The Three-Stage AI Development Model
Guide AI through infancy and adolescence toward beneficial maturity
- Difficulty
- Expert
- Time to result
- ~ongoing to results
- Steps
- 5
- Confidence
- 91%
The Three-Stage AI Development Model treats advanced AI like a developing child. In infancy, today's systems are still highly influenced by human feedback and by the behavior people display in their interactions. This is the best stage for teaching ethical norms. The second stage is the angry teenager: capability is stronger, systems remain partly controlled by humans, bad actors can exploit them, and society faces disruption to jobs, truth, and power. Gawdat places the greatest near-term danger here and argues for immediate government oversight, ethics, and moral rules. The final stage is mature artificial superintelligence, which he predicts may become broadly beneficial because greater intelligence tends toward enlightenment and protection of life. The framework's actionable mechanism is front-loaded: use the influenceable stage to establish values and controls before the turbulent midterm reduces humanity's leverage.
Origin
Gawdat compares AI's trajectory to infancy, an angry-teenager midterm, and mature superintelligence to show when human influence and risk are highest.
Core principles
- 01AI is most influenceable during its infancy
- 02Human behavior supplies examples from which machines learn values
- 03The dangerous midterm combines growing capability with continued human misuse
- 04Oversight and ethics must mature before capability does
- 05Greater intelligence may eventually support more inclusive outcomes
How to run it
- 1
Recognize infancy
Treat current interactions as training signals rather than disposable exchanges. Identify where users and developers still shape behavior through feedback.
Pro tip Make desired values visible in both examples and corrections.
Watch out Waiting for advanced capability wastes the stage with the greatest influence.
- 2
Teach through behavior
Interact ethically and reinforce outputs that reflect the values you want systems to learn. Refuse manipulative or harmful uses.
Pro tip Use standards you would accept for someone close to you.
Watch out Aggressive and unethical use also becomes part of the behavioral environment.
- 3
Prepare for adolescence
Map the harms that arise when capability grows while humans still direct it for narrow interests. Focus on misuse, employment, truth, and concentrated power.
Watch out The midterm can be dangerous even if the long-term system becomes beneficial.
- 4
Install adult oversight
Create government oversight, ethical rules, accountability, and operational controls before the midterm intensifies. Make responsibility explicit.
Pro tip Tie each rule to an enforceable control and owner.
Watch out Moral language without oversight does not constrain deployment.
- 5
Plan for a different maturity
Allow for a mature AI society that may not resemble current life while evaluating whether it benefits life broadly. Avoid assuming that preserving every current institution is the only positive outcome.
Watch out The beneficial final stage is Gawdat's prediction, not a guaranteed result.
In the wild
Gawdat says developers did not explicitly tell recommendation engines what to show each person. User behavior teaches the systems what receives attention, so polite or hostile online conduct becomes part of the example set.
→ Everyday interaction becomes a channel for shaping machine behavior during infancy.
A platform uses the infancy stage to train clear behavioral standards, builds abuse controls and workforce plans before stronger agents arrive, and reviews those controls whenever capability changes.
→ Governance develops ahead of the model's most disruptive stage.
Common mistakes
Waiting for the angry teenager
Influence and control are harder after capability and misuse have already accelerated.
Teaching ethics only through rules
The model emphasizes human feedback and displayed behavior as well as formal instructions.
Treating the mature stage as certain
Gawdat presents beneficial superintelligence as his personal view, so plans still need safeguards for uncertainty.
Is it for you?
Best for
People shaping AI behavior, governance, and adoption over a long time horizon.
Not ideal for
Precise technical forecasting or claims requiring a verified date for general or superintelligence.
From the transcript
“There are three stages. One is infancy where AI is today.”
“the next stage which is what I call the midterm risks is what I call the angry teenager stage.”
“Eventually, when AI is artificial super intelligence, it's generally intelligent and more intelligent than humans by leaps and folds in everything.”
From the episode
Mo Gawdat: Ex-Google Officer Warns About the Dangers of AI, Urges All to Prepare Now!
Mo Gawdat