AI 2040’s Missing Layer: Content Provenance in the Basin of Sanity

The Sequel Everyone Is Misreading

On July 9 the AI Futures Project — the team behind AI 2027, the scenario document that dominated AI discourse last year — published its sequel: AI 2040: Plan A. It reached Axios and the Washington Post the same day, and it will reach a conversation near you shortly, probably introduced with some version of: “the AI 2027 people have pushed their prediction back to 2040.”

That reading is wrong twice, and the document says so itself. First, Plan A is a recommendation, not a prediction — the authors’ own words. It is a policy proposal stress-tested by writing it as a concrete, dated narrative: a verified international slowdown of frontier AI development, enforced through radical research transparency. Second, the capability timeline inside the document barely moved. Human-range AI still arrives around 2030 in their telling; 2040 is merely the year scaling to superintelligence resumes in their preferred, slowed-down world. The authors put their own odds of Plan A being implemented at 3–15%. Their modal world is still the race.

I read the whole thing, supplements included, because Numonic builds content provenance infrastructure — and the document’s most compelling idea quietly depends on a layer it never names. That gap is worth walking through, because it is the layer working studios, agencies, and platform builders can actually do something about — starting now, regardless of what happens to anyone’s treaty.

What Plan A Actually Proposes

The mechanics, compressed: the US and China agree in 2029 to stop racing to superintelligence. Enforcement rests not on trust but on total transparency of frontier AI R&D — mutual auditors inside each other’s datacenters, compute accounting precise enough to make covert projects detectable. The authors call it “mutually assured compute destruction.” The world scales AI within the human range through the early 2030s, pauses at top-human-expert level around 2035, and resumes the climb to superintelligence around 2040 with alignment research having had years instead of months.

The document compares this against four rival strategies — racing at maximum speed, sabotaging the competition, domestic regulation alone, or shutting everything down — and scores each on how much safety margin it buys. Whatever you think of the treaty’s feasibility (the authors themselves are sceptical), the interesting move for builders is structural: every plan they consider stands or falls on transparency infrastructure. Verification, disclosure, auditability — the load-bearing mechanism is always the same, only the scope changes.

The Best Idea in the Document: the Basin of Sanity

Tucked into the supplements is the concept I found most durable: the basin of sanity. The argument runs like this. As AI floods the information environment, society lands in one of two self-reinforcing loops. In the vicious loop, degraded collective reasoning leads people to adopt sycophantic, agenda-pushing AI — which degrades reasoning further. In the virtuous loop, sound epistemics lead to the adoption of truth-seeking AI tools, which strengthen collective reasoning. The basin of sanity is the region where society remains “sane enough to make itself more sane.”

The supplement’s toolkit for staying in the basin is genuinely thoughtful: AI-generated community notes, track-record scoring for public figures, calibrated AI forecasters, shared knowledge bases “trusted by a wide range of actors” to prevent society from fracturing into alternate realities, and reliable ways to know whether you are talking to a human or a machine.

Notice what all of those have in common: they are institutional epistemics. They operate on claims, arguments, forecasts, and reputations. And that matters because of what the document predicts elsewhere about the world those institutions will inhabit.

The Missing Layer

Here is the strange part. By the document’s own 2027 and 2028 chapters, companies are spending ten billion dollars a month on fleets of AI agents that can do anything on a computer, and white-collar work broadly is experiencing what software engineering experienced in 2026. In that world — their world, never mind ours — the overwhelming majority of content a human encounters is AI-generated or AI-touched. Yet the epistemics supplement makes no specific predictions about synthetic media, and none of its proposals operate at the level of the content itself.

Institutional epistemics assumes you can trust the substrate. A calibrated AI forecaster citing fabricated sources is worse than no forecaster, because the calibration launders the fabrication. A community note attached to an image nobody can trace is an opinion, not a correction. Track-record scoring — their own proposal — requires knowing reliably who said what, when. That is an attribution problem, and attribution at scale is precisely what provenance infrastructure does.

The content layer has to answer three questions, and they are questions we work on daily:

• Where did this come from? Origin and identity — which tool, which account, which organisation produced this asset.
• How was it made? The creation record — model, prompt, parameters, workflow. The full generative context, not a “made with AI” sticker.
• What happened to it since? Lineage — edits, derivations, exports, and every hop across tools where metadata traditionally goes to die.

None of this is speculative. C2PA content credentials, IPTC metadata, and cross-tool lineage graphs exist and ship today. They are unevenly adopted and imperfectly durable — we have written before about the three layers of content trust and where each current standard falls short. But the foundations are real, which is more than can be said for most of the infrastructure in any 2040 scenario.

The Transparency Act That Already Exists

Plan A’s 2027 chapter imagines Congress passing an “AI Transparency Act of 2027” — an omnibus bill forcing frontier labs to publish model specifications and internal usage statistics. Model-layer transparency, aimed at the labs.

What the scenario does not mention is that the content-layer twin of that act is already law. Article 50 of the EU AI Act becomes enforceable on August 2, 2026 — twenty days after this post. It requires machine-readable marking of synthetic content, disclosure of deepfakes, and telling people when they are interacting with an AI system. We have covered what Article 50 requires and what remains unsettled in detail; the short version is that the direction of travel is unambiguous even where the operational guidance is not.

To be precise about what Article 50 is not: it is a marking and disclosure rule, not a full provenance mandate. It does not require creation records or lineage. But machine-readable marking is the regulatory on-ramp to provenance infrastructure — once content must carry a durable, machine-readable signal of its synthetic origin, the tooling that captures, preserves, and verifies that signal stops being optional. The scenario writers imagine transparency regulation arriving top-down, in Washington, at the model layer, in 2027. It is actually arriving bottom-up, in Brussels, at the content layer — this summer.

Where This Leaves a Builder

I run an AI-native company and spend my days closer to the deployment coalface than to the treaty table, so I will offer the operating view. The strongest published critique of AI 2040, from Richard Ngo, points at what he calls the capabilities-impact gap: benchmark scores keep soaring while measured real-world impact lags well behind. From inside an AI-leveraged engineering organisation, I can report the gap is real — and that it is made of trust, not capability. The friction I fight daily is verification, provenance, and accountability for machine output. Not whether the model can do the work, but whether anyone can safely rely on what it did.

That diagnosis is why AI 2040 changed nothing about what we build and something about our conviction in it. Consider the two branches honestly. In the fast world — theirs — synthetic media volume and transparency regulation both explode, and provenance becomes load-bearing public infrastructure. In the slow world — Ngo’s — enterprises adopt generative tooling piecemeal over a decade, and every step of that adoption needs the same audit trails, disclosure records, and lineage. Content provenance is one of the few assets that compounds in both branches. You do not need to bet on a timeline to build it; you need only bet on the transition itself.

Concretely, our lane today: capturing the creation record at ingest — prompts, models, parameters, and workflows from tools like ComfyUI and Midjourney; reconstructing lineage across tools; reading and detecting C2PA credentials where they survive; and producing the compliance-ready exports that Article 50 workflows will lean on. This month that includes running our own Article 50 readiness audit, which we will write up honestly — including what tools like ours can and cannot yet do. Deeper C2PA integration, including signing on export, is specified and on our near-term roadmap. When we say “roadmap,” we mean it in the honest tense — not the marketing one.

What to Take From AI 2040, Even If You Never Read It

Five things worth carrying out of the document and its reception:

1. Correct the misreading. Plan A is a recommendation with self-assessed 3–15% odds of adoption, not a forecast. The capability timeline inside it still points at roughly 2030 for human-range AI.
2. Steal the method, not the conclusion. The authors call it scenario scrutiny: most plans fall apart when forced through a concrete, dated narrative. It works just as well on a product roadmap as on geopolitics. Write the story in which your plan fails, with dates.
3. Watch the transparency agenda, not the treaty. The treaty is a long shot by its authors’ own numbers. The incremental asks — disclosure, marking, verification, compute accounting — are live policy material on both sides of the Atlantic, and Article 50 is the first enforcement of the content-layer half.
4. Build what survives both branches. Fast takeoff or slow grind, the transition needs attribution, lineage, and audit infrastructure. Assets that compound in every plausible world are rare; prefer them.
5. The basin of sanity needs a floor. If the epistemic infrastructure the scenario calls for ever gets built, content provenance is a load-bearing wall — and unlike most of Plan A, it does not require a treaty. It requires standards, tooling, and people willing to ship the unglamorous layer first.

The AI Futures Project wrote a book’s worth of scenario and supplements about how humanity might stay sane through the transition. The content layer got a parenthesis. That footnote is our day job — and after reading theirs, I am more convinced than ever that it is the right one.

— Jesse M. Blum, co-founder and CTO, Numonic Labs