# The Island in the Map | Deriss Research

Canonical: https://deriss.com/articles/the-island-in-the-map
Description: Social Explainable AI grew 18× in four years. The only strand touching collective life sits at the edge of the map — and two papers in 478 mention contestability.

---
# The Island in the Map: What 647 Studies Reveal About the Missing Social in Social Explainable AI

Source: https://deriss.com/articles/the-island-in-the-map
Author: Soheill Deriss
Published: 2026-08-06
Canonical context: https://deriss.com/llms-full.txt

> A field called Social Explainable AI grew roughly 18× in four years. Cluster it by its own language and the only strand that touches collective life sits alone at the edge of the map — and when you read it, it turns out to be individual classification wearing social clothing. Two papers in 478 mention contestability. The science of explaining a system to a collective does not yet exist. Cities are about to need it.

A research field tells you what it believes by what it measures. It tells you what it has forgotten by what sits at the edge of its map — and by the words that never appear in it.

**TL;DR:** We assembled 647 peer-reviewed journal articles on Social Explainable AI published between 2021 and 2026, and clustered 640 of them by the language of their abstracts. The field is accelerating hard and its centre of gravity has moved from algorithms to accountability. But seven of its eight thematic clusters are **dyadic**: one human, one model, one decision. The eighth — the only one that touches collective life — is a 36-paper cluster on social media that sits measurably apart from everything else. It is also one of the fastest-growing regions of the field, which is the good news. The bad news is what it contains: in 23 of those 36 papers the task is detecting or classifying a property of a single person or a single item. The social is the **data source**, not the subject of the explanation. Across the corpus, 52% of abstracts invoke an individual decision-maker. **0.4% — two papers — mention contestability or recourse.** The field has built a rich science of showing people reasons and almost none of letting them argue back. That is the layer cities need, it is the return path missing from the civilian node, and it will be supplied from a foreign cloud by default unless it is built locally.

---

## I. A field arrives

Four years ago this literature barely existed. In 2021 we count 15 journal articles that jointly address explainability, a human or social dimension, and AI. In 2025 we count 277 — a roughly **eighteen-fold increase**, with the break between 2023 and 2024, when annual output leapt from 42 to 111. The corpus closes mid-2026 with 165 papers already recorded in a partial year, on pace to exceed 2025. Nothing here has plateaued.

<!-- FIGURE 1 -->

Volume is the least interesting thing about that curve. The composition underneath it is the story.

In 2021 and 2022, close to half of all annual output — 47% in both years — sat in a single cluster: studies of how explanation changes user trust, reliance and mental models. By 2024 that share had fallen to 18%. It has since settled around a fifth. The cluster never shrank; it grew every year in absolute terms. It was outpaced.

What outpaced it, measured by the share of each theme's output published in 2025–26:

| Theme | Share published 2025–26 |
|---|---|
| Ethics, Governance & Accountability | 80% |
| XAI for Detection & Security | 79% |
| **Social Media & Misinformation** | **78%** |
| Human-Centered AI Design | 77% |
| Explanation → Trust & Cognition | 64% |
| Human–AI Interaction & Decision-Making | 64% |
| Trust-building in Domain Systems | 60% |
| Black-Box Interpretability in Healthcare | 56% |

Read plainly: the early field asked *how do we explain a model?* The field maturing now asks *who is accountable for the explanation, and what does the human do with it?* That is a migration out of computer science and into governance and design — the migration that makes this literature strategically relevant rather than merely technical.

Note the third row. We will come back to it.

---

## II. The map, and the island

Cluster all 640 abstracts by their own vocabulary and lay them out in semantic space, and eight coherent themes appear.

<!-- FIGURE 2 — HERO -->

| Theme | Papers | Citations |
|---|---|---|
| Explanation → Trust & Cognition | 153 | 4,429 |
| XAI for Detection & Security | 100 | 785 |
| Black-Box Interpretability in Healthcare | 95 | 908 |
| Ethics, Governance & Accountability | 80 | 508 |
| Human–AI Interaction & Decision-Making | 77 | 1,274 |
| Human-Centered AI Design | 69 | 943 |
| Social Media & Misinformation | 36 | 268 |
| Trust-building in Domain Systems | 30 | 296 |

Seven of these form a single connected continent. Healthcare interpretability bleeds into ethics; ethics bleeds into human-centered design; design bleeds into interaction and trust. The vocabulary is shared, the boundaries are soft.

One cluster does not participate, and this is measurable rather than impressionistic. Projections like the one above can manufacture separation that is not there, so we tested it in the original high-dimensional term space instead. Each cluster's centroid was compared to its nearest neighbouring cluster by cosine similarity. For the seven continental themes, that figure runs between **0.60 and 0.69** — they genuinely overlap. Trust-building in Domain Systems sits further out at 0.49. **Social Media & Misinformation sits at 0.355**, roughly half the cohesion the continent shares with its neighbours, and it carries the highest mean silhouette score of any theme in the corpus.

The island is real. It is not an artefact of the picture.

Now consider what separates it. It is not method, and it is not application domain. It is **scale of subject**. Every cluster on the continent studies a dyad: one clinician and one diagnostic model, one analyst and one alert, one user and one interface. Healthcare, security, decision support, design — all individual-scale. The island is the only cluster in the corpus whose subject matter reaches beyond a single person at a time.

Which raises the obvious question: what is actually in it?

---

## III. The island is not what its name suggests

Here is the finding that reorganised this entire analysis.

Of the 36 papers on the island, **23 are detection, classification or prediction tasks**. They detect fake news. They detect hate speech. They detect depression, suicidal ideation, emotional subtext, humour, and sentiment in individual posts. They classify disaster imagery and predict default risk on social lending platforms.

This is capable, useful work. It is not a science of collective explanation. In almost every case the unit being explained is a single item or a single person, and *social media is simply where the data came from*. Detecting depression in one user's posting history is an individual-scale classification problem that happens to draw on a social corpus. The social is the dataset, not the subject.

Read the whole island and the genuinely systemic papers number roughly five: an agent-based model of sociotechnical transparency in platform algorithms, a framework for misinformation diffusion across networks, a study of the personalisation paradox, a large-scale analysis of public climate discourse, and — tellingly — a 2024 paper in *AI & Society* whose entire argument is a plea for interdisciplinary research between explainable AI and the social sciences. The field's own most collective-minded contribution is a request that someone start.

So the corrected finding is sharper than the one the map first suggested. It is not that the field ignores the social. The social cluster is one of its fastest-growing regions — 78% of it published in the last two years. **The problem is the door it is entering through.** The collective is arriving in this literature framed as a *detection* problem, and a science that begins in detection builds forensic tools: instruments for identifying what went wrong after coordination has already failed. It does not build infrastructure for coordination that works.

This lands exactly where [Instalment III](https://deriss.com/articles/europes-human-centric-experience-layer) drew the line between the **engagement stack** — optimised for attention, value captured centrally — and the **experience layer**, optimised for participation and belonging. And it echoes the sixth of the seven fractures in [Instalment I](https://deriss.com/articles/human-centric-era): algorithmic optimisation for engagement over meaning. The research community is studying the wreckage of the engagement stack with real care. Almost nobody is studying the road.

### What the vocabulary confirms

If the field had a collective turn underway, its language would show it. We searched all 478 records carrying full abstracts:

| Concept | Papers | Share |
|---|---|---|
| An individual decision-maker (clinician, user, operator, analyst, patient) | 249 | 52.1% |
| A collective unit (community, citizens, population-level) | 32 | 6.7% |
| Civic or urban context | 15 | 3.1% |
| Allocation | 14 | 2.9% |
| Calibration | 19 | 4.0% |
| Overreliance | 11 | 2.3% |
| **Contestability or recourse** | **2** | **0.4%** |

Two papers. In a literature of 478 abstracts about explaining artificial intelligence to human beings, two mention whether the human can argue back.

That single number is the thesis of this article. Explanation, as this field currently practises it, is something delivered *to* a person. It is almost never something a person can *use against* the system that produced it.

---

## IV. The accelerant problem

A reasonable objection: perhaps the individual case is simply the tractable one, and the collective follows once the fundamentals are solid.

The corpus argues otherwise, through its two most-cited papers, which point in opposite directions.

The most-cited study in the dataset — 1,178 citations — finds that explainability and causability drive user perception, **trust and acceptance** [1]. The second, at 821 citations, is titled *To Trust or to Think*, and finds that explanations induce **overreliance**: people accept an AI's suggestion even when it is wrong, adding explanations does not reduce this and may increase it, and deliberate cognitive forcing is required to counteract it — at the cost of users rating the forcing designs least favourably [2]. Later work replicates the ambivalence in high-risk decision tasks [9], and a systematic review of clinical settings is titled, plainly, *How Explainable AI Can Increase or Decrease Clinicians' Trust* [7].

The mechanism is identical in every case. Explanation raises confidence. Whether that confidence is *warranted* is a separate variable that explanation does not touch — and which 96% of this literature does not measure.

**Explanation is a trust accelerant, not a trust proof.** This is the same structural error we identified one layer up in [Guarded Globalization, One Year On](https://deriss.com/articles/guarded-globalization-one-year-on), where trust was inferred from a supplier's *location* rather than verified as a property. Here it is inferred from a system's *articulation*. A friendly border is not provenance; a fluent rationale is not reliability.

So the collective case is not a scaled-up version of a solved problem. It is the same unsolved problem with no one left to catch it. An uncalibrated explanation given to one clinician is a risk a colleague, an audit or a second opinion may still intercept. An uncalibrated explanation given to a city — *this is why the service went there and not here* — has no second reader, and on the evidence above, no route of appeal.

---

## V. The return path

[Instalment IV](https://deriss.com/articles/the-intelligence-convergence) named the Context Problem: superintelligence is only as powerful as the context it can reach, which is why the platforms are building devices that capture context continuously and ambiently. It argued that when a citizen's context flows into a model owned elsewhere, what has been surrendered is not convenience but informational sovereignty. And it proposed the **civilian node** — the citizen as a sovereign participant in the city's network rather than a data point inside someone else's.

That instalment closed on a question it deliberately left open: *who builds the civic intelligence layer that sits between the superintelligence labs and the citizen?*

Here is the part of the answer this evidence supplies. **A node requires two pipes.** Instalment IV specified the upward one — context leaving the citizen under consent and control. It did not specify the downward one. A civilian node that emits context and receives nothing it can interrogate is not a node. It is a sensor with better manners.

The return path is explanation: the mechanism by which a person, a neighbourhood or a municipality asks a system *why* and receives an answer they can test, contest and act on. Social Explainable AI is the science of that return path — and the measurements above show it has been built almost entirely for one person at a time, with the contest step missing at a rate of 499 papers to one.

Meanwhile demand arrives on schedule. The EU AI Act reached its main application milestone on 2 August 2026 — four days before this was published — and with it Article 86, which gives any person affected by a decision taken on the output of a high-risk system listed in Annex III the right to obtain from the deployer *clear and meaningful explanations of the role of the AI system in the decision-making procedure and the main elements of the decision taken*.

Read the scope carefully. The right belongs to an **affected person**, singular. It is triggered by a decision that adversely affects **that person's** health, safety or fundamental rights. It is silent on whether the explanation is calibrated to the system's actual reliability. And it says nothing at all about explaining an allocation to the population it was made about — the case where no single person's rights are obviously infringed, but a community's are shaped. Europe has legislated the accelerant, and regulated neither the brake nor the collective case.

This is the gap. It is a research gap, a regulatory gap, and — for anyone building civic infrastructure in the next thirty-six months — a market gap.

---

## VI. What changes for leaders

**Founders and product leaders.** If your system allocates, matches or coordinates across people rather than advising one of them, you are operating in the part of this literature that does not exist yet. Treat explanation as a product surface with its own specification, not a compliance string rendered after the fact. The newest work in the corpus is already migrating from one-shot explanation toward **interactive** systems a human can interrogate [3]. Build for that endpoint, not the legal minimum.

**Enterprises and boards.** Stop reporting explanation *coverage* — the share of decisions that emit a rationale. Coverage became table stakes by law this month. Report **calibration**: whether the confidence your explanations produce matches the accuracy your system actually delivers. Only 4% of this literature measures it, which means the honest prior is that yours is unmeasured too.

**Cities and public institutions.** Civic AI procurement specifies transparency as a document. Specify it as a capability instead: can a resident ask why, get an answer at the scale the decision was actually made, and challenge it? Two papers in 478 address that last step. If you assume the market has solved it, you are assuming something the literature does not support.

**Policymakers.** Article 86 is a floor, and it is built for a dyad. The open work is collective explanation — a standard for explaining allocations to the populations they affect. That standard does not exist anywhere. Whoever writes it first sets the default for everyone else.

---

## A Stockholm note

Sweden has the rare preconditions for building the return path rather than importing it: high institutional trust, mature digital identity rails, municipal governance with real capacity, and a research base in federated and privacy-preserving methods strong relative to population.

It also carries the specific exposure named in Instalment IV. Language is the operating system of culture. An explanation delivered to a Swedish resident by an English-language model trained on Silicon Valley data is not merely a translation problem — it encodes another society's assumptions about what counts as a sufficient reason, and about whether a reason can be disputed at all.

The pilot that matters is small and concrete: one city, one coordination system, one population, with calibration and contestability instrumented from the start. That is the wedge Instalment III prescribed, applied to the layer Instalment IV left open.

---

## Key metrics

- **Explanation Calibration Ratio** — user confidence after explanation divided by the system's actual accuracy on that class of decision. Above 1.0 is manufactured trust.
- **Overreliance Delta** — change in a user's error-catching rate once explanations are switched on. On this literature's evidence, frequently negative.
- **Contestability Rate** — share of consequential decisions a person can actually interrogate or challenge, as distinct from merely receiving a rationale for. The corpus suggests almost nobody is measuring this.
- **Collective Explanation Coverage** — share of *allocation* decisions (who received what, and who did not) that produce any explanation to the affected population. Currently, in most civic systems, zero.
- **Time-to-Understanding** — lead time from decision to an affected person who can restate why it was made. The citizen-facing analogue of Time-to-Assurance.

---

## How we know this

This analysis draws on 647 peer-reviewed journal articles retrieved from the Crossref scholarly index across four query strands spanning explainability, human-centricity, social transparency and trust, covering 2021 to mid-2026. The 640 records carrying usable text were clustered into eight themes by term frequency (TF-IDF with k-means, k=8) and laid out in two dimensions by t-SNE. Cluster separation was verified independently of that projection, by cosine similarity between cluster centroids in the original term space and by silhouette scores. Keyword analysis was run on the 478 records carrying full abstracts; the remaining 162 are title-only, so every share reported in the vocabulary table is a **lower bound**.

Three limits are worth stating plainly, because they bear on the central finding.

**The corpus is journal-only.** Conference proceedings — CHI, FAccT, IUI — are the highest-velocity venues in this field and are under-represented here. That is where much social-computing and contestability work is published, so the two-paper contestability count is certainly an undercount of the wider field. A conference-inclusive replication is the obvious next test, and we intend to run it. Our expectation is that it raises the absolute numbers without changing the ratio, because explanation research in those venues is also overwhelmingly framed around an individual decision-maker facing a single system. That expectation is falsifiable, which is the point of stating it.

**Keyword presence is a proxy for concept presence**, and a coarse one. A paper can address recourse without using the word. We report it because the disparity is not marginal — 249 against 2 — and a measurement error of that magnitude is not plausible.

**The corpus carries no author names**, only author counts. We mention this because it produced an instructive failure during preparation: an automated drafting pass filled the gap with plausible-looking attributions, three of which we caught and confirmed to be wrong before publication. The references below are therefore cited by title, venue, year and DOI only. It is a small illustration of the article's own argument — a fluent output, confidently rendered, that no one had checked. Explanation is not verification.

**2026 is partial**, ending mid-year, and is excluded from the growth chart. It is included in theme-share and vocabulary analysis, where the partial year affects all themes equally.

Every figure and claim on this page traces to the published dataset, and every reference below resolves to a DOI. This is the disclosure posture we have argued for elsewhere: verifiable, machine-readable, citable. An argument about explainability should be inspectable — which is why the map above is, too.

---

## Call to action

The field named *social* has spent four years explaining machines to individuals, and has only just begun looking at collective life — through the narrow door of detecting what has already gone wrong. Cities will not wait for the literature to widen it. They will procure the layer that exists.

If you are building civic coordination infrastructure, instrument the return path now: calibration, contestability, and an explanation that survives contact with the population it was made about.

Interested in piloting a calibrated, human-centric civic intelligence layer out of Stockholm? [Get in touch](/contact), or run a [Terminal](/terminal) scan to see where your own systems currently sit.

---

## Sources

*Cited by title, venue, year and DOI. Our corpus records author counts but not author names, so attributing these papers to named authors would mean asserting something the dataset does not contain — which is the error this article is about. Every DOI below resolves; follow it for the full author list.*

1. *The effects of explainability and causability on perception, trust, and acceptance: Implications for explainable AI.* International Journal of Human-Computer Studies, 2021. 1,178 citations. https://doi.org/10.1016/j.ijhcs.2020.102551
2. *To Trust or to Think* (Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-Assisted Decision-Making). Proceedings of the ACM on Human-Computer Interaction, 2021. 821 citations. https://doi.org/10.1145/3449287
3. *From explainable to interactive AI: A literature review on current trends in human-AI interaction.* International Journal of Human-Computer Studies, 2024. 141 citations. https://doi.org/10.1016/j.ijhcs.2024.103301
4. *AI Transparency in the Age of LLMs: A Human-Centered Research Roadmap.* Harvard Data Science Review, 2024. 143 citations. https://doi.org/10.1162/99608f92.8036d03b
5. *Explainable medical imaging AI needs human-centered design: guidelines and evidence from a systematic review.* npj Digital Medicine, 2022. 256 citations. https://doi.org/10.1038/s41746-022-00699-2
6. *Measures for explainable AI: Explanation goodness, user satisfaction, mental models, curiosity, trust, and human-AI performance.* Frontiers in Computer Science, 2023. 211 citations. https://doi.org/10.3389/fcomp.2023.1096257
7. *How Explainable Artificial Intelligence Can Increase or Decrease Clinicians' Trust in AI Applications in Health Care: Systematic Review.* JMIR AI, 2024. 193 citations. https://doi.org/10.2196/53207
8. *In AI We Trust? Effects of Agency Locus and Transparency on Uncertainty Reduction in Human–AI Interaction.* Journal of Computer-Mediated Communication, 2021. 191 citations. https://doi.org/10.1093/jcmc/zmab013
9. *Effects of Explainable Artificial Intelligence on trust and human behavior in a high-risk decision task.* Computers in Human Behavior, 2023. 170 citations. https://doi.org/10.1016/j.chb.2022.107539
10. *AI Ethics: Integrating Transparency, Fairness, and Privacy in AI Development.* Applied Artificial Intelligence, 2025. 176 citations. https://doi.org/10.1080/08839514.2025.2463722

*Citation counts are Crossref "is-referenced-by" as of the corpus build (1 July 2026) and lag real-world impact, particularly for 2025–26 work.*

**Policy and framework references**

- European Commission — [AI Act regulatory framework](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai)
- EU Artificial Intelligence Act — [Article 86: Right to Explanation of Individual Decision-Making](https://artificialintelligenceact.eu/article/86/) (applicable from 2 August 2026 under Article 113)
- European Commission — [Industry 5.0: towards a sustainable, human-centric and resilient European industry](https://research-and-innovation.ec.europa.eu/knowledge-publications-tools-and-data/publications/all-publications/industry-50-towards-sustainable-human-centric-and-resilient-european-industry_en)
- European Data Protection Supervisor — [TechDispatch on federated learning](https://www.edps.europa.eu/data-protection/our-work/publications/techdispatch/2025-06-10-techdispatch-12025-federated-learning)

Full dataset — 647 records, 640 clustered across eight themes — available with this article.

---

## Related

- [The Intelligence Convergence](https://deriss.com/articles/the-intelligence-convergence) — Instalment IV
- [Europe's Human-Centric Experience Layer](https://deriss.com/articles/europes-human-centric-experience-layer) — Instalment III
- [The Deriss Society 5.0 Vision](https://deriss.com/articles/the-deriss-society-50-vision-a-blueprint-for-human-centric-european-innovation) — Instalment II
- [Design for the Human-Centric Era](https://deriss.com/articles/human-centric-era) — Instalment I
- [Guarded Globalization, One Year On](https://deriss.com/articles/guarded-globalization-one-year-on) — the Verification Layer
- [Beyond ESG: Bridging Corporate Responsibility and Industry 5.0 for Human-Centered Growth](https://deriss.com/articles/beyond-esg-bridging-corporate-responsibility-and-industry-50-for-human-centered-growth)
- [Deriss Terminal](/terminal)

---

*This is the fifth instalment in the Deriss research series on human-centric infrastructure. Previous publications explored the human-centric era, the Society 5.0 blueprint for European innovation, Europe's experience layer, and the intelligence convergence. This piece supplies the series' first primary evidence base, and names the return path the civilian node requires.*

*Deriss is a research-driven studio and product company guiding leaders through growth, brand transformation, and AI-native innovation in the challenges of Industry 5.0. Based in Stockholm.*