Incentive-to-Outcome Forecast
Trace the incentive to predict the probable outcome before harm appears
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 5
- Confidence
- 98%
Tristan Harris argues that technological direction is not wholly unpredictable when the governing incentive is visible. Start with the prize an actor is optimizing, such as engagement, market dominance, or being first to a capability. Then identify the shortcuts and externalities that become rational under that reward. Extend those choices across every competing actor and ask what repeated behavior makes probable. The social-media example is an engagement incentive that, in Harris's account, rewarded attention capture while costs landed on users and society. He applies the same reasoning to AI labs racing for capability. The framework does not prove every forecast; it creates an early-warning hypothesis that can be tested and used to redesign incentives before delayed harms become entrenched.
Origin
Extracted from The Diary of a CEO
Core principles
- 01Incentives make some technological futures more probable than others
- 02A narrow success metric can create broad social costs
- 03Delayed harm does not make a system safe
- 04Changing the incentive can change the path
How to run it
- 1
Find the decision-maker
Identify the company, institution, team, or person whose repeated choices shape the system.
Pro tip Separate public messaging from the metric that actually determines success.
- 2
Name the governing incentive
State the reward, competitive threat, or loss the actor is trying to maximize or avoid.
Pro tip Use operational measures such as engagement, revenue, speed, or strategic advantage.
Watch out Good intentions do not cancel a stronger structural incentive.
- 3
Trace rational shortcuts
Ask which safety, quality, labor, or social costs become easier to ignore while pursuing the reward.
Watch out Do not assume an externalized cost will appear on the actor's own balance sheet.
- 4
Scale the behavior
Imagine every competitor responding to the same pressure and forecast the resulting system rather than one actor's intent.
Pro tip Look for collective-action problems where individually rational choices create a shared bad outcome.
- 5
Redesign the incentive
Move harms onto decision-makers, change the success metric, or establish shared constraints before the pattern becomes entrenched.
Pro tip Treat the forecast as a testable hypothesis and update it with evidence.
Watch out Waiting for conclusive long-term harm may remove the easiest intervention window.
In the wild
Harris says publishers and platforms measured success through attention and engagement. In his account, designs that increased scrolling succeeded even when addiction, loneliness, polarization, or other social costs were borne elsewhere. He argues that the incentive made those outcomes foreseeable rather than accidental surprises.
→ The incentive becomes an early warning for likely attention-capture harms.
Harris applies the model to AI companies that believe being second could mean permanent strategic defeat. He argues that this perceived winner-take-all prize rewards speed and makes safety, security, energy, and labor effects easier to discount.
→ The forecast points toward shared rules that alter the race rather than relying on voluntary restraint by one lab.
Common mistakes
Forecasting from stated intentions
An actor may care about people while still responding to a metric that rewards harmful behavior.
Treating the forecast as certainty
The incentive supports a probable-path hypothesis, not proof that every predicted outcome must occur.
Is it for you?
Best for
Leaders assessing platforms, AI systems, policies, or markets driven by strong competitive incentives.
Not ideal for
Situations where actor incentives are unknown or outcomes are largely determined by stable physical constraints.
From the transcript
“if you know the incentives you can actually know something about the future that you're heading towards”
“If you know the incentive which is for these companies AI to race as fast as possible”
From the episode
AI Expert: We Have 2 Years Before Everything Changes! We Need To Start Protesting! - Tristan Harris