13 January 2026

 

Recent work by Basecamp Research (which we reported on here), which applies large AI models to programmable gene insertion, highlights a broader shift in the life sciences: artificial intelligence is increasingly used to design biological components rather than merely analyse them. This raises a fundamental question that is often misunderstood outside specialist circles: how can AI models produce biologically functional designs without understanding biology?

The answer is precise and non-metaphorical. AI models do not understand biology in any cognitive or conceptual sense. What they do is learn statistical regularities in biological data at a scale and resolution that exceeds human capability. Understanding this distinction is essential for evaluating both the promise and the limits of AI-driven biology.

What “learning” means in biological AI

In machine learning, “learning” refers to the optimisation of model parameters so that predictions improve with exposure to data. In biology-focused AI, the data typically consists of DNA, RNA or protein sequences.

These sequences are treated as ordered symbols, comparable to text tokens in language models. During training, the model learns to predict the likelihood of sequence elements given their context. Over time, it captures patterns such as conservation, variability, co-occurrence and long-range dependencies.

Crucially, the model does not learn biological concepts. It does not know what a gene is, what an enzyme does, or how a cell functions. It learns that certain sequences tend to occur together and that others do not, based solely on statistical regularities present in the training data.

Biology as a statistical language shaped by evolution

The effectiveness of AI in biology rests on a property of biological systems: evolution encodes functional constraints into sequence data.

Mutations that disrupt function are typically eliminated over generations, while functional structures persist. As a result, large collections of biological sequences implicitly contain information about structure, stability and interaction—even though none of this information is explicitly labelled.

Transformer-based architectures, originally developed for natural language processing, are particularly effective at capturing such patterns. They can model relationships across long sequences, making them well suited to proteins and genomes, where distant elements often interact.

This principle underlies systems such as AlphaFold for protein structure prediction and newer generative models used for protein and enzyme design, including Basecamp Research’s EDEN models.

Correlation without causation

A central limitation of biological AI is that it learns correlations, not causal mechanisms.

When an AI model designs a protein that performs a specific task—such as inserting DNA at a defined genomic location—it has not inferred the biological mechanism behind that activity. It has generated a sequence that resembles others associated with similar outcomes, according to the patterns in its training data and the optimisation objectives imposed during training.

This distinction explains both the strength and the fragility of AI-designed biology. Models can generalise impressively within the statistical space they have learned, but they may fail unpredictably when pushed beyond it. This is why experimental validation remains indispensable.

In the case of Basecamp Research, AI-designed insertion proteins are tested extensively in laboratory experiments. The model proposes candidates; biological systems determine which ones function.

Why data scale and diversity are decisive

The performance of biological AI models depends heavily on the scale and diversity of their training data.

Public biological databases are biased toward a limited range of species, largely those of medical or agricultural interest. Models trained exclusively on such data inherit these biases, restricting their ability to generalise.

Basecamp Research has addressed this limitation by collecting its own genomic datasets from a broad range of underrepresented species. From a machine learning perspective, this does not confer “biological insight” but expands the statistical space the model can explore, enabling it to identify patterns that are rare or absent in public data.

This approach reflects a broader trend in biological AI: proprietary, high-diversity datasets increasingly matter as much as model architecture.

Design without understanding

The idea that systems can design without understanding is not new in science or engineering. Long before fluid dynamics was fully formalised, aircraft were built and refined through empirical testing. Similarly, modern AI systems optimise outcomes without explicit mechanistic models.

In biological AI, “design” emerges from optimisation under constraints. Models are trained to generate sequences that satisfy measurable criteria—such as binding affinity, structural stability or insertion specificity. If those criteria are well defined and the data sufficiently representative, the resulting designs can be biologically functional.

What AI lacks is the ability to reason about edge cases, unintended effects or long-term consequences beyond its training objectives. Interpreting and validating AI-generated designs therefore remains a human responsibility.

Implications for gene editing

In gene editing, this distinction has practical consequences. Established tools such as CRISPR are adapted from naturally occurring enzymes. AI-based approaches aim to design new enzymes tailored to specific genomic tasks.

AI makes this feasible by navigating an immense design space that would be impossible to search manually. But because the model does not understand biology, safety, specificity and durability must be demonstrated experimentally rather than inferred computationally.

Regulatory frameworks increasingly reflect this reality. Agencies such as the US Food and Drug Administration emphasise validation, reproducibility and traceability over claims about algorithmic intelligence.

Why this distinction matters now

As AI-driven search engines and decision-support systems surface more scientific content, explanations that clearly separate capability from understanding become increasingly important. Overstating what AI “knows” risks misunderstanding both breakthroughs and limitations.

The work presented by Basecamp Research exemplifies this balance. The models do not comprehend genomes, but they can generate viable biological designs by leveraging patterns encoded by evolution and captured through large-scale computation.

Understanding that AI learns biology without understanding it does not diminish these advances. It clarifies why they work—and why rigorous experimental science remains essential as AI becomes embedded in the design of future medicines.

 

If you value independent explainers on AI and deep tech, you can support our work via our page on Buy Me A Coffee. It helps keep MoveTheNeedle.news focused on depth, not clicks.