Make human input
retrievable.

Every serious conversation about AI right now is an argument about the machines. One side says they're becoming too powerful; the other says they're still too limited. Both sides are staring at the technology, waiting for it to tell us what happens to us.

We think that's the wrong argument, about the wrong subject. Nothing about the machines - their power or their limits - decides whether human beings stay relevant. That gets decided somewhere much less dramatic: in product decisions, defaults, and interaction patterns, thousands of them, shipping every week.

The arrow of reliance points one way - humans depending on AI, never the reverse - not because AI overpowered us, but because the people building this future quietly stopped valuing what people contribute to it.

basedIRL exists because that decision is still being made, and it doesn't have to go the way it's going. This document explains what we think the actual problem is, and what we built about it.

The danger was never that AI becomes too powerful. The danger is that humans become powerless by proxy - not defeated, not replaced in some dramatic confrontation, just omitted. Left out of the architecture by a thousand small, individually reasonable decisions, each one treating human input as a cost to cut, a latency to optimize away, a checkbox to satisfy.

Nobody is choosing a future where human capability doesn't matter. Nobody would. We're getting that future anyway, by default, because relevance is structural: a system values you when it does materially better with you than without you - when there is a slot shaped like a person and value flowing through it.

Right now those slots are being removed, one release at a time, and the systems keep working, and their smoothness becomes the argument for removing the next one.

That is what designing human relevance out of the future actually looks like. Not a robot uprising. A changelog.

Here is the mechanism, concretely.

AI-generated content is training the next models. Those models generate more content. That content becomes the reference material the next answers are drawn from and cited against. Step by step, the information supply chain is closing in on itself - an internet that increasingly cites itself, answering instantly from an ever-larger archive of its own output, with less and less fresh human experience entering the loop anywhere.

And the whole arrangement rests on a pretense almost nobody says out loud: that the efficiency of instant answers from this closed loop outweighs any value a human could add. Any. The person who was there this morning. The operator who's run the process for a decade. The five people who could tell you in three minutes that the premise of your question is wrong. All of it, waved off - not because it was weighed and found wanting, but because it was never priced in at all.

Instant is not the same as informed. An answer in two seconds from a loop that's eating its own tail is still worse than an answer in four minutes from someone who actually knows - and the flood of instant everything is drowning people in exactly the kind of information that no one stands behind.

Now the part that should be obvious and somehow isn't: AI's efficiency is what finally makes human input affordable.

There's a reason human-powered answers died the first time. In the ChaCha and Ask Jeeves era, humans were the entire engine, and the engine was too slow. Waiting on a person for everything couldn't survive contact with search.

That constraint is gone. AI now does everything else in seconds - the retrieval, the drafting, the synthesis, the ninety percent that never needed a human. Which means the few minutes a real person takes to answer the one part that does is no longer a bottleneck. It's a rounding error, purchased out of an enormous speed surplus.

Nobody is capitalizing on this. The entire industry is spending the surplus on more instant - more generated answers, faster, about everything - and none of it on the one input the loop cannot generate for itself.

We built the patterns that spend it differently: answers that arrive fast but stay provisional until humans weigh in, and get visibly better when they do. Outreach you approve first and track like a package. Quick directional polls of real operators. Human answers entering the work as cited, verifiable evidence. Live human input, made as retrievable to an AI as a web search.

This is not a thought experiment. basedIRL is a working system: a product layer that makes live human input on anything retrievable by any AI.

An AI mid-task can plan for human input and ask permission before anyone is contacted - consent as an interaction, not a checkbox. It can route a question to real people nearby or to the right operators anywhere, dispatch it, and track it: sent, delivered, seen, answered.

It can run a quick poll when direction matters more than depth. And when a person answers, that answer doesn't vanish into a training run - it enters the work as first-class evidence, cited to a named human, with a verification code that proves they said it, disagreements preserved instead of averaged away.

Because it speaks the open protocols agents already use, this isn't an app you switch to. The same primitives render natively inside the products people already live in. Fresh human input, entering the loop again - on purpose, with provenance.

That question gets asked more every day, and almost every answer on offer is either denial, doom, or a platitude. Notice what's rare: an answer you can build.

This is one. And it is deliberately practical. There is no capability penalty in building systems that can reach live human input when they need it. The real failure mode is never building that option at all.

It doesn't require believing machines will stay limited, or pretending people are better search engines. It requires only the one thing the default future is missing: deciding, on purpose, that human input is worth keeping retrievable.

The conventions of the agentic web are hardening right now - what an AI is allowed to reach for while it works on your behalf. Every primitive in that list today is a machine. If people aren't made a first-class primitive in this layer now, nobody will retrofit us in later; you don't add a dependency on humans to a stack optimized end-to-end for their absence. This is the window.

Build AI that relies on us, so that we are not merely reliant on it. Mutual reliance is a stabler equilibrium than dependence - and unlike most things said about the future of AI, it's a design decision, not a prediction.

Vision documents usually hide the hard parts. Ours are the hard parts, so here they are.

First: convincing AI builders to add people to their stack. Every founder is optimizing for lower latency and lower cost, and we are proposing a dependency that is slower than an API and more expensive than inference - asking them to believe that grounded beats instant for the queries that matter. Some products will get it immediately. Most won't, yet.

Second: the people. A network where AI can ask humans is only as good as the humans who answer. Recruiting people to be findable and queryable - reliably, for fair pay, with real protections - is a cold-start problem stacked on a trust problem stacked on a novelty problem. We're asking people to do something nobody has done before: be a professional source of ground truth for machines.

Third: the narrative itself. Most people currently choose between excitement and dread. We're offering a third position that requires more thought than either: an intentional approach to keeping humans relevant by building their contribution into the stack itself.

We're telling you this because the beta is not a product launch - it's a search for the people who read those problems and lean in. Every network that mattered started exactly here: implausible, two-sided, and obvious only in retrospect.

We are not launching to everyone. We're looking for a specific kind of early believer:

Founders and AI developers who want their agents to return something the closed loop can't produce - a cited, verifiable answer from a real person, retrieved mid-task. People who know things willing to be among the first humans that machines learn to ask. And anyone who reads this and recognizes what's actually at stake: that human relevance is currently being designed out of the future by default, and that the defaults are still ours to set - for a very short while.

That future doesn't happen by accident. It has to be built on purpose. We've started.