"Nothing in biology makes sense without teleology" by Michael Levin

#1

O3:

1. Tidied transcript

(Filler words, false starts and repeated phrases removed; timestamps omitted; paragraphs grouped by idea. Ellipses mark obvious short omissions where the speaker paused or digressed without changing the sense.)

If you want to download any of the primary papers, data or software, everything is free and open on our website. My personal blog explains what I think some of these findings mean.

I chose the title of this talk as a nod to Theodosius Dobzhansky’s line “Nothing in biology makes sense except in light of evolution.” Dobzhansky was pointing to the power of unifying principles. Before we had a theory of electromagnetism we thought lightning, magnets and light were unrelated. Once the spectrum was unified we could build technology that operates in parts of it our sense organs do not reach.

My lab—about forty-three people—works on interventions for birth defects, regenerative medicine, cancer, bio-engineering and some AI, but the deep question underneath is: How do minds exist in the physical world? What is embodiment? We use those ideas to guide both computational and wet-lab work.

I want a single scale-independent framework for intelligence. That includes birds and primates, whales and octopuses, but also swarms, synthetic life forms, AIs, perhaps exobiological agents and even purely informational patterns… Rosenblueth, Wiener and Bigelow already sketched such a spectrum in 1943.

Unification matters for two reasons. First, practical progress in medicine and bio-engineering. Second, ethics. Our current moral categories are based on ancient language that tracks nothing real in the modern world; we will share Earth with cyborgs, hybrids and synthetic beings and need an ethic that is not obsessed with human-style brains.

A brief metaphysics detour. When I say “teleology” I do not mean mystical vitalism and I do not restrict it to second-order human purposefulness. Cybernetics already gives us a non-mysterious science of goal-directed systems. My empirical test for any framework is simple: Does it predict new experiments and capabilities?

My entire argument today fits on one slide:

  1. Whatever goals are, humans clearly have them.
  2. We know a bit about the brain mechanisms that enable goal-directed behaviour.
  3. Those mechanisms are ancient, highly conserved and not confined to nervous tissue. There is a deep evolutionary homology between cognition and morphogenesis, revealed most clearly by developmental bioelectricity.

I will show biological examples of goal-directed problem solving and then argue that teleology is just the first rung on a much broader ladder of cognition—“cognition all the way down.” Cognitive terms are really claims about how an observer can interact with a system. If the standard conceptual and bench tools of behavioural neuroscience work outside the brain, the boundary we draw is arbitrary.

Model system: groups of cells as a collective intelligence that “behaves” by navigating anatomical morphospace. Developmental decisions happen on the timescale of minutes not milliseconds, and in morphological rather than 3-D locomotor space, yet the algorithms—perceptual bistability, active inference, attractor dynamics—are the same.

Goal-directed anatomical homeostasis.
Example 1. The axolotl: amputate anywhere along the limb and the cells sense their departure from a target morphology, rebuild the limb and—crucially—stop when the error is minimised.
Example 2. Early embryos can be cut into pieces; each fragment makes a perfect twin because the collective converges on the same morphogenetic attractor from many start states.
Example 3. “Picasso” tadpoles with eyes, nostrils and jaws scrambled still become normal frogs; each organ follows novel trajectories to the correct final arrangement, overshooting and correcting if necessary.

Non-local control. Graft a tail tip to a salamander’s flank and it remodels into a limb—even though, locally, tail cells were in the “right” context. Large-scale pattern memories override local identity.

Prevailing textbooks treat development as feed-forward genetic chains plus local physics, hoping complexity “emerges.” But if you want three-fold symmetry instead of bilateral, that inverse problem is hopeless. Bioelectric pattern memories are stored set-points. We can detect, decode and rewrite them, letting the tissue do the heavy lifting while we merely update its goal.

Reading the body’s thoughts. Using voltage-sensitive dyes and reporters we film the embryo-wide electrical network—the “electric face” map appears hours before gene markers of the eye or mouth. Gap-junction-mediated injury waves link not only cells within an embryo but separate embryos in the same clutch, giving collective teratogen resistance.

Writing new goals. By expressing specific ion channels in gut progenitors we induce eye organs—complete with lens, retina, optic nerve—far from the head. The signal is instructive, modular, and recruits uninjected neighbours. Correct “prompts” reveal latent competence even in tissues long deemed incapable of eye formation.

Latent pattern memories. In planaria, transiently altering membrane voltage rewrites the default “one head, one tail” memory to “two heads.” The worm looks normal until amputated; then each fragment regenerates two heads indefinitely, without any genomic change. Head shape—triangular, flat, round—can likewise be rewritten, tapping a vast hidden morphospace.

Learning at multiple scales.
• Gene-regulatory networks alone show associative conditioning, habituation and counting.
• At the other extreme, deer antler “trophic memory” records a bone injury and reproduces an ectopic tine at that spatial coordinate in every regrowth for five years.

Creative problem-solving.
• Planaria bathed in barium lose their heads yet re-grow heads tolerant to barium, converging independently on the same scant handful of gene-expression changes in days.
• Highly polyploid newt embryos form kidney tubules with fewer but larger cells; extreme polyploidy makes one giant cell wrap around itself—a shift to an entirely different molecular mechanism to reach the same tissue-level goal.

Novel beings and new goals.
• Xenobots—frog epidermal cells liberated from the embryo—self-assemble into motile spheroids that push loose cells into piles; the piles mature into the next generation: kinematic self-replication, a life strategy unseen in nature. Their transcriptome diverges sharply from tadpole cells and includes genes for sound perception; indeed they alter behaviour in response to audio playback.
• Anthrobots—aggregates of human tracheal cells—self-organise into ciliated swimmers that can repair themselves and knit closed wounds in cultured neural layers. They express ~9 000 genes at different levels from their tissue of origin—no genomic edits, just a new lifestyle.

Implications.
• Competent matter changes how evolution works: selection can rely on cellular problem-solving (“the intelligence ratchet”).
• Cancer may be a breakdown of multicellular goal coherence—a bioelectric dissociative identity disorder—and can sometimes be reversed by forcing electrical reconnection.
• Stress-sharing circuits, cognitive “lycones” (the scale of attainable goals), and even simple deterministic algorithms show unexpected agency when examined with the right lens.

Thank-you slide omitted.

2. Concise summary of the argument

  1. Goal-directedness is not unique to brains.
    The same ion-channel and gap-junction networks that implement neural intent operate in non-neural tissues, guiding cells and organs toward stored anatomical set-points.

  2. Morphogenesis = collective intelligence.
    Embryos, regenerating limbs, planarian fragments and even disaggregated epithelial cells behave like problem-solving agents navigating “morphospace” rather than physical space.

  3. Bioelectric patterns are writable memories.
    They can be read with optical reporters and rewritten with ion-channel drugs, optogenetics or gap-junction modulation—changing what the tissue tries to build without altering DNA.

  4. Plasticity enables novelty.
    When freed from embryonic constraints, cells invent new bodies (xenobots, anthrobots) and new reproductive strategies, showing that genomes encode toolkits, not finished blueprints.

  5. A wider cognitive spectrum demands new ethics and new medical strategies.
    Regenerative medicine should focus on updating pattern memories; oncology should treat tumours as goal-confused sub-agents; moral status should turn on capacities, not substrates.

3. Critique

Aspect Strengths Weaknesses / Open questions
Empirical foundation Dozens of elegant experiments (axolotl limbs, planarian polarity, induced ectopic eyes) convincingly demonstrate bioelectric control and its rewritability. Some headline-grabbing feats (two-headed worms, xenobot replication) rely on specialised protocols and have not yet been replicated by independent groups.
Conceptual clarity Re-frames development as cybernetic control, elegantly linking cognitive science and morphogenesis. Provides a testable definition of “goal”: it must be rewritable. At times the talk conflates homology with analogy (shared ancestry vs. shared logic). Ion-channel signalling is ancient, but whether the higher-level algorithms are truly conserved or merely similar remains to be shown.
Technological promise Offers non-genetic avenues for regenerative medicine and cancer therapy—reset the target pattern instead of micromanaging cells. Translating frog or planarian bioelectric interventions to mammals is still largely speculative; human tissues have more layered regulation and immune constraints.
Philosophical reach Challenges anthropocentric ethics; forces reconsideration of what counts as an “agent.” Provides a spectrum-based framework rather than binary categories. The leap from cellular homeostasis to moral patiency is asserted, not argued. A principled way to map degrees of agency to concrete rights or duties is still missing.
Methodological caution The “use any framework that predicts new experiments” stance keeps the program grounded in empiricism. Occasional rhetorical overreach (“cognition all the way down,” “agential material”) may alienate mainstream biologists who could adopt the practical tools without the metaphysical baggage.
Reproducibility & statistics Many datasets are publicly available; voltage dyes and gap-junction modulators are standard reagents, aiding reproducibility. Some demonstrations (e.g., antler trophic memory) are based on legacy natural-history datasets that are difficult to reproduce under modern lab conditions.
Integration with genetics Shows convincingly that DNA supplies parts lists, not blueprints; bioelectric patterning adds the missing software layer. Ignores other “software” layers—mechanotransduction, epigenetics, metabolic gradients—which likely interact with bioelectric cues in complex feedback loops. A full model must integrate all.

Overall assessment

Michael Levin’s work is a rare blend of bold theory and hands-on bench science. The evidence that bioelectric circuits store and recall anatomical set-points is robust, and the ability to rewrite those memories is a genuine paradigm shift for developmental biology and regenerative medicine.

Where the programme is least mature is in scaling to mammalian systems, quantifying degrees of proto-cognition, and integrating bioelectricity with other non-genetic information layers. The philosophical extrapolations (ethics for hybrid intelligences) are stimulating but currently outrun the data.

Still, the core claim—that teleology and rudimentary cognition emerge whenever problem-solving networks organise, no matter the substrate—is now on firm empirical footing and deserves serious attention across biology, robotics and philosophy of mind.

Recommended next reads

  • Levin, M. & Dennett, D. “Cognition all the way down” (Aeon, 2020)
  • Pezzulo & Levin, “Top-down models in biology: the guiding hand of goals.” Trends in Cognitive Sciences (2023)
  • Fields et al., “The bioelectric code.” Quarterly Review of Biology (2022)

(All open access on Levin lab website.)