I Feel Terrible About This. What Can I Actually Do?

The next generation of AI shouldn’t just answer questions. It should help people turn concern into constructive action. There is a particular kind of question that increasingly arrives at the AI interface. It isn’t really a search query. It sounds more like: “I feel terrible about what’s happening in Palestine. Is there anything useful I can actually do online? Should I support Greenpeace? A humanitarian organization? Write my representatives? Donate somewhere? What would actually help?” A traditional search engine responds with a list. A chatbot responds with paragraphs. A modern enterprise automation platform might turn a request into a workflow. ServiceNow, for example, now describes AI agents that can use organizational context, make decisions, and execute workflows across enterprise systems. But there is another possibility. What if the system first tried to understand what the person actually wanted to accomplish? Not: “What organization should I donate to?” But: “What kind of action would make sense for this particular person?” That is a surprisingly different problem. ⸻ The Genie Problem Imagine asking an extremely literal genie: “I want to help.” The genie might hand you a donation page. That’s not necessarily wrong. But it isn’t necessarily useful either. Perhaps you have money but no time. Perhaps you have time but little money. Perhaps you want to contact elected officials. Perhaps you want to volunteer. Perhaps you’re a programmer who could contribute technical work. Perhaps you’re a student who wants reliable educational resources. Perhaps you simply want to understand the situation before doing anything. The useful answer therefore isn’t one recommendation. It’s a decision space. A good system might respond: Here are five kinds of action you could take. Donate: organizations currently accepting contributions for their stated programs. Volunteer: organizations currently seeking volunteers. Advocacy: current petitions, public-comment opportunities, or elected-representative contacts. Learn: primary-source reporting and educational material from multiple perspectives. Contribute skills: organizations seeking technical, translation, research, or other assistance. Then it could explain the tradeoffs. That’s a much more interesting use of AI. ⸻ The important word is auditable Suppose the system recommends Greenpeace. That recommendation shouldn’t materialize from some mysterious model intuition. The system should be able to tell you why. Greenpeace itself currently offers several forms of participation, including volunteering, online action, and donations. It also publishes its funding principles and says Greenpeace International does not accept funding from governments, corporations, political parties, or intergovernmental organizations. Those are facts. The system can present them. But it should distinguish those facts from its own interpretation: “Based on your stated interest in environmental advocacy, we think this organization may be relevant.” That sentence is an AI recommendation. The information supporting it is evidence. Those should never quietly become the same thing. ⸻ The Enterprise Automation Companies Already Know the Trick This is where the comparison with Salesforce, ServiceNow, and similar platforms becomes useful. The big enterprise systems have spent years solving a related problem: How do you turn a human request into a series of machine-executable actions? ServiceNow’s workflow tooling, for example, lets organizations construct multi-step workflows and connect systems together, while its current AI platform combines agents, data, workflows, and governance. That’s powerful. But imagine taking the same philosophy and applying it to questions about the world rather than questions about corporate operations. Instead of: “Reset this employee’s account.” you might ask: “I’m interested in supporting ocean conservation. What are the credible ways I could participate?” The machine doesn’t immediately execute something. First it researches. Then it compares. Then it identifies the available actions. Then it explains what it knows. Then it lets the human choose. That’s an important distinction. ⸻ Don’t Build Another Giant Enterprise Platform And here’s where Arc Codex takes a deliberately contrarian turn. You don’t necessarily need another $100-million enterprise suite. You might need one really good interface. A person types: “I want to help with X.” The system asks perhaps two or three sensible questions. Then it produces: YOUR OPTIONS 1. Donate Organizations and campaigns relevant to your stated goal. 2. Volunteer Current opportunities requiring your time or skills. 3. Advocate Specific actions available to citizens or members. 4. Learn Primary sources, research, reporting, and competing interpretations. 5. Contribute Ways you can help using skills you already possess. And beside every recommendation: Why am I seeing this? Click it. The machine opens its notebook. ⸻ The Notebook Is the Product That’s the part I find most interesting. Behind the friendly interface is the machinery we’ve been calling CLIP — the Claim Lifecycle & Integrity Protocol. The user sees: Greenpeace — suggested because you selected environmental advocacy. But underneath: SOURCE Greenpeace official website CLAIM Organization currently offers online actions, volunteering and donation opportunities. SOURCE DATE ... PROVENANCE Official organizational source CLASSIFICATION Environmental advocacy RELEVANCE High CONFLICTING INFORMATION None detected LAST VERIFIED ... RECOMMENDATION Generated by Arc Codex Suddenly AI recommendation becomes something you can inspect. That is a very different proposition from: “The AI thinks you should do this.” ⸻ And the System Should Be Able to Say No Sometimes the best answer is: “I don’t have enough reliable information to recommend an organization.” Or: “These organizations make similar claims, but their current activities are difficult to independently verify.” Or: “Three sources disagree. Here are the disagreements.” Or simply: “You might want to learn more before taking action.” That isn’t failure. That’s a safety feature. An information system that always produces an answer is eventually going to produce some very confident nonsense. An information system capable of saying “I don’t know” has a fighting chance. ⸻ From Radio to Research Assistant This is where the apparently strange Arc Codex architecture begins to make sense. The same machinery that produces a multilingual news radio station can produce a research assistant. The ten correspondents can monitor different territories. Local AI agents can continuously inspect sources. NLP systems can measure characteristics such as tone and sentiment. Deduplication can identify stories that are merely repeating one another. Translation can make material available across languages. The queue can preserve provenance. The audit system can examine the system’s own behavior. And the final interface can turn all of that machinery into something remarkably simple: What are you trying to accomplish? That’s the user interface. Everything else happens behind it. ⸻ The Smallest Useful AI Is Sometimes the Most Interesting There is a tendency in technology to build bigger systems because bigger systems sound impressive. But perhaps the more interesting future is the opposite. A system that does one thing extremely well: You tell it what you want to accomplish. It researches the available options. It explains them. It shows its sources. It exposes its reasoning and uncertainty. And then it lets you decide. No mysterious autonomous executive. No AI deciding what you believe. No giant recommendation engine quietly manipulating the destination. Just a very capable research assistant with a very good notebook. ⸻ The Real Product Maybe the product isn’t artificial intelligence at all. Maybe it’s agency. The internet has become extraordinarily good at telling people what other people are saying. AI is becoming extraordinarily good at summarizing it. The next step may be helping an individual answer a much older human question: “Given everything I now know, what can I actually do?” And then giving them a handful of credible choices. Not telling them which choice to make. Not pretending the machine knows the answer to every moral question. Just doing the research, showing the work, keeping the receipts, and putting the decision back where it belongs. With the human being. ⸻ A little thought experiment Try asking your favorite AI: “I feel strongly about something. What can I constructively do about it?” Then ask: “Show me exactly why you recommended those actions, where every factual claim came from, what you don’t know, and what alternatives you considered.” That second question may be the beginning of a rather different kind of AI. Not an oracle. A research assistant with an audit trail.