": "high"Among other aspects, the tool recognizes Marxist terminology — without the author quoting or referencing Karl Marx outright. It is able to recognize political language even if a piece of content is less explicit in its policy positions.
At the end of the analysis of a dimension, the full dimension receives a composite score based on scores determined by excerpts and an explanation summarizing why it arrived at its classification:
"summary": "The author consistently demonstrates a far-left (-3)
perspective on source treatment and institutional authority.
They explicitly reject and delegitimize traditional institutional
authority (particularly police) while elevating worker/activist
perspectives and experiences. The author frames institutional
power as inherently corrupted by capitalist interests and consistently
validates revolutionary/radical interpretations of events over
official narratives.",
"composite_score": -3,
"classification": "Far Left",
"min": -3,
"max": -3,
"confidence": "high"2. Far right rating example
This piece called “The Chaos Factor” was written and published by Kira Davis on her Substack, Just Kira Davis. She describes her newsletter as sharing her opinions about “culture, politics, and faith.”
This content received a far right classification with the following summary:
"composite_score": 2.7915107547409956,
"classification": "Far Right",
"confidence": "high",
"min": 2.0864197530864197,
"max": 3,
"summary": "The piece portrays immigration as a source of crime and chaos,
citing examples like armed gangs taking over apartments and violent
attacks by illegal immigrants. It uses alarmist language to frame
current events as part of an intentional plot to destabilize America,
urging citizens to distrust mainstream media (\"foul legacy media\")
and government institutions. By emphasizing individual action over
government intervention and expressing deep skepticism of authority,
the piece reflects right-wing positions."This is an example of a rating within the “source treatment” dimension of Davis’ piece:
"excerpt": "Many of the people you talk to don't even know the other
side of the equation, having been dependent on the foul legacy media
for all their information.",
"tier": "direct_statement",
"markers": [
"Treats mainstream institutions as compromised",
"Rejects conventional academic authority"
],
"reasoning": "The author explicitly dismisses mainstream media as
'foul' and positions it as an unreliable source of information,
suggesting alternative sources are more trustworthy",
"scale_level_classification": 3,
"classification": "Far Right",
"confidence": "high"The tool recognizes that the excerpt is a direct statement Davis makes about a source of information — “the foul legacy media.” That wording triggers markers associated with the far-right classification for source treatment that we set: “treats mainstream institutions as compromised” and “rejects conventional academic authority.”
3. Balanced vs. Centrist example
The first two examples were clearly on the right and left. But, as mentioned, the tool can distinguish near-zero scores as either centrist (if the points are all clustered toward the middle of the spectrum) or balanced (if the points are evenly dispersed between right and left).
This piece titled “Trump Is Quickly Being Reminded He Must Work With Democrats” written by Josh Barrow for his Substack, Very Serious, gets a centrist rating. Its composite score is near zero (-0.33), which could be a score for a centrist or balanced piece. Here, the tool identifies subtleties showing Barrow’s piece falls in the middle of the spectrum, rather than points out arguments from both sides. Here’s the summary:
"composite_score": 0.06111111111111111,
"classification": "Centrist",
"confidence": "high",
"min": -0.33333333333333337,
"max": 1,
"summary": "The piece presents a centrist perspective by focusing
on practical governance and bipartisan solutions, such as proposing
debt ceiling reform to facilitate legislative progress. It employs
neutral and analytical language centered on legislative mechanics
and political realities, emphasizing the need for cooperation across
party lines. By highlighting procedural challenges and advocating for
workable solutions without ideological bias, the writing reflects a
centrist approach to political analysis."On the other hand, a piece titled “Too many Americans still fear the future” by Noah Smith on his Substack, Noahopinion, is rated balanced. The piece is about how trends like increased housing construction and a surge in entrepreneurship can be an antidote to Americans’ apprehension toward change.
This piece, while not seeming overtly political based on the headline, deals with political trends and topics. The political spectrum rating tool recognizes that and finds the opinions “balanced” with the following summary:
"composite_score": 0.09211595448470444,
"classification": "Balanced",
"confidence": "high",
"min": -1.0714285714285714,
"max": 1,
"summary": "The piece presents a balanced perspective, emphasizing
practical, market-based outcomes over ideological positions. It
critiques both Republicans and Democrats for resisting technological
advancements—Republicans for opposing renewable energy and electric
vehicles due to cultural biases, and Democrats for attempting to
regulate the tech industry and AI out of concern for corporate
power—highlighting how entrenched interests hinder progress.
By advocating for pragmatic solutions and collective progress while
acknowledging the need for balanced government involvement, the piece
reflects centrist views with both center-left and center-right elements."It points out and responds to both parties’ takes on advancement. This is an example of the tool identifying a “moderate right” policy perspective, based on the following excerpt:
"excerpt": "Those restrictions were, in my assessment, pretty pointless,
especially because China would have just proceeded with largely
unfettered AI research anyway — as it ended up doing with DeepSeek.",
"tier": "direct_statement",
"markers": [
"Prioritizes practical over theoretical knowledge",
"Validates truth through market/practical success"
]This is the summary of that excerpt’s moderate right rating:
"reasoning": "The author directly opposes AI restrictions and regulation,
favoring a more market-driven approach to AI development, consistent
with moderate right policy positions",
"scale_level_classification": 1,
"classification": "Moderate Right",
"confidence": "high"The tool also provides examples of Smith taking a “moderate left” policy perspective, based on this excerpt:
"excerpt": "But applying the pause to renewables broadly is an
escalation — pausing solar energy action as well.",
"tier": "pattern_evidence",
"markers": [
"Promotes environmental protection",
"Supports expanded social programs"
]This is the summary of that excerpt’s moderate left rating:
"reasoning": "The author implies opposition to pausing renewable
energy development, suggesting support for government enabling of
clean energy, aligning with moderate left policy preferences",
"scale_level_classification": -1,
"classification": "Moderate Left",
"confidence": "medium"Questions about Particle’s political rating tool? Visit our website to get in touch. And to understand more about the news faster, download the Particle app on iOS devices.
