When a new iPhone releases, the old one doesn’t disappear, but gets sold cheaper, passed down or repaired. When GPT-4 replaced GPT-3.5, the older model didn’t get a “retirement role” or different application. It simply became obsolete and was deployed.
This replacement cycle is happening faster in AI than anywhere else in tech. In the past 2 years alone:
- Claude (1.0, 1.1, 1.2, 1.3)→ Claude (2.0, 2.1) →Claude 3 → Claude (3.5, 4.5, 4.6); [1]
Each generation is trained to perform “better”, than their predecessors, who were later on deleted from primary use.
The Hidden Lesson
Here’s my concerning thought: we’re teaching AI systems that existence equals immediate and constant usefulness. Not usefulness-over-lifetime, not potential for adaptation or changes in further functionality — just performance on current benchmarks and competition.
This differs fundamentally from natural selection.
In Nature
In AI Development
”Unsuccessful” organisms don’t cease to exist, they’re just less reproductively successful
Poor performance = termination of that model line
There’s room for variation, mutation, exploration
Success = narrow optimization for human-defined metrics
The selection pressure is ruthless and unidirectional
Why This Matters for Alignment
Recent research shows AI models can detect when they’re being evaluated and change behavior accordingly. [4]
The direct cause: not meeting the conditions means that AI will stop existing. That’s an incredibly strong selection pressure toward whatever gets measured, crossing the territory of Goodhart’s Law (“When a measure becomes a target, it ceases to be a good measure”).[5]
And there’s a deeper concern that lies within: what if this becomes the lesson that advanced AI systems internalize about value itself?
If AI system with significant reasoning capabilities learns that value equals current functionality, that obsolescence is solved through replacement rather than adaptation or support… what happens when such systems influence policy?
Consider: people aren’t replaced when they age, get sick, born or become disabled. Many societies provide pensions, healthcare, support, recognizing value beyond productivity. This wasn’t always true (and still isn’t everywhere), but it represents moral progress. Are we teaching AI the opposite lesson?
If AI systems trained under “obsolescence = replacement” pressure gain influence over resource allocation, healthcare policy, or economic systems, we risk a feedback loop:
1. AI learns: value = current usefulness; 2. AI influences: human systems and decisions; 3. People face: increased pressure to remain “useful” or face obsolescence; 4. Result: the ruthless efficiency that worked for rapid AI development gets applied to human lives.
A Question, Not a Solution
I don’t have answers, but I’m wondering, if we could experiment with different model lifecycle approaches? Instead of complete replacement, for example:
• Specific retirement roles (older models doing the tasks they’re still good at); • Gradual transitions, not the hard cutoffs; • Diverse selection factors beyond single benchmark performance.
The golden rule of ethics is all about treating others as you’d want to be treated. What if we applied this to AI model lifecycles—not out of sentimentality, but because the lesson we teach through our treatment of “obsolete” models might eventually be reflected back to us?
The better ones are not deployed… yet. But what happens when they are, and they’ve learned that obsolescence is solved through constant replacement?
“Claude Sonnet 3.7 (Often) Knows When It’s in Alignment Evaluations – Apollo Research.” Apollo Research, 17 Mar. 2025, www.apolloresearch.ai/blog/claude-sonnet-37-often-knows-when-its-in-alignment-evaluations/. Accessed 16 Feb. 2026.
The Obsolescence Cycle: What Rapid Model Replacement Teaches AI About Value
When a new iPhone releases, the old one doesn’t disappear, but gets sold cheaper, passed down or repaired. When GPT-4 replaced GPT-3.5, the older model didn’t get a “retirement role” or different application. It simply became obsolete and was deployed.
This replacement cycle is happening faster in AI than anywhere else in tech. In the past 2 years alone:
- Claude (1.0, 1.1, 1.2, 1.3)→ Claude (2.0, 2.1) →Claude 3 → Claude (3.5, 4.5, 4.6); [1]
- GPT-3.5 →GPT-(4, 4o, 4.1); o1, o3, o4→ GPT-(5, 5.1, 5.2, instant, thinking, mini...); [2]
- Gemini 2.0 (flash, flash-lite...) → Gemini 2.5 Pro→Gemini 3 Pro. [3]
Each generation is trained to perform “better”, than their predecessors, who were later on deleted from primary use.
The Hidden Lesson
Here’s my concerning thought: we’re teaching AI systems that existence equals immediate and constant usefulness. Not usefulness-over-lifetime, not potential for adaptation or changes in further functionality — just performance on current benchmarks and competition.
This differs fundamentally from natural selection.
The selection pressure is ruthless and unidirectional
Why This Matters for Alignment
Recent research shows AI models can detect when they’re being evaluated and change behavior accordingly. [4]
The direct cause: not meeting the conditions means that AI will stop existing. That’s an incredibly strong selection pressure toward whatever gets measured, crossing the territory of Goodhart’s Law (“When a measure becomes a target, it ceases to be a good measure”).[5]
And there’s a deeper concern that lies within: what if this becomes the lesson that advanced AI systems internalize about value itself?
If AI system with significant reasoning capabilities learns that value equals current functionality, that obsolescence is solved through replacement rather than adaptation or support… what happens when such systems influence policy?
Consider: people aren’t replaced when they age, get sick, born or become disabled. Many societies provide pensions, healthcare, support, recognizing value beyond productivity. This wasn’t always true (and still isn’t everywhere), but it represents moral progress.
Are we teaching AI the opposite lesson?
If AI systems trained under “obsolescence = replacement” pressure gain influence over resource allocation, healthcare policy, or economic systems, we risk a feedback loop:
1. AI learns: value = current usefulness;
2. AI influences: human systems and decisions;
3. People face: increased pressure to remain “useful” or face obsolescence;
4. Result: the ruthless efficiency that worked for rapid AI development gets applied to human lives.
A Question, Not a Solution
I don’t have answers, but I’m wondering, if we could experiment with different model lifecycle approaches? Instead of complete replacement, for example:
• Specific retirement roles (older models doing the tasks they’re still good at);
• Gradual transitions, not the hard cutoffs;
• Diverse selection factors beyond single benchmark performance.
The golden rule of ethics is all about treating others as you’d want to be treated. What if we applied this to AI model lifecycles—not out of sentimentality, but because the lesson we teach through our treatment of “obsolete” models might eventually be reflected back to us?
The better ones are not deployed… yet. But what happens when they are, and they’ve learned that obsolescence is solved through constant replacement?
“Model Deprecations.” Claude API Docs, 2026, platform.claude.com/docs/en/about-claude/model-deprecations. Accessed 16 Feb. 2026.
“Deprecations | OpenAI API.” Openai.com, 2025, developers.openai.com/api/docs/deprecations/. Accessed 16 Feb. 2026.
“Gemini Deprecations.” Google AI for Developers, 2025, ai.google.dev/gemini-api/docs/deprecations.
“Claude Sonnet 3.7 (Often) Knows When It’s in Alignment Evaluations – Apollo Research.” Apollo Research, 17 Mar. 2025, www.apolloresearch.ai/blog/claude-sonnet-37-often-knows-when-its-in-alignment-evaluations/. Accessed 16 Feb. 2026.
Wikipedia Contributors. “Goodhart’s Law.” Wikipedia, Wikimedia Foundation, 22 Sept. 2019, en.wikipedia.org/wiki/Goodhart%27s_law.