Maria Ressa says the third AI wave is coming for human agency

01

WHAT ACTUALLY CHANGED

Reuters published an interview with Ressa on September 15 after speaking with her during a visit to Oslo for a conference on threats to democracy. She argued that public pressure and legislation are needed to limit harmful AI systems and framed regulation as a public-safety question rather than only a speech question. These are her policy positions, not enacted rules or consensus findings.

Ressa divided the technology's public impact into three waves. The first is the capture of attention by social media, which she connects to polarization and isolation. The second is intimacy through chatbots. The third is agency through agentic and multi-agent systems that can act for people. The framework is a narrative model for connecting several technologies, not a controlled study, universal taxonomy, or proof that each wave causes the next.

Her focus moves the argument from what a system says to what it can do. A chatbot without tools produces conversation. An agent with an account, memory, payment method, browser, contacts, calendar, workplace access, or publishing permission can turn a recommendation into a state change. The risk depends on the exact capability, permission, user, context, reversibility, and consequence, not on the word agent alone.

Ressa called for technology companies to be accountable for harm and argued that age limits do not solve design features that can manipulate adults as well as children. She also rejected the claim that Western regulation necessarily hands an AI advantage to China, citing Chinese restrictions as an example. That cross-country comparison is her argument in the interview and is not independently established by the cited materials.

She proposed communication spaces built by public-interest groups rather than controlled solely by large technology companies. Reuters describes Rappler Communities, launched in 2023, as a mobile app using the open-source Matrix protocol to connect Rappler's newsroom and readers. Ressa presents it as one alternative for safer democratic conversation. The interview provides no comparative evidence that it improves listening, participation, trust, or civic outcomes at scale.

02

WHY THIS MATTERS

The three-wave frame joins problems that are often regulated separately. A recommender system may fall under platform rules, a companion chatbot under consumer or child-safety rules, and an autonomous assistant under financial, employment, health, or cybersecurity rules. A person experiences them as one sequence: notice me, know me, persuade me, then act for me. Oversight needs to follow that complete path.

Intimacy can quietly change the quality of consent. A person may disclose more to a system that sounds patient, affectionate, or certain, then accept its suggested action without noticing the transition from conversation to authorization. Products should signal when the mode changes, restate the consequence, separate emotional engagement from execution, and require fresh approval outside the persuasive flow.

Agency is not binary. Software can suggest a restaurant, draft a message, choose the recipient, send the message, spend a budget, or keep repeating the action. Each step is a larger delegation. Good design exposes that ladder, lets people choose the rung, limits duration and scope, and prevents a general instruction from becoming permanent authority over unrelated tasks.

A human-in-the-loop label can hide a rubber stamp. If the system produces a polished plan, applies time pressure, buries alternatives, or makes rejection difficult, the final click does not prove meaningful control. Reviewers need the source, uncertainty, important alternatives, requested permission, expected consequence, and enough time to decide. They also need an independent stop path that the agent cannot negotiate away.

Ressa's public-infrastructure analogy deserves a real test. Democratic communication spaces do not become healthy merely because they are decentralized, nonprofit, or open source. They still need sustainable funding, moderation, security, accessibility, due process, privacy, interoperability, and evidence that communities can disagree without manipulation or harassment. Public interest is an operating model, not a decorative label.

03

WHERE IT COULD HELP

• Create a capability ladder for every assistant that separates suggesting, drafting, selecting, approving, executing, repeating, and delegating, then let people grant only the lowest level needed for the current task
• Require consent renewal when the recipient, amount, data source, destination, duration, public visibility, legal consequence, or physical effect changes, even when the agent believes the new step serves the original goal
• Place a neutral action preview outside the conversational persona showing what will happen, which account and tool will be used, what data will leave, what it may cost, when permission expires, and how to cancel or undo
• Keep an accessible receipt connecting the user request, system recommendation, sources, uncertainty, permission, tool call, resulting change, correction, and rollback without preserving unnecessary intimate conversation forever
• Evaluate public-interest platforms with published measures for safety, viewpoint access, moderation appeals, privacy, accessibility, community health, concentration of power, funding independence, and the ability to leave with contacts and content intact