Why speedometers don't steer
Public AI rankings measure how capable models are, not whether people can still oversee them. Our Observatory tracks both, and publishes its forecasts so they can be scored against what actually happens.
Leaderboards and capability benchmarks report how much AI systems can do and how fast that is growing. They say nothing about whether the institutions using those systems can see what they do and stop them when needed. A dashboard with only a speedometer tells you how fast you are going, not whether you still have control.
The single dial on the public dashboard
Every few months, software companies and benchmark trackers release new charts. The lines sweep upward across benchmark scores, parameter counts, and task duration. We are told that a new model solves math competition problems twice as fast, or completes a five-hour coding task without human intervention.1
These numbers represent real engineering achievements. But they measure only horsepower and acceleration. What they leave out is the second dial on the dashboard: the steering wheel and the brakes.
Can a human supervisor see what the model did before its decisions take effect? Does the assistant leave a tamper-proof receipt for every file it altered? Can a city council pull an emergency lever to halt an automated workflow without crashing local public services?2
When public rankings ignore those questions, they treat raw speed as if it were the only property that matters.
The widening oversight gap
When an engine grows more powerful while the steering mechanism remains unchanged, danger compounds with every mile. In technology governance, we call the distance between these two dials the oversight gap.
On one side, autonomous software executes longer chains of actions with less supervision. On the other side, human inspection remains slow, fragmented, and underfunded. AI agents' success rates fall as tasks get longer, so a system that is reliable on short tasks can fail on long ones.1 By the time an inspector notices a deviation, twenty dependent actions have already executed.
Relying on an AI assistant to monitor its own mistakes does not fix the problem. An algorithm cannot impartially grade its own work. If a system has the authority to act, its safety guarantees must be enforced by independent external mechanisms.3
How an honest observatory works
An observatory should never be a marketing department or a panic horn. Its purpose is to provide dependable navigation tools for communities, researchers, and public officials.
To provide clear instruments, our Observatory follows three straightforward commitments:
First, we track both curves side by side. We record advances in model reasoning and task execution duration. But we give equal prominence to indicators of human control: verifiable action receipts, decentralized data ownership, and democratic appeal procedures. If model speed surges while human oversight tools stagnate, our public charts display that divergence immediately.
Second, we score our forecasts in daylight. Anyone can issue vague warnings or bold claims about the future. An honest observer writes down concrete forecasts with exact dates, unambiguous resolution criteria, and probabilistic confidence intervals before events unfold.4 We score these predictions under standard scoring rules. When our estimates miss the mark, we preserve the record rather than rewriting history.
Third, we anchor every observation in immutable floor commitments. No amount of speed or technical elegance justifies ignoring personal consent, scoring human worth, or denying due process. Our nineteen constitutional floor commitments define non-negotiable boundaries that no technological emergency can override.
Keeping human hands on the steering wheel
We measure AI so that powerful tools stay accountable to the people whose lives they affect.
A locomotive needs an engine, but it also requires steel rails, signal lanterns, and reliable air brakes. Without those safety structures, an engine is merely an unguided projectile. Before communities delegate healthcare triage, public budgets, or legal administration to automated systems, they have a right to inspect both dials.
We invite you to explore our live Observatory dashboard, review our nineteen floor principles, or inspect our verified governance protocols in the Shared Horizon library.
Related essays: When pricing bots bid on public contracts · Paying for words instead of the repair. Further reading: Why long errands go off the rails · How to tell if an AI assistant works for you · Why an assistant shouldn't tap your shoulder for every small question.
1. Thomas Kwa, Ben West, Joel Becker et al. (METR), Measuring AI Ability to Complete Long Tasks, METR Research Report (March 2025; arXiv:2503.14499).
2. Harleen Kaur Sidhu, Rebecca Scholefield, Nour Annan et al., Open Problems in AI Incident Governance, arXiv:2607.05163 (July 2026).
3. Yuhang Wang, Why LLM Agents Collapse Without Oversight: The Enforcement Gap as the Mechanism Behind Emergence World Failures, arXiv:2609.15293 (September 2026).
4. Lukas Nel, Do Large Language Models Know What They Don't Know? KalshiBench: A New Benchmark for Evaluating Epistemic Calibration via Prediction Markets, arXiv:2512.16030 (December 2025).
