Can human-AI research teams improve faster than autonomous AI systems?

The dominant framing of AI progress puts autonomous self-improvement at the center — models that can improve themselves without human involvement. But co-improvement — collaboration between human researchers and AIs to achieve co-superintelligence — may be both faster and safer.

The historical evidence: every major AI paradigm shift required a tandem of data innovation and method innovation, both discovered through significant human effort with many wrong directions:

• ImageNet + AlexNet (curated data + architecture)
• Web data + scaled transformers (data collection + model scaling)
• Instruction-following data + RLHF (labeling + training objective)
• Verifiable reasoning tasks + RLVR (task curation + training method)

Each tandem took human researchers significant effort, including dead ends and intermediate results. Co-improvement with AI systems built to collaborate should accelerate finding the unknown next paradigm shifts.

Three advantages over autonomous self-improvement: (i) faster paradigm discovery — human intuition about what matters combined with AI's ability to explore solution spaces, (ii) more transparency and steerability — human involvement creates checkpoints where misalignment can be detected and corrected, (iii) human-centered safety — the system is designed around human needs by construction, not by post-hoc constraint.

Since What limits how much models can improve themselves?, co-improvement sidesteps the gap by using humans as external verifiers. The generation-verification gap limits pure self-improvement; it does not limit systems where humans provide the verification signal.

Since Does incremental AI replacement erode human influence over society?, co-improvement explicitly preserves implicit alignment (claim 2 in the disempowerment thesis) by keeping human researchers in the loop. The disempowerment thesis predicts what happens when humans are removed; co-improvement is the architectural choice to keep them in.

The practical agenda: measuring AI research collaboration skills with new benchmarks covering problem identification, data/benchmark creation, method innovation, experimental design, and evaluation — then training to improve those benchmarks specifically. This is What capabilities do AI systems need for autonomous science? reframed from an autonomy checklist to a collaboration skill inventory.

Source: Human Centered Design