confiance statutaire citizen science; Participatory Science; participatory research; citizen science; participatory science; community-based participatory research; data quality; trust; validity; objectivity; metrology; bias; QA/QC; validation; verification; fitness for purpose; provenance; FAIR; open science; knowledge co-production.

confiance démontrable

A significant part of the controversies surrounding citizen science rests on an apparently simple question: “can data produced by citizens be trusted?”

Yet this formulation already contains an assumption that should be demonstrated: the existence of a particular class of data whose defining property would be the identity of the person collecting them.

From the perspective of metrology, statistics and the epistemology of science, this category is fragile. A temperature measured at 18.4 °C does not acquire a new physical nature because the thermometer was read by a salaried hydrologist, a farmer, a teacher, a student or a retiree. Keywords:

Contemporary literature tends to shift the assessment away from the identity of the producer and towards the properties of the procedure. Kosmala and colleagues had already concluded, in their review that became a reference in the field, that very different participatory projects can produce data whose accuracy equals or exceeds that of professionals when the system combines appropriate design, training, validation, replication and statistical treatment of errors [44].

Lewandowski and Specht contributed a particularly important result to this discussion: in studies where professional and volunteer data were compared against the same independent benchmark, professionals were more accurate in only four of the seven available comparisons [45].

This obviously does not demonstrate universal equivalence between experienced and inexperienced people, but rather that

1. A good or bad question that has become historically structuring

.

A first conclusion can therefore be drawn.

What does change, however, is the probability of error associated with different components of the process: site selection, instrument calibration, compliance with stabilization time, reading, transcription, contextual qualification or application of the protocol.

Frontiers in Ecology and the Environment

The difficulty appears as soon as four different objects are confused:

• the real phenomenon;
• the observation or measurement;
• the recorded data;
• the evidence used to support a scientific proposition.

An animal presence, a nitrate concentration, a temperature, a perception, a social practice or a pathological episode exist or occur independently of their entry into a database. Their observation then results from an interaction between a phenomenon, a material or cognitive observation system, a protocol and an observer. Recording transforms this observation into data. It is only after additional operations — qualification, selection, aggregation, modelling, comparison — that certain data become evidence supporting a proposition.

From this perspective, the expression “citizen data” describes its professional status cannot be used as a substitute for a measure of quality, not its essence.

It is more accurate to speak of data:

• observed by a participant;
• measured using a specified instrument;
• produced according to a defined protocol;
• accompanied by a given level of metadata;
• possibly verified by a second observer;
• corrected, weighted or modelled;
• and intended for an explicit set of uses.

This ontological distinction is far from being purely philosophical. It determines the technical architecture of systems. Data standards dedicated to Participatory Science seek precisely to preserve the link between observation, project, protocol and context. PPSR Core thus distinguishes the “Project”, “Dataset” and “Observation” levels in order to make data provenance interpretable [46].

The same principle has long existed in other scientific infrastructures. Darwin Core, for example, provides a common vocabulary for specifying occurrence, taxon, event, location, identification and the evidence associated with a biodiversity observation [47].

Well-documented Participatory Science data are therefore much less an isolated number than an Scientifically justified trust must apply to a data production chain, not to a social category of producers..

2. Ontology: what exactly is “participatory data”?

Trust is not a metrological property.

One can trust false data; one can distrust accurate data. At least five notions must therefore be distinguished.

production history refers to the proximity between an observation and a value or classification considered correct.

information object endowed with provenance concerns, in particular, the dispersion of repeated measurements.

3. Epistemology: trust, truth, validity and quality are not synonymous

refers to the stability of a process or measurement under comparable conditions.

Accuracy concerns the possibility of supporting a conclusion on the basis of the available data.

Precision is a relationship in which an actor agrees to depend, at least partially, on information or an information producer whose work they cannot fully verify themselves.

This latter situation is by no means specific to Participatory Science. Modern science relies extensively on a cognitive division of labour. A climatologist trusts metrology laboratories; a biologist trusts instrument manufacturers; an epidemiologist trusts administrative databases; a researcher trusts authors they cite; an evaluator trusts data they did not collect themselves.

The relevant question therefore becomes that of the Reliability.

Hendriks, Kienhues and Bromme notably distinguish three dimensions of epistemic trust placed in an expert: expertise, integrity and benevolence or intentions [48].

Participatory Science adds a particular feature recently highlighted under the term “dual nature of trust”: scientists must be able to trust the data produced by participants, while participants must be able to trust the scientists, institutions and uses made of their contributions [49].

Trust is therefore bilateral and, in highly participatory research, multilateral.

Validity of inference

A naïve conception of objectivity sometimes equates professionalization with neutrality: the scientist would be objective because they are trained and disinterested; the participant would be suspect because they have interests, values or personal experience of the problem.

Contemporary philosophy of science makes this opposition difficult to sustain.

Elliott and Rosenberg specifically analysed three recurring objections made against citizen science: the supposed absence of hypotheses, insufficient data quality and activist involvement. They show that none provides an argument allowing citizen science to be discredited Trust [50].

Scientific objectivity depends largely on procedures: making hypotheses explicit, critical scrutiny, the possibility of replication, documentation of decisions, plurality of viewpoints, examination of uncertainties and exposure of results to criticism.

In some cases, the involvement of participants possessing situated knowledge can even reveal implicit assumptions that had remained invisible to specialists.

This question is an old one. Wynne showed, in his work on Cumbrian sheep farmers after the Chernobyl accident, that the opposition between “scientific expertise” and “lay knowledge” concealed the fact that local actors possessed indispensable knowledge about agricultural practices and local conditions [51].

In participatory environmental or health research, the same situation regularly reappears: people living with the phenomenon know certain dimensions of exposure, uses, temporalities or behaviours that are not necessarily included in the initial models.

reasons for trust

One of the most robust outcomes of the field’s evolution is the gradual abandonment of the abstract question of “quality” in favour of

4. Objectivity is not the absence of humans

.

The ISO 19157 standard applied to geographic information proceeds in the same way: it defines dimensions and procedures for quality assessment without imposing a universally sufficient level; it is up to the user to determine whether the dataset possesses the characteristics required for the intended application [52].

This reasoning is particularly well suited to Participatory Science.

Data may be insufficient to:

• establish a regulatory concentration to within a few micrograms;

but sufficient to:

• identify a spatial anomaly requiring confirmatory instrumental measurement.

An observation may be insufficient to:

• distinguish between two morphologically very similar species;

but sufficient to:

• determine the genus or family and track a large-scale distribution.

A time series may be insufficiently standardized to:

• estimate absolute abundance;

but excellent for:

• detecting a long-term trend.

Clare and colleagues specifically propose reversing the approach: first determine how much error is compatible with the intended inference, then invest in the corrective procedures needed to reach that threshold [53].

Parrish and colleagues similarly refer to in principle and

5. The central principle: fitness for purpose

 [54].

The methodological consequence is major:

fitness for purpose

fitness to purpose

The comparative literature justifies neither the claim that “citizens produce data as good as scientists” in every case, nor the opposite claim.

Aceves-Bueno and colleagues conducted one of the first quantitative reviews of this literature. Their analysis shows a generally favourable trend for citizen-generated data, but also considerable heterogeneity between tasks, systems and criteria used [55].

Prior experience, training, length of participation, task difficulty and protocol design explain part of the variation.

Field studies provide particularly instructive illustrations.

Crall and colleagues show, in invasive plant monitoring, that volunteers can achieve good performance, but that some tasks, particularly certain abundance estimates, remain sensitive to expertise [56].

In urban tree inventories, Roman and colleagues observed approximately 90% agreement with experts for several variables, but lower performance for precise species identification and certain assessments of tree condition [57].

In Snapshot Serengeti, aggregating classifications made by more than 28,000 participants made it possible to produce animal-image classifications extremely close to those of experts [58].

In Mosquito Alert, Palmer and colleagues showed that a system combining citizen reports, photographs, entomological validation and correction for sampling effort could provide monitoring of tiger mosquito expansion comparable to conventional methods while offering considerably greater geographical coverage [59].

Across 34 tropical sites in four countries, Danielsen and colleagues compared independently produced results from local communities and scientists for 63 taxa and several types of resource use. The two systems produced similar conclusions regarding the status and trends of the resources observed [60].

These results should not be turned into a slogan. They indicate that intentional design.

the required quality must be specified on the basis of the inference and intended use before the protocol is defined.

Participatory water-quality monitoring has several decades of history and provides some of the most direct tests of comparability.

A review specifically focusing on volunteer water-monitoring programmes identified fifteen studies comparing citizen and professional physicochemical measurements. Most reported comparable quality for the parameters studied, although gaps remained regarding very large historical and heterogeneous datasets [61].

The subsequent study of the Texas Stream Team is important because it examines large-scale existing data from 1992 to 2016 rather than an artificially controlled experiment. Overall agreement with professional measurements approaches 80% statewide for the parameters studied and 91% in the local subset where conditions were better controlled [62].

In the Yukon Basin, the Indigenous Observation Network similarly shows that a community-based network can operate according to a data cycle including planning, collection, quality assurance and documentation compatible with professional scientific requirements [63].

The 2023 review by Blanco and colleagues of 72 studies on citizen monitoring of surface-water quality nevertheless confirms the considerable heterogeneity of instruments, parameters, spatial scales and temporalities [64].

The field of water therefore perfectly illustrates the general principle: the question of trust must be asked parameter by parameter and use by use.

6. What empirical comparisons between volunteers and professionals show

The main difficulties identified in the literature can be organized into several families.

the scientific competence of a system is distributed among people, protocols, instruments, controls and processing methods

It may result from:

• an insufficiently precise instrument;
• poor calibration;
• sensor ageing;
• incorrect reading;
• a poorly applied protocol;
• a delay between sampling and analysis.

These problems also exist in professional systems. This brings us back to the training of observers on this aspect, or even their supervision if, for example, they are being supported as a group.

7. The case of water: a particularly instructive laboratory

This particularly concerns:

• taxonomic identification;
• image classification;
• symptom characterization;
• habitat categorization;
• interpretation of behaviour.

It can be reduced through supporting photographs, double classification, consensus or secondary expertise. The design and sizing of the participatory research project will need to address this aspect.

8. The real sources of fragility in participatory data

This is often the most important problem.

Participants frequently choose:

• accessible locations;
• areas close to home;
• times compatible with their activities;
• objects considered interesting;
• certain species rather than others.

A dataset can therefore consist of highly accurate individual observations while still providing a biased representation of the phenomenon studied.

8.1. Measurement error

An absence observed after two hours of searching does not have the same meaning as an absence recorded after thirty seconds. The lack of information about effort is therefore a major source of difficulty for opportunistic data.

8.2. Classification error

Experience, age, prior knowledge, motivation, fatigue or sensory acuity can affect performance.

8.3. Sampling bias

People who participate in a project generally do not constitute a random sample of the population. This bias becomes particularly important when researchers wish to infer the attitudes, behaviours or perceptions of the general population.

8.4. Unknown observation effort

A participant directly exposed to pollution may be accused of being “interested”. But the existence of a stake does not demonstrate the presence of falsification. Rather, it justifies an architecture that allows observation, control, analysis and interpretation to be separated.

The existence of conflicts of interest, motivations or values is not specific to non-professionals either. The relevant epistemology is therefore procedural: which mechanisms make it possible to detect and correct influences likely to distort inference?

8.5. Observer heterogeneity

One of the most operational findings in the literature is the identification of a genuine engineering approach to participatory data quality.

Freitag, Meyer and Whiteman identified twelve strategies for building credibility distributed across preparation, collection and analysis [65].

A robust system can combine:

8.6. Participation bias

• explicit definition of the intended use;
• pilot study;
• simplification of the protocol;
• determination of tolerances;
• calibration;
• training;
• possible participant certification;
• test data or “gold questions”.

8.7. Stake-related bias

• value-range checks;
• automatic timestamping;
• geolocation;
• supporting photograph or recording;
• repeated observations;
• duplication of a fraction of measurements;
• calibration checks;
• structured data entry;
• logging of operations.

9. QA/QC: trust can be engineered technically

• automatic detection of outliers;
• comparison with reference data;
• second reading;
• community validation;
• expert review;
• observer weighting;
• detection models;
• statistical correction;
• explicit qualification of uncertainty.

Baker and colleagues studied 259 ecological systems. For the 142 whose verification methods could be established, expert verification remained the dominant method, followed by community consensus and automated procedures. The authors propose a hierarchical architecture: automated or community processing of the main volume and referral of ambiguous cases to experts [66].

This architecture is particularly important in the age of artificial intelligence: scarce human expertise can be concentrated on observations offering the greatest information gain.

Before collection

Participatory data have contributed to changing another intuition: imperfect data are not necessarily unusable.

Bird and colleagues show that the main biases encountered in large citizen databases — observer heterogeneity, preferential sampling, spatial dependencies, irregular repetitions — are largely similar to problems already known from other ecological databases [67].

Methods that can be used include:

• hierarchical models;
• mixed-effects models;
• occupancy models;
• estimation of detection probability;
• effort weighting;
• observer calibration;
• spatio-temporal models;
• machine learning;
• integration of multiple data sources.

The principle is a classic one: During collection.

This distinction between the ecological process and the observation process lies at the heart of occupancy models and much of contemporary quantitative ecology.

It is also a response to the idea that a database containing errors would be intrinsically unusable. The question is rather whether the errors are sufficiently characterized to be incorporated into inference.

After collection

An opposite misunderstanding would be to believe that massive data volume automatically corrects every flaw.

It does not.

Random error may decrease as the number of observations increases. Systematic bias, on the other hand, can become extremely precisely wrong.

One million observations concentrated around urban areas do not constitute a random sample of a territory.

One million sensors with the same calibration defect do not produce a better measurement of the true value.

One million participants using an ambiguous definition can reproduce the same conceptual bias.

10. Statistics: noise does not prevent inference

explicitly model the observation process in addition to the observed phenomenon

Trust can be increased without changing a single observation, simply by making its genealogy accessible.

Ideally, a user should be able to know:

• who or what system produced the observation;
• according to which protocol;
• on what date;
• with which instrument;
• with which version of the protocol or software;
• which corrections were applied;
• which validation was performed;
• which uncertainty is associated with it;
• under which licence the data can be reused.

Bowser and colleagues observed that many projects had developed quality practices, but that documentation, interoperability, open access and infrastructure sustainability frequently remained insufficient [68].

The FAIR principles —

11. Quantity does not, however, replace quality

— provide a general reference here [69].

Balázs and colleagues also recommend the use of international standards and systematic documentation of quality-assurance procedures [70].

An important implication follows: Volume reduces certain types of uncertainty; it abolishes neither biases nor validity problems..

12. Documentation, provenance and open science

Even when quality procedures exist, participatory data can remain underused.

Burgess and colleagues surveyed biodiversity researchers and project managers. Their work identifies a preference for certain sources and certain types of producers that is not always justified by demonstrated differences in quality [71].

Turbé and colleagues examined 503 European projects and 45 in-depth studies. Lack of trust in data quality appeared as one of the main obstacles to their integration into environmental policy, even though many of the projects studied had high scientific standards [72].

The distinction between Findable, Accessible, Interoperable, Reusable and The Science of Citizen Science becomes central here.

A Participatory Science quality policy therefore cannot be limited to producing good data. It must produce transparency about uncertainty can increase trust more than claiming that uncertainty does not exist.

13. A persistent difficulty: perceived quality is not measured quality

Institutional recognition involves questions of professional boundaries.

The use of data produced by an administration, a university, an association or a community does not necessarily carry the same symbolic cost for the decision-maker.

Yet several studies show that Participatory Science can contribute to research, conservation and public policy when properly integrated.

McKinley and colleagues conclude that a citizen-based system that is properly designed, implemented and evaluated can produce robust science and contribute to natural-resource management [73].

In the marine field, Hyder and colleagues emphasize that each dataset should be assessed individually rather than rejected on the basis of its origin [74].

Institutional integration nevertheless requires continuity, governance, standardization, documentation and responsibility. The question is therefore not only: “are these data accurate?”, but also: “which institution can guarantee their preservation, interpretation and long-term use?”

actual quality

The potential changes scale again when citizen science is considered for monitoring biodiversity or the Sustainable Development Goals.

Chandler and colleagues show that citizen and community programmes already cover several Essential Biodiversity Variables, including distribution, abundance, phenology and certain ecosystem functions, while still presenting significant geographical and taxonomic gaps [75].

Fritz and colleagues subsequently proposed a roadmap for integrating citizen science into monitoring the Sustainable Development Goals, emphasizing the importance of quality, access and interoperability [76].

This perspective changes the very function of certain data. Their advantage is not always to achieve the greatest possible precision at a given point, but to provide spatial, temporal or social resolution that a conventional professional network could not economically achieve.

perceived quality

The literature on contributory citizen science is not sufficient to understand participatory research in the strong sense.

In Community-Based Participatory Research, Participatory Action Research and related traditions, participation can concern:

• formulation of the problem;
• choice of the research question;
• definition of variables;
• protocol design;
• collection;
• analysis;
• interpretation;
• dissemination;
• decision;
• action.

The potential epistemic effect is then deeper.

Jagosh and colleagues define participatory research as the co-production of research between researchers and people affected by the phenomenon or responsible for action. Their realist review highlights benefits that may emerge through partnership mechanisms: improved recruitment, protocol adaptation, cultural relevance, better interpretation and long-term effects [77].

Balázs and Morello-Frosch proposed describing this contribution through three dimensions: clear evidence of their quality. Participation can strengthen rigor by improving design and interpretation, relevance by directing research towards the right questions, and reach by facilitating dissemination and use of results [78].

Participation is therefore not merely a democratic concession made to science. In some configurations, it constitutes

14. The institutional problem: accepting data also means accepting an unusual producer

.

15. Participatory data and international indicators

Data quality is generally assessed after researchers have decided what should be measured.

Yet research can produce extraordinarily precise measurements of a poorly chosen variable.

This form of error is particularly important for complex phenomena.

People who live daily with an ecosystem, a public service, a disease, a professional activity, a risk or pollution may point out:

• an overlooked temporality;
• an exposure location absent from the initial plan;
• a behaviour not taken into account;
• an intermediate variable;
• local terminology;
• an overly broad category;
• a plausible causal mechanism;
• a consequence that was not being investigated.

Co-production can therefore act upstream of metrological accuracy.

This perspective is consistent with the post-normal science of Funtowicz and Ravetz, which proposes broadening the peer community when uncertainties are high, values are contested and stakes are high [79].

Recent work nevertheless shows that not every form of citizen science automatically constitutes an “extended peer community”. This depends on the actual degree of participation in problem framing, evaluation and decision-making [80].

16. Participatory research: when participants are no longer merely collectors

In participatory research, trust is no longer only a judgement made about data.

It becomes a mechanism for producing those data.

A community that does not trust researchers may:

• refuse to participate;
• leave the study early;
• withhold certain information;
• respond strategically;
• prevent access to certain field sites;
• refuse to authorize data reuse.

Conversely, a trusting relationship can increase continuity of participation, the contextual quality of information and the ability to interpret anomalies.

Jagosh and colleagues showed, in their realist evaluation of CBPR partnerships, that trust and partnership synergy could contribute to the sustainability of collaborations and produce long-term ripple effects [81].

Lucero and colleagues propose an empirical typology of trust ranging from rigor, relevance, reach to a mechanism for reducing framing error, intended to track the evolution of partnerships [82].

The review by Gilfoyle, MacFarlane and Salsberg nevertheless confirms the heterogeneity of definitions and measurements of trust in participatory research networks [83].

17. Framing error: a dimension often absent from discussions of quality

The systematic review published in 2025 by Yu and colleagues analyses 211 publications devoted to knowledge co-production through citizen science. It identifies 29 variables organized around participants, project design and context. Training, trust, feedback mechanisms, participatory capacities and data management appear among the factors frequently associated with outcomes [84].

The 2026 review by Israel and colleagues simultaneously confirms the maturation of the CBPR field: increasing academic and community credibility, development of validated evaluation instruments and accumulation of findings on public policy and structural change [85].

Finally, the systematic review of 124 empirical studies devoted to public trust in science shows that integrity, source characteristics, modes of communication and public involvement play a role in trust. Some of the included studies indicate that direct participation in research can increase trust in its results, particularly when the process is transparent [86].

The scientific problem can therefore no longer be reduced to the accuracy of the recorded observation. It simultaneously concerns the technical quality of the dataset, the quality of the knowledge-production process and the relationship between producers and users of that knowledge.

18. Relational trust and scientific quality become interdependent

The state of the art makes it possible to propose a nine-dimensional framework.

trust deficit

Are we measuring the relevant phenomenon?

reflective trust

Does the protocol make it possible to answer the question?

19. 2025–2026 research: towards an integrated conception of scientific and social outcomes

Do the instruments and observations possess the required accuracy and precision?

20. An integrated matrix of trust in participatory data

Does the dataset sufficiently represent the relevant space, time, populations or events?

1. Conceptual validity

Are there procedures capable of detecting errors, inconsistencies and questionable classifications?

2. Protocol validity

Are biases and uncertainties integrated into the inference?

3. Metrological quality

Can we understand how the data were produced and transformed?

4. Sampling quality

Are the conditions governing the production, ownership and use of the data explicit and acceptable?

5. Validation quality

Are these properties sufficient for the intended decision or inference?

This matrix makes it possible to move beyond an unproductive opposition between “professional data” and “citizen data”.

6. Statistical quality

Future architectures could associate each dataset — or even each observation — with a trust profile rather than a single verdict.

Such a profile could document:

• observer training level;
• protocol used;
• instrument calibration;
• associated evidence;
• number of independent validations;
• automated validation;
• expert validation;
• agreement with other sources;
• estimated uncertainty;
• metadata completeness;
• representativeness;
• correction status;
• dataset version;
• validated scientific uses.

The objective would not be to create a social hierarchy among contributors. It would be to make trust 7. Provenance and reproducibility.

The emergence of common standards such as PPSR Core, the integration of FAIR metadata and the development of hybrid methods combining community validation, expertise and artificial intelligence make this architecture technically realistic.

8. Legitimacy and governance

Several propositions appear sufficiently well supported to constitute a reference foundation.

9. Fitness for purpose, there is no scientific reason to consider data inferior a priori because they were produced by a non-professional participant.

21. Towards a trust score rather than a binary label

, there are tasks for which professional expertise greatly improves accuracy and others for which short training, an appropriate protocol or secondary validation can achieve comparable performance.

computable, explainable and reassessable, the main weaknesses of large citizen-based systems often concern representativeness, observation effort, documentation and infrastructure as much as, or more than, the basic accuracy of observations.

22. What science now allows us to assert

, professional data do not constitute an ontologically error-free benchmark. They too must undergo quality assurance, calibration and evaluation.

First, quality must be assessed relative to the intended use.

Second, explicit QA/QC, validation, documentation and modelling procedures make it possible to increase and quantify trust.

Third, the perceived quality of data and their demonstrated quality may diverge; an institutional bias against unconventional sources remains documented.

Fourth, in more participatory forms of research, participation can increase scientific quality by improving framing, relevance, access to the field and interpretation.

Fifth, this participation can also degrade research if responsibilities are ambiguous, if the required skills have not been acquired, if power asymmetries are not addressed or if participation is purely symbolic.

Sixth, trust and quality must therefore be conceived as properties of the research architecture, rather than as moral qualities attributed to a category of actors.

Seventh

The scientific literature no longer allows a serious general opposition between “real scientific data” and “citizen data” to be sustained.

It leads instead to a more demanding distinction.

There are data whose provenance is well known or poorly known; instruments that are calibrated or not; protocols that are appropriate or inadequate; observers who are trained or not; sampling designs that are representative or biased; errors that are characterized or invisible; validation procedures that are strong or absent; datasets that are documented or opaque; and inferences that are compatible or incompatible with their properties.

These distinctions cut across professional science and Participatory Science alike.

The specificity of citizen science lies less in the emergence of a new epistemological class of data than in the redistribution of knowledge-production functions among a larger number of actors. This redistribution introduces new risks, but also new capacities: spatial coverage, temporal continuity, volume of observations, situated knowledge, early detection, access to difficult field sites, critique of existing categories and multiplication of verification possibilities.

The decisive shift therefore consists in moving from Eighth — “I believe these data because they come from a scientist” — to Ninth — “I know the conditions under which these data were produced, which controls they underwent, what uncertainty they contain and for which uses they are sufficiently robust”.

This is probably where the deepest contribution of Participatory Science to contemporary epistemology lies: it forces science to make visible what, in conventional systems, often remained implicit — the precise conditions that make it possible to transform an observation into data, data into evidence and evidence into trustworthy knowledge.