Who might be left out of AI-supported democracy?

Contribution to the handbook

Priority · Assessed 2026-10-31

This question identifies consequential risks and tests for democratic participation. Barriers to taking part differ from inaccurate representation of people who do participate: the supplied material warns of literacy and technology-wariness barriers and reports study-specific differences in preference-prediction accuracy. These findings and warnings do not establish exclusion from actual decisions.

Does this challenge the goal?

In-scope gap

The question asks how participation or representation might fail within the goal of serving citizens. It does not itself contest the goal or its beneficiary boundary. The existing definition leaves how interests are identified and protected open, but the handbook needs to examine exclusion and how to test for it.

The inquiry could reveal reasons to challenge a particular design’s desirability, but the question alone does not establish exclusion or require a definition change.

Making and interpreting laws →

Authority, accountability and resilience →

Implementation and transition →

Tests, stop conditions and unresolved work →

Distinguish barriers that prevent or constrain participation from inaccurate, incomplete, or unequal representation of people who participated.

What we can say so far

These are distinct ways people might be left out: - **Participation barriers:** Landemore warns that people who struggle to read or are wary of technology could be excluded by the digital divide. Distance, language and differing conversational customs may also make participation or mutual understanding harder. - **Representation after participation:** Even when people contribute, their perspectives may receive unequal attention or be suppressed. The Augmented Deliberation chapter reports that models have followed instructions in ways that may censor perspectives, and says capable models can be difficult for people to check. Data bias is another identified risk, although the supplied evidence does not establish its effect in a particular democratic process. The distinction matters: more opportunities to submit input do not by themselves show that participants’ views are accurately or equally reflected. Conversely, potential misrepresentation does not show that people were unable to participate.

What remains open

• The evidence does not quantify who is excluded or how often participants’ views are inaccurately or unequally represented.
• The Tokyo campaign report does not assess who participated or how well its input represented voters.
• The supplied material does not establish which safeguards or access designs reliably reduce these risks across settings.

Identify groups and mechanisms supported by the material; distinguish measured findings from warnings or plausible risks, and avoid treating study-specific differences as proof of exclusion from actual decisions.

What we can say so far

The material identifies several possible routes to exclusion, mostly as warnings rather than measured effects: - People who cannot read well or are wary of technology may be left out by the digital divide, according to Landemore. The excerpt does not estimate how many are affected. - People separated by distance or language, or whose conversational customs and styles differ, may find their perspectives harder to include or understand as deliberation becomes faster or larger-scale. The material does not measure how often this happens. - People holding particular viewpoints could have them suppressed: the Plurality chapter reports that models have followed instructions in ways that may censor perspectives, while capable models can be difficult for people to check. It does not establish the extent of suppression. - Biased data, privacy concerns and possible surveillance are broader AI risks identified by Landemore; they are not findings about exclusion in a particular democracy project. One preprint reports that accuracy changed little when the demographic details supplied to two tested models varied, and when prompts and consultation changed from English to Portuguese. This is a study-specific result, not evidence of equal performance across groups or languages, nor proof that anyone was excluded from actual decisions. Other authors propose that AI-supported deliberation might reduce unequal contributions, but that potential benefit is not established by the supplied comparison.

What remains open

• Evidence on which groups are actually excluded from participation, representation or final decisions, and how frequently.
• Evidence on the extent of viewpoint suppression or the effects of specific access and deliberation barriers across settings.

State what the evidence does and does not establish about prevalence, causes, consequences, and applicability to other populations, languages, tools, and governance settings.

What we can say so far

The evidence establishes some limited reported findings, but not how prevalent exclusion is or what its effects are across AI-supported democracies. In one preprint, the authors report little change in accuracy across specified demographic-input variations or when prompts and consultation switched from English to Portuguese. Those tests do not establish equal performance across groups or applicability to other populations, languages, models, tools, or governance settings. Other sources identify plausible causes or mechanisms of exclusion, including literacy and technology barriers, suppression of perspectives, and difficulties accommodating different languages and conversational styles. They do not measure how often these occur or establish their consequences. Proposed gains in inclusion or deliberation likewise remain possibilities rather than demonstrated general effects. Reported participation and policy-change counts in one Tokyo campaign do not show how representative participants were or establish broader effects.

What remains open

• Prevalence and distribution of exclusion across relevant populations and settings.
• Causal evidence connecting particular barriers or model behaviors to exclusion, and evidence of its consequences.
• Comparative evidence on other languages, tools, populations, and governance contexts.
• Evidence on whether proposed benefits or remedies improve inclusion in practice and for whom.