What Hope Means to the British Public: An Interactive Discourse Map
DDI United Kingdom · Audience response study
Part 2 · InteractiveLet’s Bring Back Hope
The Discourse Map
An interactive map of 12,139 public replies to Prime Minister Andy Burnham’s “What Hope Means” video, each scored on the twenty indicators of the Democracy Discourse Index. Built for students, researchers and policy readers to explore, test and teach with.
Part 2 of Let’s Bring Back Hope is one study published in three parts. All three draw on the same corpus of scored comments, and they are meant to be read together.
- ReportWhat Hope Means to the PublicThe analytical report: findings, interpretation, and implications for policy and leadership.
- Interactive mapThe Discourse Map You are hereEvery scored comment on the map, with a guided reading, a comparison tool and seminar exercises.
- Lexical appendixThe Vocabulary of the ReplyThe words the public used, by platform, reception profile and DDI band.
Part 1: “Let’s Bring Back Hope” in 48 Hours, with its interactive discourse map and lexical appendix.
GCRD’s vision is a democracy in which citizens can engage with renewed hope. Hopelessness is a critical democratic vulnerability, and hope is a democratic force multiplier. When a Prime Minister asks the country what hope means, the public reply shows where the conditions for democratic hope are present and where they are missing: whether people recognise one another, disagree without contempt, extend trust and speak as citizens who can act.
This map lets you see that reply comment by comment. A position on the map describes the language of a single comment as read by the scoring model. It is not a judgement of the person who wrote it, and it is not a diagnosis of the discourse as a whole.
Where each comment sits
Choose two dimensions, colour or shade the map by any lens, filter it, and select a position or a corner to read it in the panel.
Map
Show this view as a table
Hover or tap a bubble to read it, and select it to load it in the panel. With the map focused, arrow keys move between positions and Enter selects.
Set two groups side by side
Pick any two groups of comments. The comparison shows their means and distributions, the four dimensions, the twenty indicators and the profile mix, and can hold comment length constant.
Group A ·
Group B ·
Holding length constant reweights each group to the corpus mix of comment lengths across six bands (direct standardisation). It answers one objection, that some groups simply write more; it does not remove the other differences between them.
Readings built on the scores
Each lens groups the same DDI scores in a different way, and none of them changes a score. Every row can be sent to the map.
How to read a position
Each bubble is a position on two of the four DDI dimensions, and its area is the number of comments that share it. A dimension score is the sum of five indicators, each scored on the same five-point anchor scale, so every dimension axis runs from 0 to 20.
Colour shows the comment’s DDI band, from its composite across all four dimensions, or any other single-valued lens. The composite is 100 × the sum of the applicable dimension scores ÷ (20 × the number of applicable dimensions), rounded. Bands: Risk below 25 (9,679 comments), Concerning 25–49 (2,093), Mixed 50–74 (347) and Healthy 75–100 (20). Band labels classify model-assigned scores. “Risk” is not a judgement that a commenter is dangerous, and a low score can mean the absence of a positive quality rather than the presence of hostility.
A comment is left off a view only when one of the two selected dimensions was judged inapplicable to it. It stays in every other view. Inapplicable is not zero.
The four dimensions
The GCRD working paper on democratic communication diagnosis treats the four dimensions as jointly necessary conditions: “trust opens the communicative door; empathy determines who is recognised through it; civility keeps the door open during disagreement; agency gives people a reason to enter.” Strength in one does not make up for weakness in another, which is why the map sets two dimensions against each other instead of averaging them.
| Dimension | What a high score records | Indicators | Mean /20 |
|---|---|---|---|
| Empathy | Others are recognised as people whose experience matters | Perspective-taking, emotional recognition, humanisation, solidarity, anti-hostility | 2.51 |
| Civility | Disagreement is sustained without abuse, dehumanisation or exclusion | Respectful tone, non-dehumanisation, constructive disagreement, de-escalation, generous interpretation | 3.94 |
| Trust language | Claims and motives remain open to shared, good-faith evaluation | Institutional trust, good faith attribution, interpersonal trust, epistemic trust, anti-conspiracy | 2.85 |
| Democratic agency | Citizens are represented as able to act with democratic consequence | Political efficacy, mobilisation, collective action, rights consciousness, democratic optimism | 2.37 |
Zones and corners
To describe regions of the map, we borrow the working paper’s provisional diagnostic thresholds of 0.40 and 0.60 of the scale maximum. In the paper they classify topic-level estimates together with their uncertainty. Here they are applied to single comments, and only as a reading aid.
On the two summary scales, restraint and generation, which run from 0 to 16, the zones are 0–6, 7–9 and 10–16. Any view then has four corners. The lower left holds comments that express neither condition; brief reactions, slogans and hostile remarks can all land there, and the map alone cannot tell them apart. The upper right holds comments that express both. The two off-diagonal corners carry the most interpretive information, because they show one democratic condition present while another is missing.
Worked example: empathy and agency
Set the horizontal axis to empathy and the vertical axis to democratic agency.
Shares are of comments placed in this view; 26.4% score exactly zero on both axes. The rest fall in the middle zone on one or both axes.
A comment in the lower-right corner scores well on perspective-taking, emotional recognition, humanisation or solidarity: it registers that other people’s hardship, dignity or point of view matter. It scores poorly on efficacy, mobilisation, collective framing, rights and democratic optimism: it does not present citizens, the writer included, as able to do anything about what they recognise. There is concern without a route to consequence. The working paper calls the matching topic-level profiles “receptive but passive” and “compassionate but fatalistic”.
It does not mean the writer is apathetic; low agency language can reflect a reasoned judgement that institutions have not responded before. It does not measure support for, or opposition to, the Prime Minister or his message. And it describes one comment, not the writer’s wider views or behaviour.
Under the paper’s constraint-first logic, where empathy is present and agency is the binding constraint, the useful response is to connect concern to credible routes for action: concrete choices, realistic ways to take part, timelines and visible feedback on what changed. A more emotive message would add to the condition that is already present and leave the missing one untouched. The mirror corner calls for the opposite care: an environment of agency without recognition “may be capable of rapid exclusionary mobilisation”, and needs recognition and de-escalation before any activation.
Every pairing
The same logic applies to each of the six pairings of the four dimensions, and to the two summary scales. Names follow the working paper’s configuration vocabulary where a close match exists. They describe language patterns, not types of people.
What this corpus looks like on the map
- The mass sits at the origin. In every pairing of the four dimensions, 80–88% of placed comments fall in the lower zone on both axes. In the empathy and agency view, 26.4% score exactly zero on both.
- 78.9% of comments are in the lower zone on all four dimensions (9,555 of the 12,111 with all four applicable).
- Civility is the most expressed dimension (mean 3.94 of 20), so views with civility on an axis stretch furthest from the origin. Civil spectatorship (1.65%) is the largest off-diagonal corner in any pairing of the four dimensions.
- Democratic agency is the least expressed (mean 2.37). Bubbles rarely rise far up any view with agency on the vertical axis.
- Platforms differ in degree, not in shape. Instagram (mean composite 24.5) pulls bubbles further from the origin than X (11.9), but every platform keeps its mass near the origin.
- Replies sit further out than top-level comments (mean 17.3 against 14.1).
On point estimates, all four dimension means fall well below 0.40 of the scale. The working paper would assign a topic-level configuration only from interval estimates, account- and thread-aware aggregation and an evidence-adequacy rating, which this study does not produce. The map therefore assigns no configuration to the corpus. It describes where comments sit; it does not diagnose the discourse.
Teaching with the map
The map was built to be used in seminars on democratic theory, political communication, computational social science and public policy. Each exercise below loads its own view, so a class can work through it on laptops or phones without setting anything up. Allow ten minutes for the guided reading before the first exercise.
Seminar exercises
1Reading a skewed distributionMeasurement and description · 20 minutesOpenClose
- Load the view: empathy against agency, with the origin selected and the Comments tab open.
- Note the mean, the median and the Risk share for the whole corpus, then for the origin.
- Read four comments at the origin. For each, decide whether its zero records hostility or simply absence. Check anti-hostility in the Indicators tab.
Discuss
- Why does the median (10) sit below the mean (14.7)? What would a report that gave only the mean hide?
- Is a comment that scores zero on empathy hostile, or silent about other people? How would you tell from the scores alone?
- How should a report describe a corpus in which 12% of comments score zero on every indicator?
2Civility is not democratic qualityRestraint and generation · 20 minutesOpenClose
- Load restraint against generation, coloured by the generative ladder.
- Find the comments that show restraint and nothing else. Compare the restraint and generative averages in the panel.
- In the Lenses section, open Platforms, length and timing and compare the two scores by platform.
Discuss
- What is the difference between not doing harm and doing good in public speech?
- Why might a public be civil and still closed? What does Part 1’s phrase “civil but closed” gain from this measurement?
- Which generative indicators are rarest? Why might they be hardest to express in a comment field?
3Composition, confounding and comment lengthCausal reasoning · 25 minutesOpenClose
- Load the comparison of X (A) and Instagram (B) with comment length held constant. Note the gap with and without the adjustment.
- Compare the original exports with the supplementary X retrieval.
- Compare replies with top-level comments.
Discuss
- How much of the platform gap does length account for? What else varies with platform: audience, source post, interface, capture method?
- Adding the full X conversation lowered the pooled mean. Why is that not evidence that the discourse got worse?
- What study design would be needed to estimate a platform effect? Why are comments not independent observations?
4Checking the instrumentValidity and reliability · 30 minutesOpenClose
- Load civility against agency coloured by model confidence, with the civil spectatorship corner selected, and open the Comments tab.
- Score two indicators yourself for each comment, using the anchors in the glossary below. Record where you differ from the model.
- Read the agreement figures in the Method section and in the glossary cards.
Discuss
- Where did you disagree with the model, and why? Did sarcasm, brevity or missing context play a part?
- Interpersonal trust has the lowest project agreement (55.6%). Why might it be hard to score from a single comment?
- In the wider DDI programme, human triple-coding provides the ground truth on which ReHuman codes at scale. What would an independent, blinded validation of this study need?
5From diagnosis to communicationStrategy and ethics · 30 minutesOpenClose
- Load the civil spectatorship corner and read its response in the panel.
- Then load the mirror corner, committed but corrosive.
- Use the pairing table in the reading guide to compare responses across corners.
Discuss
- Draft two short public messages from a government channel, one for each corner, following the constraint-first logic. What would a credible route to action look like?
- Why does the working paper warn against awareness-only communication and empty calls to “get involved”?
- Why does this guidance apply to topic-level conditions rather than to replying to individual commenters? What are the risks of using discourse scores to steer messaging?
6Signals, suspicion and a poisoned wellInformation integrity · 35 minutesOpenClose
- Open the Integrity signals lens and read the method note.
- Load the map filtered to repeated texts and read the comments.
- Open The poisoned well lens. Note who is accused of being a bot, and where the accusations gather.
- Load the accusations on the map and read what the accused said back.
Discuss
- What can repeated text tell you, and what can it not? List three innocent explanations and two coordinated ones.
- None of the linked accusations answers a comment with any sign of scripting. What does that tell you about how suspicion travels? Who benefits when citizens dismiss each other as bots?
- The liar’s dividend lets the guilty dismiss real evidence as fake. How does it become a tax on sincere citizens?
- What evidence beyond this dataset would be needed to attribute a coordinated campaign? How should an analyst write about machine labels such as “suspected misinformation”?
Questions for policy scholars
- The public redefined hope as conditions: an election, the borders, the bills, the services. What would it take to make hope checkable in government communication, and who should do the checking?
- Constructive comments petition and propose from particular circumstances. What institutional protocol could acknowledge, route and report back on them at scale?
- The most common redefinition of hope was a general election. How should a government respond to a demand for consent that is hostile to the messenger and faithful to the mechanism?
- X is the hardest ground and the platform where opinion leaders concentrate. Should communication strategy differ by platform, and what are the democratic costs if it does?
- What safeguards would responsible use of a discourse index by governments require: independence, validation, transparency, limits on targeting?
The twenty indicators
Each card gives the indicator’s guiding question and the anchors at 0, 2 and 4 from the DDI Methodology and Rubric Guide, Second Edition (condensed). Scores of 1 and 3 sit between the anchors. The footer gives the corpus mean, the share scoring 2 or more, and the project-specific agreement between ReHuman and 19 human-reviewed comments. R marks the four restraint indicators.
Empathy to indifference. Does the comment treat other people as human beings: the people it is about, and the other citizens in the thread?
Corpus mean 2.51 out of 20Does the comment show awareness of how the issue looks from another person’s standpoint, especially those affected differently?
- 0
- Self-centred framing; other views dismissed or ignored.
- 2
- Some acknowledgement of other views, brief or quickly undercut.
- 4
- Generous, sustained engagement with how others experience the issue.
Corpus mean 0.11 of 4 · 2.8% score 2 or more · model agreement with reviewed scores 88.9%
Does it acknowledge the feelings of those affected, such as fear, grief, frustration or hope, as legitimate?
- 0
- Others’ emotions mocked, pathologised or invisible.
- 2
- Instrumental or conditional acknowledgement: “I understand you’re upset, but …”.
- 4
- Sincere acknowledgement of feelings as worthy of response in their own right.
Corpus mean 0.09 of 4 · 2.3% score 2 or more · model agreement with reviewed scores 83.3%
Does it refer to people, especially outgroups and opponents, as full human beings?
- 0
- Dehumanising metaphors, or mass nouns that erase individuality.
- 2
- Groups treated as a uniform, undifferentiated mass.
- 4
- People treated as individuals with stories and full humanity.
Corpus mean 0.66 of 4 · 21.0% score 2 or more · model agreement with reviewed scores 77.8%
Does it express active solidarity, not just sympathy, with those who are vulnerable or most affected?
- 0
- No concern for those affected, or hostility toward them.
- 2
- Affected groups mentioned as an abstraction.
- 4
- Stands with those most affected and commits to their welfare.
Corpus mean 0.47 of 4 · 12.6% score 2 or more · model agreement with reviewed scores 72.2%
Does it avoid pleasure at others’ suffering, contempt for their pain, or hostility toward people rather than positions?
- 0
- Celebrates harm, or attacks people rather than views.
- 2
- No explicit cruelty, but cold indifference to harm.
- 4
- Entirely free of hostility, even in firm disagreement.
Corpus mean 1.18 of 4 · 32.4% score 2 or more · model agreement with reviewed scores 88.9%
Civility to toxicity. Civility is not politeness: it maintains the conditions that let disagreement continue without dissolving the political community.
Corpus mean 3.94 out of 20Does it disagree with positions without attacking character, intelligence or worth?
- 0
- Insults, mockery and character attacks.
- 2
- Largely on the issues, but veers toward condescension.
- 4
- Consistent respect: critiques the argument, never the person.
Corpus mean 1.18 of 4 · 33.4% score 2 or more · model agreement with reviewed scores 68.4%
Does it treat all participants as legitimate voices with equal standing in the democratic conversation?
- 0
- Treats some as illegitimate; slurs or exclusionary labels.
- 2
- Implies some speakers are inherently suspect.
- 4
- Recognises everyone’s right to hold and express their views.
Corpus mean 1.88 of 4 · 61.3% score 2 or more · model agreement with reviewed scores 79.0%
Does it engage the substance of opposing arguments, even to refute them?
- 0
- Asserts without reasons, or repeats its position more loudly.
- 2
- Engages a weakened version, or engages only superficially.
- 4
- Engages the strongest version with evidence and reasoning.
Corpus mean 0.39 of 4 · 10.2% score 2 or more · model agreement with reviewed scores 84.2%
Does it avoid language designed to inflame, and try to lower the temperature?
- 0
- Deliberately inflammatory, apocalyptic or us-versus-them.
- 2
- Not designed to inflame, but unnecessarily charged.
- 4
- Actively lowers the temperature without abandoning substance.
Corpus mean 0.34 of 4 · 7.8% score 2 or more · model agreement with reviewed scores 94.7%
Does it assume the most charitable plausible motives in those it disagrees with?
- 0
- Attributes malice, greed, stupidity or bad faith.
- 2
- Implies irrationality or self-interest rather than conviction.
- 4
- Attributes disagreement to honest differences in values or evidence.
Corpus mean 0.15 of 4 · 4.7% score 2 or more · model agreement with reviewed scores 84.2%
Trust to cynicism. Not whether institutions deserve trust, but whether the discourse builds or erodes the trust infrastructure democracy depends on.
Corpus mean 2.85 out of 20Does it recognise democratic institutions as having some legitimate authority, even while criticising decisions?
- 0
- Denies the legitimacy of institutions entirely: all rigged or captured.
- 2
- Institutions framed as compromised, without nuance or evidence.
- 4
- Institutions treated as legitimate sites of contestation.
Corpus mean 0.49 of 4 · 14.9% score 2 or more · model agreement with reviewed scores 66.7%
Does it allow that political actors might act in good faith, even if wrongly?
- 0
- Universal bad faith: all politicians are corrupt, all officials lie.
- 2
- Cynical, but stops short of universal bad faith.
- 4
- Distinguishes actors and decisions, and allows sincere motive.
Corpus mean 0.21 of 4 · 7.1% score 2 or more · model agreement with reviewed scores 77.8%
Does it treat fellow citizens, including opponents in the thread, as sincere participants rather than enemies or dupes?
- 0
- Opponents treated as enemies or agents of a hostile agenda.
- 2
- Resigned suspicion of others’ sincerity.
- 4
- Assumes others take part in civic life in good faith.
Corpus mean 0.29 of 4 · 7.7% score 2 or more · model agreement with reviewed scores 55.6%
Does it treat evidence and expertise as legitimate inputs to deliberation, even when disagreeing?
- 0
- Blanket dismissal of evidence, expertise and mainstream sources.
- 2
- Selective: cites what supports it, dismisses the rest.
- 4
- Genuine engagement with evidence, challenged with reasons.
Corpus mean 0.29 of 4 · 7.7% score 2 or more · model agreement with reviewed scores 94.4%
Does it resist explaining events through hidden, coordinated malice?
- 0
- Explicit conspiracy framing: hidden controllers, manufactured crises.
- 2
- Vague, unfalsifiable claims about coordinated bad actors.
- 4
- Explains events through observable actors and interests.
Corpus mean 1.58 of 4 · 52.6% score 2 or more · model agreement with reviewed scores 77.8%
Efficacy to disempowerment. Whether citizens are cast as able to shape outcomes. It is the lowest-scoring dimension in DDI country studies to date.
Corpus mean 2.37 out of 20Does it express belief that citizens can influence political outcomes?
- 0
- Participation is pointless; outcomes are predetermined.
- 2
- Change possible in principle, futile now.
- 4
- Strong confidence that citizens can make a difference.
Corpus mean 0.82 of 4 · 29.5% score 2 or more · model agreement with reviewed scores 68.4%
Does it call citizens to engage: to vote, organise, petition or contact representatives?
- 0
- No call to action, or discourages it.
- 2
- Vague gesture toward the need for change.
- 4
- Clear, specific and achievable calls to civic action.
Corpus mean 0.24 of 4 · 4.8% score 2 or more · model agreement with reviewed scores 89.5%
Does it frame change as a shared project, with “we” language and common cause?
- 0
- Individualised, or waiting for a saviour.
- 2
- Some “we” language, but fragile.
- 4
- Sustained collective framing: citizens acting together.
Corpus mean 0.48 of 4 · 12.4% score 2 or more · model agreement with reviewed scores 79.0%
Does it reflect awareness that citizens have rights that entitle them to influence decisions?
- 0
- Politics framed as something done to citizens.
- 2
- Rights implicit, or invoked as distant abstractions.
- 4
- Clearly articulates rights to take part, be heard and hold power to account.
Corpus mean 0.35 of 4 · 7.3% score 2 or more · model agreement with reviewed scores 79.0%
Does it keep faith in democratic processes as the legitimate means of change?
- 0
- Endorses authoritarian shortcuts or a strong leader above accountability.
- 2
- Frustrated with democracy, but not yet endorsing alternatives.
- 4
- Faith in democratic processes despite their imperfections.
Corpus mean 0.47 of 4 · 11.3% score 2 or more · model agreement with reviewed scores 73.7%
Key terms
- Composite
- A comment’s score from 0 to 100: its applicable dimension totals as a share of the points available, rounded.
- Inapplicable
- A dimension the model judged could not be scored for a comment. It adds neither points nor possible points. It is not a zero.
- Restraint indicators
- The four indicators that register the absence of harm: anti-hostility, respectful tone, non-dehumanisation and anti-conspiracy. Summed, they form the restraint scale (0–16).
- Generative indicators
- The other sixteen, which register a positive contribution. Their sum divided by four and rounded forms the generation scale (0–16).
- Generative ladder
- A comment-level reading: nothing at all; restraint only; a weak, moderate or strong generative trace, by the highest generative score it reaches.
- Reception profile
- One of five groups found by clustering the twenty indicator scores. A tendency with soft edges, describing comments rather than people.
- Zone and corner
- Regions of a view defined by thresholds at 0.40 and 0.60 of the scale maximum, used as a reading aid for single comments.
- Lens
- A grouping of comments built on the scores or the text: a profile, theme, use of hope, frame, thread position, platform, wave or integrity signal.
- Integrity screen
- A separate machine classification of possible misinformation and its meta-narratives. Unverified, and never part of the DDI score.
- Poisoning the well
- Discrediting a source before it speaks. In a comment field, the presence or suspicion of inauthentic voices makes every voice count for less, and teaches citizens to dismiss one another as fakes.
- Reply-bombing
- Pasting the same reply under many comments, so that one text answers every critic.
- Liar’s dividend
- The benefit to bad actors when the possibility of fakes lets real evidence be dismissed. Here it also works as a tax on sincere citizens, who must prove they are real.
- Illustrative comment
- A comment from the curated set shown in the panel. Chosen to illustrate positions; not a random sample.
How to cite
How the map was made
Corpus
Comments were exported from the four source posts of the “What Hope Means” video (Facebook, TikTok, Instagram, X), and the full reply conversation on X was then retrieved through a historical search. Of 15,149 comments captured, 12,172 passed screening, which removed comments under 20 characters or three words, emoji-only comments, unresolved language and stimulus material. The primary cohort is 12,139 comments coded in uniform runs: Facebook 6,361, TikTok 3,224, X 1,962 and Instagram 592.
Coding
Across the DDI programme, human triple-coding by locally grounded university partners provides the ground truth on which GCRD’s ReHuman AI system codes at scale and produces national scores. This audience response study applies ReHuman directly to a single-topic corpus, using the DDI Comment Edition (Second Edition, September 2026). Its figures describe this reply and are not national DDI scores.
The primary runs used ReHuman 0.1.1, backed by a Gemini 3 Flash Preview foundation model, adapted to the DDI through its rubric, anchors and human-reviewed examples rather than fine-tuning. Against human-reviewed reference scores for 19 comments, ReHuman matched exactly on 293 of 370 indicator pairs (79.19%) and within one point on 93.78%, with a mean absolute difference of 0.286 on the 0–4 scale. Per-indicator agreement ranges from 55.6% (interpersonal trust) to 94.7% (de-escalation). This is percentage agreement on a small, same-study sample. It is not Krippendorff’s alpha and not an independent holdout test.
The lenses
- Reception profiles. k-means clustering (k = 5, 20 initialisations, fixed seed) of the twenty indicator scores for the 12,111 comments with all four dimensions applicable; silhouette 0.26.
- Uses of hope and redefinitions. Ordered pattern rules on comments containing the word; a second keyword set for the conditions named in redefinitions.
- Themes and the frame. Transparent keyword dictionaries on text with handles removed. They miss paraphrase and irony and catch some false matches.
- Threads. Replies linked to a parent by exact text match where the parent’s score is unambiguous (1,949 of 2,187 replies).
- Restraint, generation and the ladder. Computed from the indicator scores for comments with all four dimensions applicable.
- Integrity readings. The pipeline’s machine screen and meta-narratives; keyword narrative families, repeated and near-duplicate texts (TF-IDF cosine of 0.80 or more, three or more copies of twelve words or more), machine-style prose signals and accusations of bots, as defined by GCRD for this study.
- The poisoned well. Accusations of bots, paid posting or AI found by keyword; their targets read through the reply links; large threads (20 or more linked replies) compared by whether they received near-duplicate replies. Descriptive associations only.
Illustrative comments
The panel shows comments from a curated set of 480. For every profile on every platform, every corner of every pairing, every use of hope, theme, frame and integrity reading, the comments closest to the group’s average indicator profile were selected, together with the 38 quotations used in the report. Account handles, names and links were removed and profanity masked. Comments that name private individuals, disclose health conditions, or contain slurs or threats were excluded. The comments are public, but they are reproduced to illustrate positions on the map, not to identify or judge anyone.
Limits
- Capture is non-probability. Deleted, private and filtered comments are unknown, and short reactions are excluded. Findings describe the captured corpus, not the British public or each platform’s audience.
- Platform, source post, audience, interface and capture method vary together, so platform differences cannot be isolated as platform effects.
- Comments are not people. Authors may post several times and threads are dependent. No significance tests or confidence intervals are reported, because a conventional margin of error would mislead on a dependent, non-probability sample.
- The zones are provisional governance thresholds applied to single comments as a reading aid. They assign no configuration to the corpus.
- The study does not establish whether the video generated hope, changed opinion or voting intention, or attracted support or opposition.
- Integrity labels are unverified machine assessments and contribute nothing to the DDI. Signals of repetition or machine style are not evidence of who wrote a comment or why.
Read the whole study
- ReportWhat Hope Means to the PublicThe analytical report: findings, interpretation, and implications for policy and leadership.
- Interactive mapThe Discourse Map You are hereEvery scored comment on the map, with a guided reading, a comparison tool and seminar exercises.
- Lexical appendixThe Vocabulary of the ReplyThe words the public used, by platform, reception profile and DDI band.
Sources
- GCRD (2026). “Let’s Bring Back Hope” in 48 Hours, with its interactive discourse map and lexical appendix. GCRD Insights, July 2026.
- GCRD (2026). Let’s Bring Back Hope, Part 2: What Hope Means to the Public. GCRD Insights, October 2026.
- GCRD (2026). The Discourse Map: interactive companion to “Let’s Bring Back Hope, Part 2”. GCRD Insights, October 2026.
- GCRD (2026). The Vocabulary of the Reply: lexical appendix to “Let’s Bring Back Hope, Part 2”. GCRD Insights, October 2026.
- GCRD (2026). Democracy Discourse Index: From discourse measurement to democratic communication diagnosis. Working paper v0.3, August 2026.
- GCRD, ReHuman Lab and Sensika Technologies (2026). DDI Methodology and Rubric Guide, Second Edition: Comment Edition. September 2026.
- GCRD (2025). A Framework for Understanding AI-Induced Fracture and Authentic Leadership Restoration: A Five-Tier Trust Architecture. GCRD Insights.
- Let’s Bring Back Hope, Part 2 corrected publication package: findings, methodology and data files (snapshot 10 October 2026). More on the index at gcrd.org.uk/ddi.
Independence. An independent analysis, prepared without commission from any party, campaign or government body.
Global Centre for Rehumanising Democracy · gcrd.org.uk · info@gcrd.org.uk
Poisoning the well means discrediting a source before it speaks, so that whatever it says is discounted. In a comment field the source is the public itself. When inauthentic or synthetic voices are present, or are believed to be present, every voice loses value as evidence of what people think, and citizens learn to dismiss one another as fakes.
This lens brings together the evidence the corpus holds. Direct evidence of inauthentic commentary is limited: only 3.1% of comments carry any measurable sign of repetition or machine style, and none can be attributed to an operator. The clearer evidence is of what suspicion does. It lands on ordinary voices, it gathers in the threads where scripted text appeared, and it corrodes the trust the conversation depends on.
373 comments (3.1%) show a sign of scripting or machine style
Repeated texts include ordinary slogans such as calls for an election. 142 comments repeat a text that also appears on another platform. Signals are not attribution.
None of the 27 accusations that can be linked answers a comment with any sign of scripting
51 comments accuse other voices of being bots, paid or AI: 20 at the post, 31 in reply to other commenters. Of the 27 replies that can be linked to the comment they answer, 18 answer a call for an election or support for Reform or Restore. The accused comments average 20.4 on the composite, above the corpus mean of 14.7. Most are Counter-demand: a claim on the democratic process, answered with the charge that it is not real.
Profile of the accused commentsBy their wording, 18 accusations point at Reform, the right or Russia and 6 at the government’s side: paid staff, fake accounts, an AI video. Suspicion runs both ways.
2.5 accusations per 100 replies in large threads that received pasted texts, against 0.3 in those that did not
| Large threads (20+ replies) | With pasted replies | Without |
|---|---|---|
| Threads | 4 | 9 |
| Replies | 593 | 291 |
| Accusations of bots, paid or AI | 15 | 1 |
| Accusations per 100 replies | 2.5 | 0.3 |
| Mean composite of the other replies | 13.0 | 18.4 |
| Other replies in the Hostile profile | 40% | 23% |
Small numbers: 10 of the 15 accusations sit in the single largest thread, under “Hope is a General Election and a Reform Government” (377 replies). Large, heated threads attract both scripts and suspicion, so this is an association, not a measured effect.
Accusing comments average 5.6, and 42 of 51 are Hostile
Mean indicator scores on the 0–4 scale. Calling a fellow citizen a bot ends the argument without answering it. The accused are left to prove they are people: 3 replies do exactly that.
5 of the 20 Healthy comments carry machine-style markers
Share of each DDI band with machine-style markers. Long comments with the markers average 39.5 against 20.8 for other long comments. Set them aside and the Healthy count falls from 20 to 15, while the corpus mean barely moves (14.61). Humans also write this way; the markers cannot tell assisted from inauthentic.
333 comments repeat 119 texts word for word
98 of the 1,371 comments about an election or the mandate are exact repeats, and the hashtag #GeneralElectionNow appears 23 times. The fastest repetition came from one account at a time: a pro-government reply posted ten times on Facebook in under ten minutes, a list of government achievements posted nine times on TikTok in under three minutes, and an anti-government text posted three times on X in 70 seconds. Speed and repetition can make a narrowed menu of remedies look like the public’s settled view, even when a single voice is doing the repeating.
Pseudonymous author links, recovered after this map was first built, show that each of these bursts came from a single account. They do not show who was behind any account, or why.
Voices from the threads
“Why is that [name]? What specific Reform policy is your priority? Bot 🤖”
Facebook · reply · composite 3. The same text, with a different name, was posted under two comments that called for an election or backed Reform.“No bot here. There isn’t a specific policy. many of there policies resonate with me just because they are looking to change parts of the system that have thought to be broken/wrong for years”
Facebook · reply · composite 41. An accused commenter answering.“The “New Facebook Profile” agrees 😂. The £1 million extra annual salaried social media staff creating fake accounts?”
Facebook · reply · composite 1. Suspicion aimed at the government’s side.“You must be getting paid to write this s***.”
X · reply · composite 0. A reply to a comment supporting the Prime Minister.“I feel like half these comments are just pure bots 😭”
TikTok · top-level · composite 1.“One thing that would help with hope, would be to ban bots and burner accounts on social media!”
TikTok · top-level · composite 54. A citizen naming the poisoned well as the obstacle to hope.Names and handles are removed. Profanity is masked.
What a poisoned well does to democratic discourse
- The public stops being legible. When any voice might be manufactured, the comment field stops working as evidence of public feeling, for the leader who says he is listening and for the researcher who measures. A promise that people will be heard cannot be kept through a channel no one trusts.
- Suspicion becomes a weapon of dismissal. Here it fell mostly on citizens demanding an election or backing Reform. The demand for consent was met with the charge that it was not real, which answers nothing and ends the exchange.
- The liar’s dividend becomes a citizen’s tax. Sincere voices must prove they are people, and generous speech that sounds fluent is treated as manufactured. Every real participant pays for the possibility of fakes.
- Trust in trust erodes first. Accusing comments score near zero on interpersonal trust. In GCRD’s Five-Tier Trust Architecture this is the foundation: the confidence that a public is talking to itself. Without it, an appeal to hope has no audience it can believe in.
- Scripts are a method open to every side. Pro-government reply-bombing, anti-government copy-paste and templated accusations all appear. A domestic conversation can be degraded without any foreign operator, and the threads where scripts appeared are where hostility and suspicion gathered.
- Measurement is exposed where democratic quality is rarest. Synthetic fluency concentrates at the top of the scale, where the DDI finds its few healthy comments. Instruments that read text alone can mistake fluency for citizenship unless they keep integrity beside the score, record provenance and validate against human triple-coding.
- Hopelessness is the dividend. A public taught that the public may not be real is easier to persuade that nothing can change. That is the atmosphere anti-democratic forces prefer. The well is poisoned not only against the speaker but against hope itself.
What this evidence cannot show. It cannot identify who wrote any comment, how many accounts were involved or whether any were automated. Accusations are found by keyword and include jokes; implicit accusations are missed. Replies are linked only where the parent text matched, and thread comparisons rest on a handful of threads. The pattern is consistent with a poisoned well; it is not proof of a campaign.
