Generative AI and the Public Sphere: Transparency and AI Literacy

1) Introduction

The advent of generative AI has dramatically lowered the barriers to entry for content production, making accessible and inexpensive what once required significant resources: the industrial creation of highly credible text, images, audio, and video. This is not merely a technological evolution, but an institutional shift that directly affects the functioning of the public sphere, understood as the vital space in which citizens inform themselves, debate, and consolidate their opinions. Contrary to common perception, the main problem lies less in AI’s ability to generate falsehoods than in its ability to produce plausibility: when plausibility circulates without context and without adequate signals of transparency, it becomes a form of power capable of shaping perceptions of truth, urgency, and trust.

2) The public sphere as infrastructure

It is essential to recognise that the public sphere is not an abstract entity; in the digital ecosystem, it materialises through specific technical and organisational choices. Seemingly neutral elements such as the ordering of content, the labels applied, the ease of tracing sources, and the frictions imposed on sharing constitute the architecture of this space. Consequently, the governance of visibility is intrinsically political, because technical decisions determine which narratives gain social traction and which remain confined to the margins.

3) What changes with generative AI: cost, scale, mimicry

The introduction of generative AI radically alters three structural variables in the information landscape. First, it transforms scale, enabling the generation of thousands of items of content and variants in extremely short timeframes. Second, it perfects mimicry, making style, tone, and even personal identity replicable—as demonstrated by the phenomenon of deepfakes. Third, it enables the creation of micro-narratives, whereby the same theme is tailored and “coloured” differently for distinct audience segments, while concealing a single underlying direction. In this scenario, disinformation moves beyond the single false news item and corrodes systemic trust: if any content can be synthetic and yet plausible, the real risk is that citizens, overwhelmed, will abandon verification and orient their choices by mere belonging or cognitive fatigue.

4) Transparency: what must become visible (and to whom)

In this context, transparency should not be understood as explaining every process to everyone, but as the capacity to make intelligible the minimum information necessary to contextualise what one encounters. Three fundamental levels can be distinguished:

Content transparency: the ability to identify whether a piece of content has been generated or materially manipulated by AI systems.
Context transparency: clarity about who bears responsibility for publication, including authorship, channel, and purpose.
Infrastructure transparency: understanding the logics of distribution and targeting—a crucial dimension for research and auditing.

This last dimension is the most democratic, because it does not limit itself to examining a single false item but interrogates the entire architecture of attention.

5) EU governance: the AI Act, the DSA, and the principle of transparency

Within the European regulatory framework, transparency emerges as a cornerstone of governance. The AI Act (Regulation (EU) 2024/1689) establishes a risk-based governance model, introducing specific transparency obligations for different contexts of use. Alongside this, the Digital Services Act (Regulation (EU) 2022/2065) imposes stringent requirements on the management of systemic risks by very large platforms. At a supranational level, the Council of Europe’s framework convention likewise reiterates the nexus between AI, human rights, and democracy, underscoring the need to identify synthetic content.

However, for these norms to be effective, a cultural gap must be closed: transparency is functional only if citizens possess the competences to decode it—hence the relevance of AI literacy.

6) AI literacy: the civic competence to use AI without being used by it

If transparency reduces the opacity of content and channels, AI literacy reduces the opacity of our decisions, teaching us to treat the output of generative AI as an instrumental aid rather than as an epistemic authority. In practical terms, AI literacy can be articulated into four competences applicable to everyday participation in the public sphere:

Recognise the nature of the output: understand that it is plausible text, not guaranteed evidence or a certified source.
Verify factual steps: check dates, numbers, quotations, names, links, and attributions with care.
Manage risk and privacy: discern when it is inappropriate to enter sensitive data or information traceable to third parties.
Recognise framing and persuasion: identify omissions, emotional cues, stereotypes, and cognitive biases.

These competences are not ancillary; they are what makes transparency genuinely useful. A label that signals “AI-generated content” is of limited value if the reader does not know what to do with that information.

6.1 The most common risks affecting the public sphere

At this stage, AI literacy can no longer mean the mere technical ability to write prompts; it must evolve into a genuine civic competence that enables conscious, critical, and responsible use of AI systems—fully aware that generated output is plausible but not guaranteed. The AI Act itself operationalises this principle in Article 4 by requiring providers and deployers to ensure an adequate level of AI literacy for their staff. This competence is necessary to mitigate four everyday risks that affect the public sphere:

Hallucinations and false details: confidently stated but unverifiable claims, or plausible errors in facts and attributions.
Bias and stereotypes: the automated reproduction of pre-existing cultural patterns and distortions in judgement or social representation.
Privacy and sensitive data: the inadvertent inclusion—within prompts or documents—of third-party data or confidential information.
Manipulation and persuasion: emotional framing and selective topic choice, where persuasion masquerades as neutral exposition.

Without the ability to recognise these risks, any transparency label risks becoming a bureaucratic formality with no real impact.

7) Practical guidelines: risk traffic-light model and checklists

To keep the public sphere intelligible and contestable, translating these principles into practice requires lean, repeatable procedures proportionate to the level of risk. Two practical tools are proposed below: a traffic-light model and two operational checklists (one aimed at publishers, the other at readers and sharers).

7.1 Risk traffic-light model (when to trust the output)

Level	Typical scenarios	What to do
Green	Creative activities, brainstorming, drafting and non-critical rewrites.	Use AI as a co-author, while retaining control over coherence and tone.
Yellow	Summaries, formal communications, technical explanations, and numerical analyses.	Verify key points against independent sources and request precise references.
Red	Decisions with an impact on rights, health, finances, or reputation.	Do not delegate judgement: AI may support, but verification and final responsibility must remain human.

7.2 Checklist for readers and sharers: AI literacy in 7 questions

Before accepting an output, ask seven fundamental questions:

What type of output am I requesting (facts, opinions, creativity, or decisions)?
What standard of proof is appropriate here? Do I need verifiable sources?
What is the potential harm if the output turns out to be wrong?
Am I entering proprietary data, third-party data, or sensitive information?
Which biases or distortions are plausible in this specific context?
Can I triangulate at least two independent sources on the crucial points?
Who ultimately bears responsibility for the decision and the communication?

A single rule synthesises everything: if the impact is high, AI must support rather than act as the decision-maker.

7.3 Checklist for publishers: transparency guidelines for creators, newsrooms, public administrations, and companies

For those who produce information (creators, newsrooms, public administrations), it is essential to adopt a mini-policy on transparency that includes:

Disclosure: indicate the use of AI when it affects meaning, facts, or identity;
Context: state who is responsible (byline or editorial responsibility);
Citations: provide reliable sources for factual claims;
Corrections: disclose corrections transparently to increase accountability;
Separation of registers (information, advocacy, satire);
Avoid overclaiming: do not turn probabilistic outputs into certainties;
Privacy: do not publish or reuse outputs that incorporate unnecessary sensitive data.

8) Conclusion

The quality of the public sphere cannot be preserved through polarising slogans (“ban AI” / “AI will save everything”), but through two mutually reinforcing pillars: on the one hand, transparency (signals, context, and auditability), and on the other, AI literacy (the widespread capacity not to delegate critical judgement). Only through this synergy will transparency avoid remaining a bureaucratic checkbox, and AI literacy will move beyond technique, becoming a robust democratic infrastructure.