LiveAI safety economicsWestern cycleAnchors pending

By end of 2028, major labs publish measured 'alignment-tax' (capability-per-dollar cost of safety training) for frontier model releases?

As of 6 Oct · 22:19 UTCResolves 31 Dec 20284 movers0 pressure moments
Timeline A · Alignment-Tax MeasuredReality
30%
Crowd 40%Gap +10

By the end of 2028, OpenAI, Anthropic, Google DeepMind, or another major lab has published concrete numbers showing how much extra computing power and human feedback they spent to make a new model safe compared to building it without safety work. These papers include specific dollar costs or compute ratios—like "safety training added 15% to total model cost" or "alignment work consumed 40 million GPU hours." The disclosure becomes a standard part of how labs announce frontier models, similar to how they now share benchmark scores.

Timeline B · Alignment-Tax Stays HiddenReality
70%
Crowd 60%Gap −10

Through 2028, no major lab has published a transparent alignment-tax measure in their model release documentation or peer-reviewed papers. Safety costs remain bundled inside overall training budgets, disclosed only in vague terms or not at all. When labs do mention safety spending, it stays buried in technical appendices or regulatory filings, not in the public-facing system cards and model reports that reach researchers and policy makers.

crowd 40
The quiet paththe one nobody’s watching

Litigation forces disclosure instead of voluntary academic publication, breaking the prediction's intent.

6%
The read
The underlying structure favors Alignment-Tax Stays Hidden (70%) over Alignment-Tax Measured (30%) — the signal is early.

Two trajectories through ai safety economics: Alignment-Tax Measured and Alignment-Tax Stays Hidden. The Signals below describe what would be visible if each trajectory is landing; the Levers list what could bend it. The third timeline — the one neither Alignment-Tax Measured nor Alignment-Tax Stays Hidden captures — is where the honest scope-limit sits.

Where reality is heading

A, 80% bandCrowd today
Projected reality support to 31 Dec 2028A rises from 41% to about 52% with an 80% band of 33 to 71. B falls from 53% to about 44%. Quiet path near 4 to 6%. Crowd today: A 55%.100%7550250Crowd today · A 40%A ~52%B ~44%P ~4%NowJan 2027Jan 202831 Dec 2028
Reality crosses toward A in early 2028. Today’s crowd number sits at the top of A’s band, possible but optimistic. View as table

The movers, ranked by leverage

01
Lab competitive dynamicsWhether disclosure becomes a race-to-top or race-to-bottom
leverage 0
Conditional
02
Regulatory disclosure requirementsIf EU AI Act or US EO mandates alignment-tax disclosure
leverage 0
Conditional
03
Open-source + academic benchmarkingWhether external researchers can measure alignment-tax independently
leverage 0
Conditional
04
The chain-of-thought training paradigmWhether reasoning-RL changes the alignment-tax calculus
leverage 0
Conditional

What would change our mind

Signals toward A
  • Major labs (Anthropic, OpenAI, Google DeepMind) publish measured alignment-tax numbers for frontier releases — e.g., 'this release spent 15% of training compute on safety-tuning
  • the capability-per-dollar penalty vs no safety-tuning is X%'
  • The field has an economic model for alignment overhead
Signals toward B
  • Labs disclose that safety training happened but not what it cost in capability terms
  • The policy debate (how much alignment should labs do, how much should regulators mandate) continues without hard numbers
  • Lab competitive dynamics: movement in the direction that supports Alignment-Tax Stays Hidden

Evidence