SmartAGI · Regulatory analysis · Issue 1 · 4 October 2026
*Conceptual illustration generated for this analysis; it does not depict a real event or campaign.*
On 2 September, Bill 18,669-06 on synthetic propaganda and automated accounts in elections was introduced in the Chamber of Deputies. A week later, on 9 September, five senators introduced Bill 18,642-07, which would punish concealed coordinated networks capable of simulating public participation. Both address digital manipulation in politics, but regulate different objects: who funds and distributes an election item; who directs a network designed to look like a collection of independent voices. The copy of the first motion included in the Chamber’s 70th session on 23 September identifies it as being at its first stage. The second remains under consideration too: neither is law. The text of Bill 18,669-06 shows which duties and prohibitions the Chamber proposes.
The motion leaves room for campaigns to use AI to draft, translate or design. Its concern arises when technology alters authorship or the appearance of events: coordinated automated accounts intended to influence a vote, words attributed to a candidate who never spoke them, or a video presented as a record of an invented scene. The bill seeks to regulate who is speaking and what happened as represented by the item; it does not give the state a general power to decide the substance of political opinions.
An army of accounts is not a crowd
Article 4 would prohibit commissioning, contracting, financing or operating, directly or through third parties, a set of automated accounts for electoral propaganda. The bill defines inauthentic coordinated behaviour as using those accounts in a coordinated way to influence the outcome of an election or plebiscite. It expressly exempts people who act together voluntarily and without payment.
A fictitious crowd is a useful image for the risk, but it is not a literal element of the motion. Article 3 defines inauthentic coordinated behaviour by the coordinated use of automated accounts to influence an outcome; Article 4 prohibits commissioning, contracting, financing or operating a group of such accounts for electoral propaganda. It does not additionally require proof that the operation meant to simulate public support. The express exemption covers people acting together voluntarily and without payment. A neighbourhood group may back a candidate, repeat a slogan and urge people to vote: that human organisation does not by itself become the automated conduct described by the bill.
The boundary becomes less clear when a campaign automates only part of the process. A person may choose the message and timing while software opens profiles, publishes variations, replies to users or likes allied accounts. The bill’s definition covers accounts operated wholly or partly by software and several types of activity, not just automatic posting. Investigators would have to examine coordination, payment, instructions, infrastructure and purpose. A high volume of messages, on its own, does not prove that a bot farm is behind them.
The exemption for unpaid human coordination helps separate mobilisation from automation, but leaves an evidential task. “Unpaid” does not mean “unorganised”: activists may receive instructions, use shared tools and follow a common schedule without being paid for each post. Conversely, a paid network may divide work among agencies, consultancies and intermediaries to conceal who commissioned the service. The motion recognises that chain by including direct and third-party operations, and by requiring declarations from digital electoral-propaganda providers. Enforcement will depend on reconstructing those links, not on treating a repeated hashtag as proof of wrongdoing.
Article 5 requires agencies, consultancies and propaganda providers to submit a sworn statement when they contract. Its wording has a textual problem: it asks them to declare that they have not engaged and will not engage in conduct prohibited “by this article”, although the article itself states no prohibition. Readers cannot tell whether the reference was meant to point to Article 4, Articles 4 and 8, or another provision; deciding by intuition would amend the motion from outside. The electoral administrator must attach the statement to the general income-and-expense account. Candidates, parties, electoral pacts and those responsible for plebiscite options would, in turn, have to report to the Electoral Service, when registering, any digital-positioning contracts made on their behalf during the previous twelve months. The declaration must identify the provider, its address and country of incorporation or residence, the amount, duration and services agreed.
That twelve-month window reaches beyond the visible campaign week. It makes sense for contracts that begin as “community management”, promotion or positioning and acquire an electoral purpose only as voting approaches. It also means the register must be clear enough to follow the money and distinguish legitimate services from an operation designed to manufacture support. Knowing that a contract existed does not yet reveal who wrote each post, which accounts distributed it or what part of the work was automated. Disclosure makes the commercial chain more visible; it does not replace an investigation into how the service operated.
Two bills, two accounts of the fictitious crowd
Bill 18,669-06 is not Chile’s only recent proposal concerning networks of accounts. On 16 September, the Senate reported that five senators had introduced Bill 18,642-07, a separate initiative to amend several statutes in order to prevent and punish digital manipulation of public opinion. The distinction matters: Article 4 of the Chamber motion concerns groups of automated accounts used for electoral propaganda; the Senate initiative defines an “inauthentic coordinated network” that may include automated, partly automated, fictitious or commonly controlled accounts, provided that the relationship is concealed and the operation is massive or systematic enough to suggest spontaneous, independent participation.
According to the Senate’s institutional account of the motion, the bill would punish anyone who creates, organises, finances, contracts, administers or uses such a network to mislead others about the existence, scale, origin, independence or spontaneity of public support, opposition or participation. For that conduct it proposes a term of presidio menor, a Chilean statutory imprisonment category, in its medium-to-maximum grades, and a fine of 20 to 100 monthly tax units. Its definition also says that a pseudonym, virtual identity or parody account is not fictitious merely because it does not use a civil name. The focus is concealed coordination and the manufactured appearance of many independent interventions, not anonymity by itself.
The legislative difference becomes clearer side by side. Bill 18,669-06 sets electoral rules on synthetic propaganda, labels, digital contracts and automated accounts; Bill 18,642-07 proposes a broader offence of manipulating public opinion through coordinated networks, using a definition that does not require every account to be a bot. One asks what is published and who pays for it during an election. The other focuses on who controls a network and what false impression of participation it seeks to create. Both are under consideration; neither is law.
Their coexistence leaves Congress to decide whether these rules should advance together, coordinate their definitions or retain different scopes and penalties. If two laws use “automated account”, “coordinated network” or “electoral propaganda” differently, campaigns and authorities may not know which threshold applies. Multiple motions do not amount to a settled design: the work is to connect the thresholds, evidence and safeguards without turning a pseudonymous identity or human mobilisation into a crime on appearance alone.
The label has to stay with the piece
Article 3 defines synthetic content as text, an image, audio or video—or a combination of them—generated or substantially altered through artificial intelligence or another digital technology in a way that appears to represent authentic facts, statements or circumstances. It excludes minor edits that do not change the meaning of what is shown. The definition tries to prevent a colour correction or sound adjustment from becoming synthetic propaganda by itself. Its focus is a piece presented as a record of something that supposedly happened.
Article 7 would require electoral propaganda that qualifies as synthetic content to carry the words “Content generated or altered with artificial intelligence” and the name of its issuer. For an image, the notice must appear on the image itself; for audio, it must be spoken at the start; for video, it must remain visible throughout. The duty also applies to text when it meets the definition. The rule does not stop at labelling the original file on a campaign account: it requires the warning to survive the way the propaganda is distributed.
Consider two hypothetical examples. A team uses a system to summarise policy proposals, then writes a post using its own sources and conclusions. That use does not automatically turn the text into a synthetic representation of facts or statements. If the team generates a video that appears to show a candidate announcing that a public benefit will be withdrawn, the invented scene, its resemblance to authentic footage and its circulation as propaganda raise precisely the problem the proposal seeks to regulate. The distinction depends on what the material claims to represent, not simply on whether an AI tool appeared somewhere in the workflow.
The motion does not ban all election propaganda made with AI. It requires disclosure when a piece falls within its definition. That raises operational questions for candidates, agencies and platforms: who decides whether a change is “substantial”? What happens to satire in the style of a news report? When does a political illustration represent an event, and when is it a metaphor? Must every user who downloads and reposts a clip preserve the label? The bill specifies where the notice should appear, but does not develop standards for borderline cases, short formats, excerpts or translations.
The definition also covers “another digital technology”, not just generative models. That avoids making the rule depend on a tool’s marketing name. It also makes it more important to explain which substantial transformation turns a piece into synthetic content. An edit can alter the meaning of a statement without using a generative model; digital production can look authentic even when it combines several tools. The proposed legal test concerns what the content appears to represent, but agencies and courts will need criteria to apply it consistently.
Article 8 moves from disclosure to a limit. It would prohibit commissioning, contracting or distributing synthetic content created or altered to invent, replace or change the actions, image or voice of a person—living, dead or fictional—so that a reasonable person might believe it records events or statements that actually occurred. Unlike a label, a prohibition addresses the kind of content and the appearance it creates. The person being imitated cannot automatically waive the rule: the text provides no such exception.
A visible warning can disclose that a video was generated; it cannot make the fictional record true. That is why the motion combines the two instruments. Article 7 reveals the synthetic origin of propaganda that may lawfully circulate; Article 8 prohibits certain realistic simulations that present invented events as fact. This is not a proposal to remove every image made with AI. The prohibition has a narrower threshold: a simulation that a reasonable person could take for an authentic record.
Not every false claim is a deepfake
The bill opens with an important principle: the proposed law would not intervene in the content of messages or in their truthfulness. That sentence seeks to separate control of fabricated records from state judgement about political ideas. A campaign may offer a mistaken reading of the economy or make a questionable promise; that argumentative error does not thereby become synthetic video. The proposal focuses on attributing an action or statement that never occurred to someone, and on simulating an automated crowd.
The separation is useful, though not mechanical. Article 8 uses an authenticity test: could a reasonable person believe the scene or statement happened? Applying it would require someone to assess format, context, distribution and the likelihood of confusion. The decision would not establish whether a policy proposal is true; it would judge whether a piece falsifies its source. The boundary protects debate about ideas while still requiring interpretation of the material.
The definition of “inauthentic coordinated behaviour” likewise requires a finding that a network aims to influence the result. The same software can schedule advertisements, automate customer-service replies or distribute propaganda. The mere presence of code does not appear to be the test: who used it, how accounts were coordinated and what role that coordination played in the election matter. Without procedures to request data, protect third parties and justify a penalty, a rule against bots could become a political label applied to rival accounts.
The motion seeks to reduce that risk by distinguishing automated behaviour from voluntary work by people. The exemption matters for civic organisations, community groups, unions, parties and grassroots campaigns. It also prevents an activity from appearing unlawful solely because several people use a shared tool or spread the same message. Even so, proving payment or automation may require contract records, payment trails, publication metadata or testimony. Transparency about providers and spending is therefore part of the investigative capacity.
The motion cites Brazil’s Superior Electoral Court Resolution 23.732/2024 as a precedent. Brazil’s framework has since added rules for the 2026 elections. Its measures on notices, deepfakes and the circulation of synthetic material near polling day give Chile’s Congress a concrete comparison; their details and limits appear below.
The European Union takes another approach. Its AI Act treats certain systems designed to influence the outcome of an election or referendum as high-risk, while excluding administrative and logistical tools whose outputs are not directly exposed to people. It also imposes transparency duties for certain synthetic image, audio and video content. One question concerns the risk posed by a system; another is whether a piece tells people that it was fabricated. Chile’s motion combines content, account and contract obligations in one proposal. To make the law workable, it must specify which authority verifies each duty and what evidence it may use.
The article requiring labels is missing from the list of fines
The bill exposes an initial mismatch between duties and penalties. Article 10 sets a fine of 30 to 300 monthly tax units for breaches of the prohibitions in Articles 4 and 8: coordinated automated accounts used for propaganda and realistic synthetic impersonation. False declarations by providers under Article 5 or digital-positioning services under Article 6 carry fines of 10 to 100 units, in addition to any applicable criminal liability. In the first case, the provider would also be barred from working in the next two electoral processes.
Article 7, by contrast, requires synthetic propaganda to be labelled but does not appear among the provisions to which Article 10 expressly assigns those fine ranges. Article 10’s first paragraph says that breaches will not be treated as electoral expenditure; that sentence does not itself establish the consequence of omitting a label. Other general laws may apply. What the proposal does not specify is a direct, dedicated sanction for the omission. In a transparency law, the difference between “show this notice” and “this is what happens if you do not” is not a drafting nicety.
Procedure and evidence also remain unresolved. The explanatory section says the Electoral Service should pursue the prohibited conduct, and Article 6 sends declarations about contracts to that service. Yet the visible articles do not set out a full procedure for detecting coordinated campaigns, obtaining platform data, receiving complaints, preserving evidence before it disappears, or distinguishing a paid operation from a civic group. Congress could specify those powers during debate. Without that detail, the agency may receive a duty to monitor whose operational scope the law does not describe.
Digital evidence vanishes quickly. An account is deleted, a story expires, an advertisement is withdrawn and the audience that saw a piece becomes harder to reconstruct. Election violations often require a swift response, but speed does not turn suspicion into proof. The framework would need to say what information is retained, for how long, who may request it, what judicial oversight applies and how a decision can be challenged. A graphic label cannot answer those questions by itself.
Labels can also become separated from context as content travels. The bill requires the warning to remain on a video throughout its duration; a platform may crop, transcode or display a preview that hides it. A user may repost a screenshot of a single frame. The propaganda’s original publisher has the initial duty, but the proposal does not set out a technical obligation for platforms to preserve or reinsert the notice when they reproduce or recommend the file. Other frameworks combine visible labels with platform procedures or declaration tools for political advertising. The committee will have to decide whether this rule works without that link in the chain.
The notice “generated or altered with artificial intelligence” also does not explain what changed. It does not say whether AI created the voice, rewrote the script, adjusted the image or produced the entire scene; nor does it identify the model or source material. That limit may be deliberate: requiring a technical dossier on every campaign item could make the message unreadable and demand information that does not change its meaning. Yet if the public is to recognise a fabricated representation, a standard legend answers only the first question—there was synthesis—and leaves the second open: which part of the supposed evidence was constructed?
The motion does add financial traceability that could fill part of the gap. Digital contracts, the provider, the amount, the term and the description of services would help investigators establish who paid for a positioning campaign. A register of synthetic products within electoral-expense accounts would be public. It could connect an item to its funding and commissioner. Even so, the bill does not define the access format, a unique identifier for each item, how a modified version links to its original, or how the record should show that a video was removed or corrected.
The spending account is filed after the audience has seen the content. It may support later auditing; stopping a simulation from reshaping debate before voting calls for timely complaints and review. If a video is copied across social media and messaging services before an authority rules, a correction reaches a different audience from the original. The motion does not create a special immediate-removal procedure like the one Brazil provides for certain violations, although general law may supply other routes. The result will depend on how the rules interact and which authority has power at each step.
Campaigns begin before the official campaign period
Chile’s statutory definition of electoral propaganda already covers public events and radio, written, image, audiovisual or analogous advertising that promotes candidates or parties for electoral purposes. Law 18,700 sets periods and conditions for such propaganda. The new motion would apply in that field: elections and plebiscites governed by the electoral laws it lists. Its temporal boundary may be decisive, because opinion is partly formed outside the formal campaign window.
The proposal tries to capture some of that activity by requiring disclosure of digital-positioning contracts entered into during the previous twelve months. Even so, the labelling duty applies to “electoral propaganda”. The text will need to clarify how it treats synthetic material posted on a political page weeks or months before the official period and then amplified during a campaign. The general law distinguishes propaganda from the expression of ideas and information; technology does not erase that line, but it lets campaigns prepare material, audiences and networks in advance.
The prohibition on automated accounts also applies to electoral and plebiscitary processes. What happens when an operation begins months earlier and remains active into the campaign? The provision describes coordinating accounts to influence the outcome of the process, so the purpose may cross different dates. The law still needs to explain how that purpose is proved and what temporal connection makes a penalty applicable. This is not a matter of extending dates by intuition; it is a matter of identifying who commissioned the operation and what it sought to achieve.
Campaigning is not only a paid advertisement with a logo. A network can gain reach through unpaid posts, themed accounts, comments and coordinated reactions that make one view look spontaneous or dominant. The bill covers accounts that post, share, react or interact, and requires disclosure of positioning services. That is broader than limiting the law to audiovisual items or advertising purchased on a platform. Implementation must not assume that every highly active account is a bot: automation, coordination and an electoral commission must be established together.
The question of proportionality does not require playing down the problem. A campaign that pays to simulate a crowd uses money and automation to distort the perceived level of social support. A network of volunteers sharing arguments shows organised support, which is part of political activity. Confusing the two could deter collective expression; ignoring the first would let money buy the appearance of many independent voices. The motion’s explicit exemption offers a principle of separation. Its strength will depend on applying the evidence equally to allies and opponents.
Brazil is already debating what to do near polling day
The Brazilian comparison deserves attention because its rules go beyond saying “add a label”. The current text of Resolution 23.610/2019, amended by Resolutions 23.732/2024 and 23.755/2026, specifies how to identify synthetic content, which minor adjustments are exempt, how chatbots are treated and what applies in the hours surrounding a vote. For certain fabricated or manipulated content that could damage electoral balance or the integrity of the process, it also prohibits deepfakes used to favour or harm candidacies and sets consequences for registration or office.
The 2026 revision added a specific time-based prohibition: from 72 hours before until 24 hours after the election ends, publishing or reposting—even without charge—and paid promotion of new synthetic content using the image, voice or statements of candidates or public figures are prohibited, even if the content is labelled. It also requires providers of political promotion to add a field where the purchaser declares AI use. These are measures adopted by Brazil’s Superior Electoral Court under its authority and for its elections. They do not transfer automatically to Chile, where Congress is designing statutory duties and penalties under a different institutional arrangement.
The practical comparison is this: what risk does a label address, and what risk arises when voters receive a false clip just before the polls close? A label helps identify generation or alteration; Brazil’s time limit starts from the premise that there may be no time to verify and respond at the end of the vote. Chile’s motion contains no quiet period for synthetic material as filed. Congress could maintain a less restrictive solution if it concludes that a warning and an authenticity-based ban are enough; it must also explain how an effective response can reach people in the final hours.
European rules separate a different issue. The AI Act treats certain systems intended to influence an election or referendum as high-risk, except administrative and logistical uses whose outputs are not directly exposed to people. At the same time, it requires transparency for certain synthetic image, audio and video. One question is the risk presented by a system; another is whether a piece tells people that it was generated. Chile brings several obligations for content, accounts and contracts into one proposal. To keep the law workable, it must state which authority checks each obligation and what evidence it may rely on.
What needs to change before this can become law
The motion starts from a defensible democratic intuition: voters may hear harsh, partial and even mistaken propaganda; they should not discover too late that an alleged recording never happened or that a flood of accounts was one contracted operation. The definition of synthesis, the labelling duty and the prohibition on realistic impersonation address distinct problems. The rules on accounts and contracts also seek to prevent private spending from masquerading as citizen deliberation.
The text needs to complete the path from conduct to a reviewable decision. What data must a platform keep, and who may request it? How is the beneficiary identified when a provider is a foreign consultancy or a chain of subcontractors? What does the disclosure of a positioning contract cover when a third party paid for it? What penalty applies to a failure to label under Article 7? What procedure allows an urgent response followed by an appeal? Answering those questions is part of the legislative work required to make the prohibition enforceable and reviewable.
The authenticity test also needs observable criteria. The phrase “a reasonable person might believe” avoids banning every parody but leaves context to be assessed. Is a satirical legend part of the image itself? Does it matter if the imitated voice belongs to a public figure? Was the video shared with an explicit statement that it is fictional? Was it cropped to remove that notice? Cases can differ, and the law should allow them to be assessed without requiring citizens to guess the author’s intention.
For coordinated accounts, evidence must also support the distinction between a campaign with many people and a campaign that simulates many people. Repeating a slogan is evidence of distribution, not necessarily automation; posting quickly is not sufficient proof of payment. Account histories, timing, payments, shared instructions and operating patterns can be indicators, but each needs context and a right of response. An automated detection system might guide an inquiry; it cannot replace a finding about who commissioned, financed or operated the network.
Useful transparency would leave a public record of the item and the person responsible: what was produced, who commissioned it, which service took part, who declared the expense, and how the item was corrected or removed. Some data may need protection for privacy, investigations or security. The bill should distinguish public from restricted information rather than assume that a “public register” solves every problem. Voters need to know the political and synthetic origin without receiving personal data that adds nothing to scrutiny.
There is a political choice behind every threshold. Congress must decide whether it is primarily regulating the falsification of audiovisual evidence, automated accounts, the money buying both, or the voter’s exposure to unverified material. The motion covers all four, though with different levels of precision. That breadth can help because an electoral simulation often combines content, distribution and funding. It also increases the risk that institutions will not know which evidence to seek first.
Citizens need to know when a campaign has fabricated a scene, who attributed it to a candidate, which company was paid and which authority can verify the claim. The law can protect that possibility without deciding which opinions are acceptable. To do so, the label needs an enforceable consequence, and the ban on bots must remain distinct from organised political mobilisation.
Congress has a concrete proposal: four sets of rules on automation, synthetic content, contracts and spending, with penalties specified for some conduct. Deliberation should add a sanction for the labelling duty, define enforcement powers, preserve evidence and establish routes of appeal. The law needs to make three things legible: when a campaign automates accounts, when an item fabricates a representation, and who answers for each. A crowd mobilised by people is a form of political action; a network of software-operated accounts is assessed under the definition and prohibition proposed in Article 4. A synthetic record impersonating events falls under the analysis of Articles 7 and 8. Congress should set those criteria and the consequences of breaching them before an authority has to improvise during an election.
Primary sources: Bill 18,669-06, legislative document for session 70 · Chile’s Law 18,700, Article 30 · Brazilian electoral regulation consolidated, TSE Resolution 23.610 and amendments · TSE rules for Brazil’s 2026 elections · EU AI Act, consolidated text in force on 27 July 2026 · Senate legislative record, Bill 18,642-07 · Senate account and text excerpts · Chile’s Law 18,700 on Ley Chile · “The colours of AI and the reasons behind the law”, external column by Rodrigo Castillo Murillo and Francisco Figueroa Ogalde.