FrIdéo: French News Outlet Ideology Scoring via Multi-Source Evidence Fusion and Distancing-Aware Lexical Scoring

Amr Sobhy

Accepted paper · ICNLSP 2026·Peer-reviewed

FrIdéo

A continuous ideology scale for thirty French news outlets, built by asking nine independent sources the same question and reporting where they agree and where they do not.

By Amr Sobhy

Oral presentation · Trento, September 2026

FrIdéo places thirty outlets across print, broadcast and digital-native formats on one continuous scale, each with a position, a confidence interval, and every input published. Editorial ideology has no ground truth, so the scale is not offered as a verdict on any newsroom: it reports what nine unrelated bodies of evidence say when asked the same question, including the places where they say different things.

Listen to the summaryLength 1:31

0:00 / −1:31

Year: 2026
Venue: ICNLSP 2026; Oral presentation · Trento, September 2026
Research area: French media, Political communication, Computational journalism, Media ideology

Read the paperExplore the scale See the method

At a glance

national news outlets, print to digital-native

independent sources of evidence per outlet

0.84

average correlation between sources sharing no data

outlets whose interval includes zero: unresolved

The results

The scale

Thirty outlets, from L’Humanité (−2.41) to Valeurs actuelles (+1.87).

Select any outlet to see its score, its confidence interval, and how much evidence sits behind it.

Point size = number of independent sources (4–8)Hatched zone = intervals here cross zero; the scale does not resolve these outlets against each other

Points are the estimated position; whiskers are 95% confidence intervals. Larger points have more independent evidence behind them. Band shading marks the seven descriptive labels. Those exist to make the chart readable, and no analysis in the paper uses them.

Each of these thirty outlets now has a page of its own: the score, the interval, what each of the nine evidence families said, and what should stop you over-reading it.

Browse all thirty outlets →

30 outlets →

Who owns it28/30
Who it links to30/30
What readers think8/30
What it says it is18/30
How it is described27/30
Who it puts on air3/30
Media Bias/Fact Check rating15/30
Ad Fontes Media rating9/30
What it actually publishes30/30

Has dataNo data

168 / 270

Each column is one outlet, each row one source

01 / 05

Nine witnesses, none of them complete

The scale asks the same question of nine unrelated sources — among them an ownership registry, a map of who links to whom, a survey of readers, editorial charters, two rating panels and half a million articles.

Not one of them covers the whole panel. Broadcast guest bookings reach three outlets; reader surveys reach eight. Only the hyperlink network and the outlets’ own articles reach all thirty. Of the 270 cells in this grid, 168 have data.

That patchiness is the honest part. A method that reported a clean number for every outlet would be hiding which ones it barely knows.

30 outlets →

Who owns it28/30
Who it links to30/30
What readers think8/30
What it says it is18/30
How it is described27/30
Who it puts on air3/30
Media Bias/Fact Check rating15/30
Ad Fontes Media rating9/30
What it actually publishes30/30

Has dataNo data

168 / 270

Each column is one outlet, each row one source

02 / 05

Whose words are these?

Counting loaded vocabulary naively produces nonsense. L’Humanité uses far-right phrases constantly, in order to attack the people who mean them. Score the words alone and the Communist daily comes out right-wing.

So for every occurrence, the text witness checks whether the outlet marked the phrase as somebody else’s: guillemets, a distancing prefix like la théorie du, explicit criticism in the same sentence. The rule runs symmetrically: « idéologie woke » in a right-leaning paper is marked the same way.

le « grand remplacement » théorisé par Renaud Camus

Marked as someone else’sIn guillemets, and attributed to the man who coined it

la théorie du grand remplacement

Marked as someone else’s“the theory of” is a distancing prefix

le supposé racisme systémique

Marked as someone else’sThe mirror case: a left-coded phrase marked by a right-leaning outlet

lutter contre l’immigration massive

The outlet’s own wordsNo marking. The outlet is using the phrase, not reporting it

Four cases, as the rule sees them

03 / 05

More evidence, less inheritance

Some outlets are exhaustively documented. Others barely exist in the record. Rather than pretend we know as much about Blast as about Le Monde, each score is pulled toward what the outlet was founded to be, and how hard depends on how much direct evidence there is.

The weight is α = n / (n + 2). Two sources and the score is half evidence, half founding tradition. Seven and it is about three-quarters evidence. The outlets with the least behind them also carry the widest error bars.

Across the panel, coverage runs from four sources to eight.

n = 40.67

n = 50.71

n = 60.75

n = 70.78

n = 80.80

Direct evidenceFounding tradition

Each dot is one outlet, at its real source count

04 / 05

They agree, and they share no data

Here is the result the whole method rests on. Take the six best-covered witnesses and correlate every pair of them: fifteen pairs, running from 0.71 to 0.94, averaging 0.84.

These sources were built by different people, in different eras, for purposes that had nothing to do with placing outlets on a scale. They agree anyway.

1Who owns it
2Who it links to.75
3What it says it is.87.84
4How it is described.94.80.88
5Media Bias/Fact Check rating.86.71.83.87
6What it actually publishes.90.77.81.92.88
123456

0.710.94

Correlationmean 0.84

Pairwise correlation between the six best-covered sources

05 / 05

The JDD, before and after

Eight of the nine witnesses are structural and update slowly; one reads the outlets’ own recent articles. That is what lets the scale register a newsroom that changes.

On 22 June 2023 the Journal du Dimanche was handed to Geoffroy Lejeune, until that week the editor of the far-right Valeurs actuelles. The newsroom struck for forty days. Dozens of journalists left.

Running the same scorer, with the same rules, on three windows of the JDD’s own articles: −0.16 before, +0.30 immediately after, +0.57 a year later. What moved was not whether the paper wrote about immigration. It was whether « immigration massive » sat inside quotation marks or not.

The JDD is one outlet, examined because the date of the handover was known. It shows the scale can register an editorial change; it says nothing about how often such changes happen.

Jan 2022 – Jun 2023before

−0.16

51213,251 articles371

Aug – Dec 2023immediately after

+0.30

592,962 articles109

Jan – Jun 2024about a year later

+0.57

674,800 articles247

Text score

Same scorer, same rules; only the paper changed

Background

What existed before

How far apart are Le Point and L’Express? Is La Croix further from the centre than Les Echos? Where does Ouest-France actually sit? Each of these questions has a partial answer in the existing literature, and the partial answers do not compose. They cover different outlets, in different years, on different scales, and several were built to answer something else.

No single source settles it

Editorial ideology has no ground truth. Nobody can hand over the true position of Le Monde: there is no experiment that settles it, no register to consult, and no number the paper itself could supply even if it wanted to. Outlets resist self-categorisation, readers conflate what an outlet is with what they think it is, and a content analysis deep enough to settle the question is too expensive to run across thirty newsrooms.

Every method therefore picks a proxy: the readers, the link structure, a panel of experts. Each is a real signal with a known failure. Reader partisanship measures the audience rather than the newsroom. A link network captures an eighteen-month window ending in 2019 and then goes stale. An expert panel inherits the expertise, and the blind spots, of whoever sits on it.

The alternative is not a better proxy. It is more of them. If nine sources built by different people, using different methods, for unrelated purposes all place Valeurs actuelles to the right of Le Monde, that agreement carries more weight than any one of them being clever. Where they disagree, the disagreement is itself information: it indicates that the question is open at that outlet, which is what an error bar is for.

“We do not claim to have recovered the true position of any outlet. We claim a scale that is internally consistent, robust to modelling choices, and convergent with independent external evidence.”
FrIdéo, §6

The evidence

Nine witnesses

Nine sources, each answering the same question about the same thirty outlets. What matters is how little they have in common.

An ownership registry, a survey of readers, a map of hyperlinks and half a million articles have nothing to do with each other. They were assembled by different people, in different decades, for purposes that had nothing to do with placing outlets on a scale. That independence is not incidental. It is the entire reason their agreement means anything.

How nine become one

The nine are put on a common footing and averaged, but not blindly. Some outlets are exhaustively documented; others barely exist in the record. Rather than pretend we know as much about Blast as about Le Monde, the more independent evidence an outlet has, the more its score is driven by that evidence; the less it has, the more it falls back on what the outlet was founded to be, and the wider its error bar becomes.

Le Monde has seven sources behind it. Franceinfo, the best-covered outlet on the panel, has eight. Blast has four. The arithmetic that turns those counts into a weighting is set out below.

No single source is load-bearing. Drop any one of the nine, rebuild the entire scale from scratch, and the ordering barely moves: the rank correlation with the full model never falls below 0.992.

The model, written out

The whole measurement is five lines. Each family is standardised on its own, the available families are averaged, that average is blended with the founding-tradition prior according to how much evidence the outlet has, and the result is re-standardised across the panel.

z_ik = x_ik − x̄_ks_kEach family k is standardised across only the outlets it covers. Missing values are left missing rather than imputed as zero, which would pull uncovered outlets toward the centre.
z̄_i^dir = 1n_iΣz_ikThe mean of whichever families cover outlet i. Unweighted, for the reason given below.
α_i = n_in_i + λ(λ = 2)How far the score follows the direct evidence. λ = 2 is a design choice, not an estimate: two sources should not override the historical record unaided.
z_i^bl = α_i z̄_i^dir + (1 − α_i) z_i^strThe blend of direct evidence and the structural prior, which encodes founding tradition on a five-point grid scaled by 1.5.
final_i = z_i^bl − μ_Bσ_BRe-standardised across all 30 outlets, giving a zero-mean, unit-variance scale. This is the published score.

2 sources

α = 0.50

4 sources

α = 0.67

7 sources

α = 0.78

What λ = 2 means in practice. An outlet with two sources is placed half by evidence and half by its founding tradition; one with seven is placed almost entirely by evidence. Outlets with no direct evidence fall back on the prior entirely.

The direct mean is deliberately unweighted rather than weighted by each family’s precision. Family dispersion is largest for the sources that diverge from the consensus, notably the article-text family, so precision weighting would quietly suppress the one contemporary signal the model is built to keep visible.

The core finding

How much the sources agree

If the nine witnesses were each measuring their own thing, readership here and ownership there, their placements would scatter. They do not scatter, and the degree to which they do not is the central result.

1
The sources track each other
Across every pair of the six best-covered witnesses, the correlation runs from 0.71 to 0.94, averaging 0.84. Ownership registries, a 2018–2019 hyperlink network, a reader survey and half a million articles agree with each other about where thirty newsrooms sit.
2
There is one dimension here, not several
A standard test for whether several measurements are really tracking one underlying thing finds that a single dimension accounts for 88.5% of what the sources share; the next accounts for 7.4%. We also went looking for the most likely second dimension: La Croix is right-of-centre through Catholic tradition, Les Echos through economic liberalism, so a distinct cultural-versus-economic axis should put them at opposite ends. They differ by 0.35 of a standard deviation. If that axis is there, this evidence does not find it.
3
It separates groups it was never shown
We removed the Media Bias/Fact Check family entirely, rebuilt the scale without it, and asked whether the outlets MBFC calls left and right separate on the rebuilt scale. They separate without overlap: every right-group outlet scores above every left-group outlet, an ordering that would arise by chance about once in 1,700 times. Repeating the test with only the five families that share no method with MBFC leaves the agreement essentially unchanged, so it is not an artefact of shared method.

Per-outlet source spread

Each dot is one independent source placing this outlet. The diamond is the published score.

Who owns it−1.06
Who it links to−0.18
What readers think−0.06
How it is described−0.78
Media Bias/Fact Check rating−0.84
Ad Fontes Media rating−0.84
What it actually publishes−0.34
Published score−0.71

7 sources have data for this outlet. They span 0.99 of a standard deviation.

One of the best-evidenced outlets on the panel: seven of the nine sources have data on it, and no single one of them is doing the work. This is what the typical case looks like.

What this does not establish

None of this shows the scale is correct, because there is nothing for it to be correct against. It shows that the scale is internally consistent, stable under every perturbation tested, and pointing in the same direction as evidence it never saw.

Read this before quoting

How to misread this

Five readings this scale does not support, each with the statement the evidence does support.

Le Monde is centre-left, according to science.
On a thirty-outlet scale where zero is the sector average, Le Monde sits at −0.71.
The label is a reading aid. The measurement is the number, and the number is relative to this specific set of newsrooms.
Le Parisien leans further right than BFMTV.
Both sit in the unresolved centre; the scale does not separate them.
Their intervals overlap heavily, and a difference smaller than the error bar is not a difference. Where an outlet is being described in print, the interval is the number to quote rather than the point estimate.
Atlantico is further right than CNews.
Atlantico’s estimate is +1.08, but it is one of five outlets with thin evidence.
About a third of Atlantico’s score comes from its founding profile rather than direct measurement, and its interval is correspondingly wide. Sparse outlets are the wrong place to draw fine distinctions.
Le Figaro scores +1.02, roughly the same as some German or American outlet.
Nothing. There is no valid version of this sentence.
The scale is defined by this panel. Add or remove outlets and every number changes. These scores are not comparable to any other country, rating system, or outlet set.
This shows which outlets are trustworthy.
This shows where outlets sit on a left–right axis.
Position is not quality. A far-left outlet and a far-right outlet can both be scrupulously accurate; a centrist one can be sloppy. This says nothing about accuracy, rigour, or good faith.

Questions

Citation

If you use the FrIdéo scores or the replication code, please cite the current version.

@inproceedings{sobhy2026frideo,
  title     = {{FrIdéo}: French News Outlet Ideology Scoring via Multi-Source
               Evidence Fusion and Distancing-Aware Lexical Scoring},
  author    = {Sobhy, Amr},
  booktitle = {Proceedings of ICNLSP 2026},
  year      = {2026}
}

Accepted for oral presentation at ICNLSP 2026, Trento. The citation will be updated when final publication details are available.

Open data

The scores are also published on data.gouv.fr, the French government open data platform, and on Kaggle, under CC BY 4.0: CSV and JSON files, data dictionary, and full metadata.

Kaggle

This work was supported by

Contact

For research questions, collaborations, or media inquiries.

Amr Sobhy

Keep reading

More from the lab

See all publications →