OpenAI Wants to Write the Rules It’s About to Break

A new OpenAI blog called AI Futures promises big thinking about power, governance, and freedom. It arrives from the lab least trusted to hold any of them.

The company that gutted its own nonprofit oversight in 2024 now wants to tell you how AI should govern the world.

That’s the subtext of AI Futures, OpenAI’s new blog exploring how transformative AI could reshape power, governance, the economy, and individual freedom. The essays are genuinely thoughtful. They’re also a bank robber publishing a paper on vault security. Both things are true at once, and the second one is why this launch matters more than the prose inside it.

What it actually is

Strip away the framing and AI Futures is a house-published series of policy and philosophy essays. No new model. No benchmark. No API change. It’s OpenAI staking a claim on the narrative of what comes after the models get scary-good — who holds power, who gets displaced, what freedom means when a machine can do most cognitive work.

Those are the right questions. OpenAI is not the only lab asking them, and it’s not the best at it. Anthropic has been doing this work with receipts for over a year. Its Economic Index ships actual anonymized data on how Claude is used across occupations — real numbers about which jobs are getting automated versus augmented, updated over time. Dario Amodei’s “Machines of Loving Grace” laid out a concrete, falsifiable vision of AI’s upside. Google DeepMind publishes a versioned Frontier Safety Framework with capability thresholds you can point to.

OpenAI’s contribution, so far, is prose. Well-written prose that arrives conveniently detached from data. That’s the tell.

The credibility problem nobody at OpenAI will name

Here’s what makes AI Futures land funny. This is the same company that:

Restructured to strip meaningful control from the nonprofit board that was supposed to keep it honest, converting the for-profit arm into a public benefit corporation with investors holding real leverage.
Watched its entire original safety leadership walk out the door — Jan Leike, Ilya Sutskever, and much of the Superalignment team, which was then dissolved after being promised 20% of compute it reportedly never got.
Used aggressive non-disparagement clauses that threatened departing employees’ vested equity, until public exposure forced a walk-back.

Now it’s publishing essays on individual freedom and the concentration of power. When the lab most responsible for concentrating AI power writes think-pieces warning about the concentration of AI power, that’s not thought leadership. It’s narrative pre-positioning ahead of regulation.

Head-to-head: who’s actually credible on this

If you’re going to read one lab’s take on AI’s political and economic future, read Anthropic’s, and it isn’t close.

Anthropic backs claims with published data (the Economic Index), a versioned Responsible Scaling Policy with hard capability triggers, and a track record of holding models back — it delayed capabilities and shipped ASL-3 protections around Claude Opus 4 when its own evals flagged uplift risk. You can argue with the conclusions, but the work is auditable.

Google DeepMind comes second: the Frontier Safety Framework is concrete and versioned, though the essays sit closer to research than public discourse.

OpenAI comes third on credibility despite frequently leading on raw capability. AI Futures is smart writing unmoored from mechanism. There’s no “here’s the data, here’s the threshold, here’s what we’ll do.” It’s vibes about governance from the company with the worst governance story in the industry.

Winner: Anthropic. Not because its models are always ahead — the GPT-5 and Claude Opus generations trade blows on SWE-bench Verified in the mid-70s, close enough that the coding-benchmark race is basically a tie — but because on the specific question AI Futures claims to own, only one lab shows its work.

What to watch in the next 3–6 months

This blog is a leading indicator, not an endpoint. Track these:

Regulatory timing. Expect AI Futures posts to align suspiciously well with EU AI Act enforcement milestones and any renewed US federal push. When a lab starts publishing on “governance,” a lobbying position is usually forming underneath it.
The agent authorization fight. As GPT-5-class agents get real permissions — spending money, sending email, touching production systems — the actual governance question isn’t philosophical. It’s OAuth. Who authorizes an AI agent to act as you, and how do you revoke it? Watch for OpenAI to tie AI Futures rhetoric to concrete identity and delegation tooling. That’s where “individual freedom” gets operational.
Anthropic’s counter. Expect a new Economic Index drop with harder displacement numbers. Data beats essays.
Whether OpenAI publishes anything falsifiable. If AI Futures ever ships a threshold, a commitment, or a dataset instead of a manifesto, take it seriously. Until then, read it as strategy.

Do something more useful than reading essays

You don’t have to wait for a lab to tell you how AI is reshaping your job. Measure it. Pull your own usage data and see what you’re actually offloading to a model versus what you still do yourself — that’s your personal economic index. Here’s a starting point against the OpenAI API:

import os
from openai import OpenAI

client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])

# Classify your last week of AI tasks: augmentation vs automation
tasks = [
    "drafted a client proposal",
    "wrote and shipped a unit test suite",
    "summarized a 40-page contract",
    "generated marketing copy I sent unedited",
]

resp = client.responses.create(
    model="gpt-5",
    input=(
        "For each task, label it AUGMENTATION (human still in loop) "
        "or AUTOMATION (no human review). Return JSON with a one-line "
        "reason each.\n\n" + "\n".join(f"- {t}" for t in tasks)
    ),
)

print(resp.output_text)

Run that weekly. The ratio of automation to augmentation in your own workflow tells you more about AI’s impact on your life than any lab’s blog about power and freedom. Then compare it against Anthropic’s public Economic Index to see where you sit versus your whole occupation.

The bottom line

AI Futures is worth reading — critically, the way you’d read a company’s sustainability report. The questions are real, some of the writing is sharp, and OpenAI does sit at the center of the story it’s narrating. That’s exactly the problem. The lab with the most power to shape AI’s political future is now shaping the conversation about it, and it’s the one whose own governance collapsed under the first real test. Read the essays. Trust the data. Those aren’t the same source.